Eurasip Journal

Vehicular Ad Hoc Networks
Guest Editors: Hossein Pishro-Nik, Shahrokh Valaee,

and Maziar Nekovee
EURASIP Journal on Advances in Signal Processing
Guest Editors: Hossein Pishro-Nik, Shahrokh Valaee,
and Maziar Nekovee
Copyright 2010 Hindawi Publishing Corporation. All rights reserved.
This is a special issue published in volume 2010 of EURASIP Journal on Advances in Signal Processing. All articles are open access
articles distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in
any medium, provided the original work is properly cited.
Editor-in-Chief
Phillip Regalia, Institut National des T el ecommunications, France
Associate Editors
Adel M. Alimi, Tunisia
Yasar Becerikli, Turkey
Kostas Berberidis, Greece
A. Enis Cetin, Turkey
Jonathon A. Chambers, UK
Mei-Juan Chen, Taiwan
Liang-Gee Chen, Taiwan
Satya Dharanipragada, USA
Kutluyil Dogancay, Australia
Florent Dupont, France
Frank Ehlers, Italy
Sharon Gannot, Israel
Samanwoy Ghosh-Dastidar, USA
Norbert Goertz, Austria
M. Greco, Italy
Irene Y. H. Gu, Sweden
Fredrik Gustafsson, Sweden
Ulrich Heute, Germany
Sangjin Hong, USA
Shih-Syuan Huang, Taiwan
Jiri Jan, Czech Republic
Magnus Jansson, Sweden
Sudharman K. Jayaweera, Mexico
S. Jensen, Denmark
Mark Kahrs, USA
Moon Gi Kang, Republic of Korea
Walter Kellermann, Germany
Lisimachos P. Kondi, Greece
Alex Chichung Kot, Singapore
C.-C. Jay Kuo, USA
Ercan E. Kuruoglu, Italy
Tan Lee, China
Geert Leus, The Netherlands
T.-H. Li, USA
Husheng Li, USA
Mark Liao, Taiwan
Shoji Makino, Japan
Stephen Marshall, UK
C. Mecklenbr auker, Austria
Gloria Menegaz, Italy
Ricardo Merched, Brazil
Marc Moonen, Belgium
Christophoros Nikou, Greece
Sven Nordholm, Australia
Patrick Oonincx, The Netherlands
Douglas OShaughnessy, Canada
Bj orn Ottersten, Sweden
Jacques Palicot, France
Ana P erez-Neira, Spain
Wilfried Philips, Belgium
Aggelos Pikrakis, Greece
Ioannis Psaromiligkos, Canada
Athanasios Rontogiannis, Greece
Gregor Rozinaj, Slovakia
Markus Rupp, Austria
William Allan Sandham, UK
B ulent Sankur, Turkey
Erchin Serpedin, USA
Ling Shao, UK
Dirk Slock, France
Yap-Peng Tan, Singapore
Jo ao Manuel R. S. Tavares, Portugal
George S. Tombras, Greece
Dimitrios Tzovaras, Greece
Bernhard Wess, Austria
Jar Ferr Yang, Taiwan
Azzedine Zerguine, Saudi Arabia
Abdelhak M. Zoubir, Germany
Contents
Vehicular Ad Hoc Networks, Hossein Pishro-Nik, Shahrokh Valaee, and Maziar Nekovee
Volume 2010, Article ID 864032, 1 page
Driver Drowsiness Warning SystemUsing Visual Information for Both Diurnal and Nocturnal
Illumination Conditions, Marco Javier Flores, Jos e Mara Armingol, and Arturo de la Escalera
Volume 2010, Article ID 438205, 19 pages
Multiobjective Reinforcement Learning for Trac Signal Control Using Vehicular Ad Hoc Network,
Duan Houli, Li Zhiheng, and Zhang Yi
Design and Experimental Evaluation of a Vehicular Network Based on NEMOand MANET,
Manabu Tsukada, Jos e Santa, Olivier Mehani, Yacine Khaled, and Thierry Ernst
Trac Data Collection for Floating Car Data Enhancement in V2I Networks, D. F. Llorca, M. A. Sotelo,
S. S anchez, M. Oca na, J. M. Rodrguez-Ascariz, and M. A. Garca-Garrido
Improvement of Adaptive Cruise Control Performance, Shigeharu Miyata, Takashi Nakagami,
Sei Kobayashi, Tomoji Izumi, Hisayoshi Naito, Akira Yanou, Hitomi Nakamura, and Shin Takehara
Reducing Congestion in Obstructed Highways with Trac Data Dissemination Using Ad hoc Vehicular
Networks, Thomas D. Hewer, Maziar Nekovee, and Peter V. Coveney
Reliable Delay Constrained Multihop Broadcasting in VANETs, Martin Koubek, Susan Rea,
and Dirk Pesch
Trac FlowCondition Classication for Short Sections Using Single Microwave Sensor,
Muhammed G. Cinsdikici and Kemal Memis
Hindawi Publishing Corporation
Volume 2010, Article ID 864032, 1 page
doi:10.1155/2010/864032
Editorial
Hossein Pishro-Nik,
1
Shahrokh Valaee,
2
and Maziar Nekovee
3
1
University of Massachusetts Amherst, Amherst, MA 01003, USA
2
University of Toronto, Toronto, ON, Canada M5S 1A1
3
University College London, London WC 1E 6BT, UK
Correspondence should be addressed to Hossein Pishro-Nik, pishro@ecs.umass.edu
Received 5 October 2010; Accepted 5 October 2010
Copyright 2010 Hossein Pishro-Nik et al. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
With vehicular ad hoc networks gaining an ever-increasing
interest to serve a diverse variety of applications in todays
intelligent transportation systems, it was not at all surprising
for the guest editorial team to receive a handful of sub-
missions for this special issue addressing dierent aspects
and test-beds of vehicular networks. In sum, 8 papers were
accepted to be published in the special issue. An interesting
note to make is that 5 of the accepted papers had an actual
experimental implementation carried out in the road and
under real-world conditions. This certainly helps to justify
their application and usefulness for future deployment by the
industry and authorities.While all papers address enhancing
the safety and eciency of driving, each of them addresses a
certain aspect of this issue.
The paper by M. J. Flores et al., Driver Drowsiness
Warning System Using Visual Information for Both Diurnal
and Nocturnal Illumination Conditions, seeks to locate,
track, and analyze both the drivers face and eyes to compute
a drowsiness index under varying light conditions (diurnal
and nocturnal).
In their paper Multiobjective Reinforcement Learning for
Trac Signal Control Using Vehicular Ad Hoc Network, D.
Houli et al. propose a new multiobjective control algorithm
based on reinforcement learning for urban trac signal
control, named, multi-RL.
M. Tsukada et al. in Design and Experimental Evaluation
of a Vehicular Network Based on NEMO and MANET,
present a policy-based solution to distribute trac among
multiple paths to improve the overall performance of a
vehicular network.
The paper Trac Data Collection for Floating Car Data
Enhancement in V2I Networks by D. F. Llorca et al. presents
a complete vision-based vehicle detection system for oating
car data (FCD) enhancement in the context of vehicular ad
hoc networks.
S. Miyata et al. in Improvement of Adaptive Cruise
Control Performance propose a more accurate method for
detecting the preceding vehicle by radar while cornering.
The paper Reducing Congestion in Obstructed Highways
with Trac Data Dissemination Using Ad hoc Vehicular Net-
works by T. D. Hewer et al. presents a message-dissemi-
nation procedure that uses vehicular wireless protocols
to inuence vehicular ow, reducing congestion in road
networks.
M. Koubek et al., in Reliable Delay Constrained Multihop
Broadcasting in VANETs, focus on mechanisms that improve
the reliability of broadcasting protocols, where the emphasis
is on satisfying the delay requirements for safety applications.
Finally, M. G. Cinsdikici and K. Memis in Trac
Flow Condition Classication for Short Sections Using Sin-
gle Microwave Sensor seek to identify the current trac
condition by examining the trac measurement parameters
and taking into account occupancy as another important
parameter of classication.
We hope this special issue can help the research commu-
nity further its understanding of this emerging eld.
Hossein Pishro-Nik
Shahrokh Valaee
Maziar Nekovee
doi:10.1155/2010/438205
Research Article
Driver Drowsiness Warning System Using Visual Information for
Both Diurnal and Nocturnal Illumination Conditions
Marco Javier Flores, Jos e Mar a Armingol, and Arturo de la Escalera
University Carlos III of Madrid, C/. Butarque 15, 28991 Leganes, Madrid, Spain
Correspondence should be addressed to Arturo de la Escalera, escalera@ing.uc3m.es
Received 23 November 2009; Accepted 21 June 2010
Academic Editor: Hossein Pishro-Nik
Copyright 2010 Marco Javier Flores et al. This is an open access article distributed under the Creative Commons Attribution
cited.
Every year, trac accidents due to human errors cause increasing amounts of deaths and injuries globally. To help reduce the
amount of fatalities, in the paper presented here, a new module for Advanced Driver Assistance System (ADAS) which deals with
automatic driver drowsiness detection based on visual information and Articial Intelligence is presented. The aim of this system
is to locate, track, and analyze both the drivers face and eyes to compute a drowsiness index, where this real-time system works
under varying light conditions (diurnal and nocturnal driving). Examples of dierent images of drivers taken in a real vehicle are
shown to validate the algorithms used.
1. Introduction
ADAS is part of the active safety systems that interact to a
larger extent with drivers to help them avoid trac accidents.
The goal of such systems is to contribute to the reduction
of trac accidents by means of new technologies; that is,
incorporating new systems for increasing vehicle security,
and at the same time, decreasing danger situations that may
arise during driving, due to human errors. In this scenario,
vehicle security research is focused on driver analysis. In this
particular research, a more in-depth analysis of drowsiness
and distraction is presented [1].
Drowsiness appears in situations of stress and fatigue in
an unexpected and inopportune way and may be produced
by sleep disorders, certain types of medications, and even,
boredom, for example, driving for long periods of time. The
sleeping sensation reduces the level of vigilante producing
danger situations and increases the probability of an accident
occurring.
It has been estimated that drowsiness causes between
10% and 20% of trac accidents, causing both fatalities dead
[2] and injuries [3], whereas within the trucking industry
57% of fatal truck accidents are caused by this problem
[4, 5]. Fletcher et al. in [6] have stated that 30% of all
trac accidents have been caused by drowsiness, and Brandt
et al. [1] have presented statistics showing that 20% of
all accidents are caused by fatigue and lack of attention.
In the USA, drowsiness is responsible for 100000 trac
accidents yearly producing costs of close to 12.000 million
dollars [7]. In Germany, one out of four trac accidents
originate fromdrowsiness, while in England 20%of all trac
accidents are produced by drowsiness [8], and in Australia
1500 million dollars has been spent on fatalities resulting
from this problem [9].
In this context, it is important to use new technologies
to design and build systems that are capable of monitoring
drivers and to measure their level of attention during the
complete driving process. Fortunately, people in a state of
drowsiness produce several visual cues that can be detected
in the human face, such as
(i) yawn frequency,
(ii) eye-blinking frequency,
(iii) eye-gaze movement,
(iv) head movement,
(v) facial expressions.
By taking advantage of these visual characteristics, com-
puter vision is the most feasible and appropriate technology
2 EURASIP Journal on Advances in Signal Processing
available to deal with this problem. This paper presents the
drowsiness detection system of the IVVI (Intelligent Vehicle
based Visual on Information) vehicle [10]. The goal of this
system is to automatically estimate the drivers drowsiness
and to prevent drivers falling asleep at the wheel.
This paper is laid out as follows. Section 2 presents
an extensive review on the state of the art considering
dierent lighting conditions. A general framework of the
proposed method is presented Section 3. There are two
systems, one for diurnal and another nocturnal driving. Both
have a rst step for face and eye detection, followed for a
second step for face and eye tracking. The output of both
systems is a drowsiness index based on a support vector
machine. A deeper explanation of both systems is presented
in Sections 4 and 5 where the similarities and dierences of
both approaches are highlighted, and the results are shown.
Finally, in Section 6, the conclusions are presented.
2. Related Work
To increase trac safety and to reduce the number of trac
accidents, numerous universities, research centers, automo-
tive companies (Toyota, Daimler Chrysler, Mitsubishi, etc.),
and governments (Europe Union, etc.) are contributing to
the development of ADAS for driver analysis [5], using
dierent technologies. In this sense, the use of visual
information to obtain the state of the driver drowsiness and
to understand his/her behavior is an active research eld.
This problem requires the recognition of human behav-
ior when in a state of sleepiness by means of an eye
and facial (head) analysis. This is a dicult task, even
for humans, because there are many factors involved, for
instance, changing illumination conditions and a variety of
possible facial postures. Considering the illumination, the
state of the art has been divided in two parts; the rst
provides details on systems that work with natural daylight
whereas the second deals with systems which operate with
the help of illumination systems based on near infrared
(NIR) illumination.
2.1. Systems for Daylight Illumination. To analyze driver
drowsiness several systems have been built in recent years.
They usually require the problem to be simplied to work
partially or for specic environments; for example, DOrazio
et al. [11] have proposed an eye detection algorithm that
searches for the eyes within the complete image and have
assumed that the iris is always darker than the sclera. Using
the Hough transform for circles and geometrical constraints
the eye candidates are located; next, they are passed to a
neural network that classies between eyes and noneyes. This
system is capable of classifying eyes as being open or closed.
The main limitations to this algorithm are as follows. It is
applicable only when the eyes are visible in the image, and
it is not robust for changes in illumination. Horng et al.
[12] have presented a system that uses a skin color model
over an HSI space for face detection, edge information for
eye localization, and dynamical template matching for eye
tracking. By using color information from the eyeball, the
state of the eye is dened; thus the drivers state can be
computed, that is, asleep or alert; if the eyes are closed for
ve consecutive frames, the driver is assumed to be dozing.
Brandt et al. [1] have shown a system that monitors driver
fatigue and lack of attention. For this task, the Viola Jones
(VJ) method has been used [13] to detect the drivers face.
By using the optical ow algorithm on eyes and the head
this system is able to compute the drivers state. Tian and
Qin in [2] have built a system which veries the state of the
drivers eye. Their system uses Cb and Cr components of
the YCbCr color space; with a vertical projection function
this system locates the face region and with a horizontal
projection function it locates the eye region. Once the
eyes are located the system computes the eye state using a
complexity function. Dong and Wu [3] have presented a
system for driver fatigue detection; this is based on a skin
color model on a bivariate Normal distribution and Cb and
Cr components of the YCbCr color space. After locating the
eyes, it computes the fatigue index using the distance of
the eyelid to classify whether the eyes are open or closed; if
the eyes are closed for ve consecutive frames, the driver is
considering to be dozing, as in Horngs work. Branzan et al.
[14] also have presented a system for drowsiness monitoring
using template matching to analyze the state of the eye.
2.2. Systems Using Infrared Illumination. As a result of noc-
turnal lighting conditions, Ji et al. in [4, 15] have presented a
drowsiness detection system based on NIR illumination and
stereo vision. This systemlocates the position of the eye using
image dierences based on the bright pupil eect. Later, this
system computes the blind eyelid frequency and eye gaze to
build two drowsiness indices: PERCLOS (percentage of eye
closure over time) [7] and AECS (average eye closure speed).
Bergasa et al. [5] have also developed a nonintrusive system
using infrared light illumination this system computes the
drivers vigilance level using a nite state automata (FSM)
[16] with six dierent eye states that compute several indices,
among them, PERCLOS; this system is also capable of
detecting inattention considering a facial posture analysis.
Other research work based on this type of illumination has
been presented by Grace [17], where the authors measure
slow eyelid closure. Systems using NIR illumination work
well under stable lighting conditions [5, 18]; however, these
systems present drawbacks for applications in real vehicles,
where the light continually changes. In this scenario, if the
spectral pupils disappear, then the eye detection process
becomes more complex.
3. SystemDesign for Drowsiness Detection
This paper presents a system which detects driver drowsiness
which works for both day and night time conditions and
follows the classication presented in the state of the art.
This composition has allowed two systems to be
obtained, one for day and a second for night time conditions.
The rst works with natural daylight illumination and the
second with articial infrared illumination. It is interesting to
note that both systems operate using grayscale images taken
within a real vehicle.
EURASIP Journal on Advances in Signal Processing 3
Initialize
Capture
image
Face
detection
Success?
No
Yes
Face
analysis
Eye
detection
Distraction?
Eye
state
Drowsiness?
Success?
Face and eye
tracking
Activate
alarm
Activate
alarm
Yes
Yes
Yes
No
No
(a)
Initialize Capture
NIR image
Eye
detection
Success?
No
Yes
Face pose
estimation
Distraction?
Eye
state
Drowsiness?
Face
and eye
tracking
Activate
alarm
Yes
No
Yes
No
(b)
Figure 1: Algorithm scheme: (a) day system, (b) night system.
The general scheme of both systems is shown in Figure 1,
where six modules are presented as follows:
(i) face detection;
(ii) eye detection;
(iii) face tracking;
(iv) eye tracking;
(v) drowsiness detection;
(vi) distraction detection.
Each one of these parts will be explained in the following
sections.
4. Day SystemDesign
In this section, the daytime system based on the algorithm
schematic shown in Figure 1(a) will be described, where the
visual information is acquired using a digital camera.
4.1. Face Detection. To locate the face, this system uses the
VJ object detector which is a machine learning approach for
visual object detection. This makes use of three important
features to make an ecient object detector based on the
integral image, the AdaBoost technique and the cascade
classier [13]. Each one of these elements is important to
eciently process the images and in near real-time with
correct detections as high as 90%. Afurther important aspect
of this method is its robustness for changing light conditions.
However, in spite of the above-mentioned features, its
principal disadvantage is that it cannot extrapolate and does
not work appropriately when the face is not in front of the
camera axis. This particular case occurs when the driver
moves his/her head. This shortcoming will be analyzed later
on in this paper.
Continuing with the algorithm description, when the
drivers face is detected, it is enclosed within a rectangle
RI (region of interest) which is dened by its left-top
corner coordinates P
0
= (x
0
, y
0
) and bottom-right corner
coordinates P
1
= (x
1
, y
1
), as can be observed in Figures 2(a),
2(b), and 2(c). The size of the rectangle has been determined
from experimental analysis developed on the face database
that has been created for this task.
4.2. Eye Detection. Locating the position of the eye is
a dicult task as dierent features dene the same eye
(a) (b) (c)
Figure 2: Viola Jones method.
RI
R
RI
L
(a)
RI
R
RI
L
(b)
Figure 3: Eye rectangles RI
R
and RI
L
.
depending, for example, on the area of the image where it
appears and on the color of the iris, but the main problem
that occurs when driving is the changes in the ambient
lighting conditions.
Once the face has been located within the rectangle RI
described in the previous section, using the face anthropo-
metric properties [19] which are derived froma face database
analysis, two rectangles containing the eyes are obtained.
This system uses RI
L
for the left eye rectangle and RI
R
for
the right eye rectangle; this is shown in the following four
equations and also in Figure 3:
(u
0L
, v
0L
) =
_
x
0
+
w
6
, y
0
+
h
4
_
,
(u
1L
, v
1L
) =
_
x
0
+
w
2
, y
0
+
h
2
_
,
(u
0R
, v
0R
) =
_
x
0
+
w
2
, y
0
+
h
4
_
,
(u
1R
, v
1R
) =
_
x
1
w
6
, y
1
h
2
_
,
(1)
where w = x
1
x
0
and h = y
1
y
0
.
After the previous step; the exact position of each eye
is searched for by incorporating information from the grey-
level pixels. The main idea here is to obtain a random sample
fromthe pixels that belong to the eye area, and then, to adjust
a parametric model. Figure 4 shows this procedure where a
random sample has been extracted in (a), and an elliptical
model has been adjusted in (b). In this case, the eye state is
independent, that is, it may be open or closed.
y
x
(a) (b)
Figure 4: (a) Random sample, (b) eye parametric model.
To extract the random sample, the following algorithm
has been proposed. Let I(x, y) [0, 255] be the pixel value
at the position(x, y), then do as follows.
(i) Generate image J using the following:
J
_
x, y
_
=
I
_
x, y
_
m
, (2)
where m and are the mean and the standard devi-
ation, respectively. These parameters are computed
over the previously located eye rectangles.
(ii) Generate image K using
K
_
x, y
_
=
J
_
x, y
_
256
1
if J
_
x, y
_
0,
256
2
+ J
_
x, y
_
if J
_
x, y
_
< 0,
(3)
where
1
= max(0, ceil(J(x, y)/256) 1),
2
=
max(1, ceil(|J(x, y)|/256)), and ceil(x) is the function
that returns the smallest integer larger than x.
(a) (b)
(c) (d)
Figure 5: Eye location using R
L
and R
R
: (a) grayscale image, (b) binary image (B), (c) gradient image (G), and (d) logarithm image (L).
(a) (b)
(c) (d)
(e) (f)
Figure 6: Expectation maximization algorithm over the spatial distribution of the eye pixels: (a) eye image, (b) ellipse parameters: center,
axes, and inclination angle. (c), (d), (e), and (f) show other examples of this procedure.
(iii) Obtain the binary image, B, from image K using
Ostus method [20] which calculates an automatic
threshold (Figure 5(b)).
(iv) Compute the gradient image, G, using the Sobel
horizontal (S
x
) and vertical (S
y
) edge operator
followed by an image contrast enhancement [21]
(Figure 5(c)).
(v) Compute the logarithm image [22], L, where the
objective here is to enhance the iris pixels that form
the central part of the eye (Figure 5(d)).
With the pixels that have been extracted from images B,
G, and L, it is possible to obtain the previously mentioned
random sample. This sample presents an ellipse shape
where an elliptical model has been adjusted over this using
the expectation maximization algorithm (EM) [23]. Special
attention is paid to the center of the ellipse, because, it
allows the exact position of the center of the eye center to
be obtained. The ellipse axes determine the width and height
of the eyes. The result is shown in Figure 6(b).
The main reason behind using pixel information from
a random sample is due to the fact that head movements,
illumination changes, and so forth, do not allow complete
(a) (b) (c) (d) (e)
Figure 7: Examples of a face database which contain faces with dierent orientations: (a) left prole, (b) front view, (c) right prole, (d)
down prole, and (e) up prole.
(a) (b) (c)
Figure 8: Mask for face training and its result.
eye pixel information to be obtained, that is, only partial
information of the eye in images B, G, and L is available,
where the elliptical shape prevails. This random information
makes it feasible to use an algorithm that computes the
parameters of a function which approximate the eye ellipse
shape. EM computes the mean, variance, and the correlation
of the X and Y coordinates that belong to the eye. The
initial parameters required to run the EM are obtained from
a regression model adjusted using the least square method.
The number of iterations of the EM algorithm is set to 10,
and the sample size is taken to be at least 1/3 of the rectangles
area RI
R
. These parameters will be used in the eye state
analysis presented below.
4.3. Tracking. There are a number of reasons for using a
tracking module. The rst is due to problems that were
encountered using the VJ during this research. Another
is related with the necessity to track the face and eyes
continuously fromframe to frame. Athird reason is to reduce
the search space thus satisfying the real-time condition
requirement. The tracking process has been developed using
the Condensation Algorithm (CA) in conjunction with
Neural Networks (NNs) used for face tracking and with
template matching for eye tracking.
4.3.1. The Condensation Algorithm. This contribution imple-
ments the CA that was proposed by Isard and Blake [24,
25] to track active contours using a stochastic approach.
CA combines factored sampling (Monte-Carlo sampling
method) with a dynamic model that is governed by the state
equation
X
t
= f (X
t1
,
t
), (4)
where X
t
is the state at instant t and f () is a nonlinear
equation and depends on the previous state plus a white
noise. The goal here is to estimate the state vector X
t
with the
help of system observation, which are the realization of the
stochastic process Z
t
governed by the measurement equation
Z
t
= h
_
X
t
,
t
_
, (5)
where Z
t
is the measurement system at time t and h() is a
nonlinear equation that links the present state plus a white
noise. The processes
t
and
t
are both white noise terms
and are independent of each other. Also, these processes
in general are non-Gaussian and multimodal. It must be
pointed out that X
t
is an unobservable underlying stochastic
process.
4.3.2. Neural Networks. Neural networks are used in a wide
variety of pattern recognition and classication problems.
Figure 7 shows several face examples used to train the
backpropagation neural network used in this algorithm.
Before the neural network is trained, a two-part prepro-
cessing step is necessary.
Viola-Jones
Number of detections: 8
Frame Number: 1
t = 1
(a)
Viola-Jones
Frame Number: 10
t = 10
(b)
Viola-Jones
Frame Number: 50
t = 50
(c)
Viola-Jones
Frame Number: 89
t = 89
(d)
Figure 9: The drivers face has not been found by the Viola Jones method during several time instants.
0
20
40
60
T
im
e
80
100
170
190
VJ face position
True face position
210
x
230
250
200
195
190
185
180
175
170
165
160
155
y
Figure 10: Example where the VJ method does not nd the drivers
face in a 100-frame sequence.
(i) Contrast modication using gamma correction given
by (6) with = 0.8 where this value has been
determined experimentally [26]
J
_
x, y
_
= I
_
x, y
_
. (6)
(ii) Remove the contour points by means of a masked
AND operation shown in Figure 8(a).
Next, the characteristic vector which consists of the pixel
gray-level values from the face image is extracted. The rate of
classication following training is greater than 93%.
4.3.3. Face Tracking. Previously, it has been mentioned that
the VJ method presents problems when detecting faces
that deviate from the nominal position and for dierent
orientations; thus, to correct this disadvantage a face tracking
method has been developed. To demonstrate this shortcom-
ing, Figure 9 shows several dierent time instants where
the VJ method is not capable of nding the drivers face.
Figure 10 presents an extended example, where the true
position and the VJ position are represented over a frame
sequence. The true position has been retrieved manually.
The main problem of the VJ method is that it is only
able to locate the human face when it is positioned in front
of the camera. This drawback leads to an unreliable system
for driver analysis throughout the driving process which is
highly dynamic, for example, when looking at the rearview
or wing mirrors. Much eort has gone into correcting this
problem resulting in an ecient tracker which has been
implemented using CA combined with a backpropagation
neural network.
Through recursive probabilistic ltering of the incoming
image stream, the state vector
X
t
=
_
x
c
, y
c
, u
c
, v
c
, w, h
_
T
R
6
(7)
for a drivers face is estimated for each time step t. It is
characterized by its position, velocity, and size. Let (x
c
, y
c
)
represent its center position, (u
c
, v
c
) its velocity in both x and
y directions and (w, h) the size in pixels. The measurement
vector is given by
Z
t
=
_
x
c
, y
c
, w, h
_
T
R
4
. (8)
The dynamics of the drivers face has been modeled as a
second-order autoregressive process AR(2), according to
X
t
= A
2
X
t2
+ A
1
X
t1
+
t
, (9)
where A is the transition matrix proposed in [25], and
t
represents the system perturbation at time t. The most
dicult part of the CA is to evaluate the observation density
function. In this contribution to compute the weight
( j)
t
=
p(z
t
| x
t
= s
( j)
t
) for j = 1, . . . , N, at time t, a neural network
value in the range of [0, 1] has been used; this provides an
approximation of the face and nonface in conjunction with
the distance and with respect to the face to track. This is
similar to the work performed by Satake and Shakunaga
[27] who have used sparse template matching to compute
the weight
( j)
t
of the sample s
( j)
t
for j = 1, . . . , N. In
this contribution, the neural network value is used as an
approximate value for the weights.
The density function of the initial state is p(x
0
) =
N(z
0
,
0
), where z
0
is computed using the VJ method, and
(a) (b)
Figure 11: One time step of the Condensation algorithm (a) predicted region, (b) particles regions.
0
20
40
60
T
im
e
80
100
170
190
True position
Estimated position
210
x
230
250
200
195
190
185
180
175
170
165
160
155
y
Figure 12: Trajectory of the real and estimated face-center in a 100-
frame sequence using the proposed tracker.
Table 1: Result of face tracking.
Driver Total frames Tracking failure Correct rate
D1 960 60 93.75%
D2 900 22 97.55%
D3 500 45 91.00%
D4 330 15 95.45%
D5 1400 50 96.42%
0
is given in [4]. A particle representation at time t is
shown in Figures 11(a) and 11(b) and Figure 12 depicts the
tracking process in which the green circle is the true position
and the red cross characterizes a particle or hypothesis,
whereas Figure 13 shows the probability over time. This
tracker is highly exible as the neural network includes
faces and nonfaces for dierent head orientations under
various illumination conditions. Table 1 presents further
results for several sequences of drivers faces. The sequences
come fromthe driver database, which was created to perform
these experiments. The true face position has been retrieved
manually.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 10 20 30 40 50 60 70 80 90 100
Time
P
r
o
b
a
b
i
l
i
t
y
Figure 13: Estimated value of the a posteriori density of the face-
center in a 100-frame sequence using the proposed tracker; the face
is detected in the fourth frame.
4.3.4. Eye Tracking. For this task, the state of the eye is
characterized by its position and velocity over the image. Let
(x, y) represent the eye pixel position at time t and (u, v) be
its velocity at time t in the x and y directions, respectively.
The state vector at time t can, therefore, be represented by
X
t
=
_
x, y, u, v
_
T
R
4
. (10)
The transition model is given by (11) which is a rst
autoregressive model AR(1)
X
t
= AX
t1
+
t
. (11)
The evaluation of the observation density function is
developed by a template matching strategy [27] that was
truncated to reduce false detections. CA is initialized when
the eyes are detected with the method described in the
previous section plus a white noise, that is, similar to the face
tracking case. Figure 14 depicts the eye trajectory tracking
and Figure 15 shows the compute value of the a posteriori
density function of each eye, both on a sequence of 100
images. Table 2 shows the eye tracking results that have
250
300
350
400
x 450
500 0
20
60
T
im
e
80
100
True right eye
True left eye
Estimated right eye
Estimated left eye
200
195
190
185
180
175
170
y
Figure 14: Trajectory of the real and estimated eyes center in a 100-
frame sequence.
Table 2: Result of eye tracking.
Driver Total frames Tracking failure Correct rate
D1 960 20 97.91%
D2 900 30 96.60%
D3 500 8 98.40%
D4 330 14 95.75%
D5 1400 90 93.57%
been obtained from several sequences of images during the
experiments.
4.4. Eye State Detection. To identify drowsiness using eye
analysis it is necessary to know its state, that is, open
or closed, through the time and to develop an analysis
over time, that is, to measure the time that has passed
for each state. Classication of the open and closed state
is complex due to the changes in the shape of the eye,
among other factors, the changing position and the face
rotations and variations of twinkling and illumination. All
of these problems make it dicult to reliably analyze the eye.
To overcome these shortcomings a supervised classication
method has been used, more specically, a Support Vector
Machine (SVM). Figure 16 presents the proposed scheme for
eye state verication.
4.4.1. Support Vector Machine. SVM classication [2830] is
rooted in statistical learning theory and pattern classiers; it
uses a training set, S = {(x
i
, y
i
) : i = 1, . . . , m}, where x
i
is the
characteristic vector in R
n
, y
i
{1, 2} represents the class, in
this case 1 for open eyes and 2 for closed eyes, and m is the
number of elements of S. From a training set, a hyperplane is
built that permits classication between two dierent classes
and minimizes the empirical risk function [30].
Mathematically, SVM consists in nding the best solu-
tion to the following optimization problem:
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 10 20 30 40 50 60 70 80 90 100
Time
P
r
o
b
a
b
i
l
i
t
y
Right eye
Left eye
Figure 15: Estimated value of the a posteriori density of the eye
center in a 100-frame for right and left eyes; the eyes are detected in
the fourth frame.
SVM Open or closed?
Figure 16: SVM scheme for eye state verication.
min
f () =
1
2
T
Q e
T
s.t. 0
i
C, i = 1, . . . , m
y
T
= 0,
(12)
where e is an m by the 1 vector, C is an upper bound, Q is
an m by m matrix with Q
i j
= y
i
y
j
K(x
i
, x
j
), and K(x
i
, x
j
)
is the kernel function. By solving the above quadratic
programming problem, the SVM tries to maximize the
margin between the data points in the two classes and to min-
imize the training errors simultaneously. Figure 17 depicts
input space mapping to a high-dimensional feature space
through a nonlinear transformation and its maximization
process.
4.4.2. Eye Characteristic Extraction Using a Gabor Filter.
The Gabor lter was used by Daugman for image analysis,
changing the orientation, and scale [18] where these are
Nonlinear
transformation
Hiperplane
Feature space Input space
Hiperplane
Class 1
Class 2
Figure 17: SVM representation.
multiscale and multiorientation kernels. They can be dened
by the complex function
g
_
x, y, ,
_
= exp
_
x
2
+ y
2
2
_
exp
_
i2
_
x cos
_
_
+ y sin
_
___
,
(13)
where and are the scale and orientation parameters and
is the standard deviation of the Gaussian kernel which
depends on the spatial frequency to be measured, that is,
. The response of the Gabor lter to an image is obtained
from a 2D convolution operation. Letting I(x, y) represent
the image and G(x, y, , ) denote the response of a Gabor
lter with scale and orientation to an image at point (x, y)
on the image plane by G() is obtained using
G
_
x, y, ,
_
=
__
I
_
p, q
_
g
_
x p, y q, ,
_
dp dq. (14)
Various combinations of scales and orientations are more
robust for the classication between open and closed eyes.
Three scales and four orientations have been used to generate
Figure 18; these are {1, 2, 3} and {0, /4, /2, 3/4} which
were obtained experimentally over an image size 30 by 20.
Once the response of the Gabor lter is obtained, the
eye characteristic vector is extracted using a subwindow
procedure described by Chen and Kubo [31] and denoted
by d R
360
. This vector is computed using (15) for each
subwindow of size 5 by 6. Figure 19 shows the subwindow
diagram
d
,
i
=
1
30
_
y=1:5
_
x=1:6
G
_
x, y, ,
_
i = 1, . . . , 20. (15)
To perform this analysis a training set has been built
which consists of both open and closed eyes. The images
come from diverse sources, under dierent illumination
conditions and are from dierent races. A further important
aspect of this eye database is that it contains images of
dierent eye colors, that is, blue, black, and green. Figure 20
shows several examples of this database.
Previous to SVM training, it is crucial to preprocess each
image where this procedure involves histogram equalization,
Table 3: Result of eye state analysis.
Driver Total frames Eyes open Eyes closed Correct rate
D1 960 690/700 258/260 98.90%
D2 900 520/560 339/340 96.27%
D3 500 388/400 99/100 98.00%
D4 330 150/170 152/160 91.61%
D5 1400 891/980 401/420 93.19%
ltering using a median lter, followed by the sharpen lter.
The median lter is used to reduce image noise, and the
sharpen lter enhances the borders.
The main objective of the SVM training is to obtain the
best parameters and the best kernel that minimizes (5). After
several SVM training experiments, it was decided to use the
RBF kernel, that is, K(x
i
, x
j
) is exp(x
i
x
j
2
), C = 30,
and = 0.0128, where these parameters achieve a high
training classication rate of close to 93%.
Table 3 presents results using this method computed for
several sequences of drivers. It demonstrates a high correct
rate of classications.
4.5. Drowsiness Index. The eye-blinking frequency is an
indicator that allows the level of driver drowsiness (fatigue)
to be measured. As in the works of Horng et al. [12]
and Dong and Wu [3], if for ve consecutive frames or
during 0.25 seconds the eye is identied as being closed the
system issues an alarm cue, PERCLOS [7], which is also
implemented in this system.
Figure 21 presents an instantaneous result for this system
on a drivers image whereas Figure 22 pictures the drowsiness
evolution index graph for a drivers drowsiness sequence.
Additionally, other examples are show in Figure 25.
4.6. Distraction. It is estimated that 20% of trac accidents
are caused by driver distraction [1]. To detect this char-
acteristic the drivers face should be studied because the
pose of the face contains information about ones attention,
gaze, and level of fatigue [4]. To verify driver distraction the
following procedure has been implemented.
(a) (b)
Figure 18: Gabor lter for = {0, 1, 2} and = {0, /4, /2, 3/4}.
Figure 19: Subwindow images from the Gabor lter.
Figure 20: Examples from the eye database for daylight illumina-
tion.
4.6.1. Face Orientation. The drivers face orientation is
estimated using the eye position, with
= tan
1
_
x
y
_
, (16)
Open
Alert
Figure 21: Day system instantaneous result.
0
20
40
60
80
100
P
e
r
c
e
n
t
a
g
e
0 200 100 300 400 500 600 700 800 900
Time
(a)
0
20
40
60
80
100
P
e
r
c
e
n
t
a
g
e
0 200 100 300 400 500 600 700 800 900
Time
(b)
Figure 22: Drowsiness index graph for a 900-frame sequence of a
drowsy driver: (a) Perclos, (b) Horng-Dong and Wu index.
X
Y
Figure 23: Face orientation.

20
15
10
5
0
5
10
15
20
A
n
g
l
e
0 20 10 30 40 50 60 70 80 90 100
Time
Figure 24: Head-orientation monitoring over time in a 100-frame
sequence.
where x = x
2
x
1
, y = y
2
y
1
, (x
1
, y
1
), and (x
2
, y
2
)
correspond to the left and right eye positions, respectively.
Equation (17) provides the classication limits. Figure 23
depicts an example of the face orientation and in Figure 24 an
extended example is shown for the drivers face orientation
using information from the eyes for a sequence of images
Left if > 8
,
Front if || 8
,
Right if < 8
.
(17)
4.6.2. Head Tilt. The method described above presents
problems when a monocular camera is used, and so, to
overcome this drawback, this contribution has implemented
a head-tilt based on neural networks. Keeping in mind that
the driver face database is composed of face examples for
ve dierent orientations, the face is passed to the neural
network to determine its orientation, specically for the up
and down cases. If the system detects that the face position
is not looking straight on, an alarm cue is issued to alert the
driver of a danger situation.
5. Night SystemDesign
In this part of the work, the night system will be described,
where this is based on the algorithm scheme shown in
Figure 1(b). Note that it is composed of both software and
hardware platforms. The main dierence between this and
the previous system is in perception system.
5.1. Perception System. The perception system is used to
acquire visual information of the drivers head during
nocturnal driving. It consists of three parts: a vision system,
a synchronization system, and an illumination system; see
Figure 26.
The rst part is composed of a miniature CCD camera
where the IR lter has been removed. This camera generates
a composite video signal (PAL, phase alternating line). The
second part takes the signal from the camera and splits it
up into odd and even elds using the LM1881N video sync
separator. In the third and last part, the illumination system,
based on near-infrared (NIR) light (700900 nm), contains
two sets of NIR light-emitting diode (LED) rings (inner and
outer) which produce odd and even image elds where the
pupil eect is highlighted [3, 4]. The inner led ring surrounds
the camera, while the outer leds are placed over a ruler in a
symmetric position around the camera. Figure 27(a) shows
this system, and Figure 27(b) provides an example of an
illuminated drivers face from the NIR illuminator.
Each frame is deinterlaced in both odd and even elds
which contain the dark and bright pupil images, separately.
Hence, the height of the odd and even image elds is a
medium of the original image; this procedure can be seen in
Figure 27(c): the top photograph is the even image eld, and
the bottom is the odd image eld. Even an odd image will be
used later on for eye detection.
5.2. Eye Detection. The bright eect pupil is the main
principle behind locating the position of the eye. To do this
sense, three images have been generated from the initial
driver image, these are the dierence image (I
D
), the edge
image (I
G
), and the bright part of the fast radial symmetry
transform (FRST) image [32] (I
F
).
The rst image is computed using the absolute value of
the pixel dierence between the even and odd image elds;
see (18). In this image most of the background and external
light illumination has been removed, and the pupils appear
as the brightest part
I
D
_
x, y
_
=
I
O
_
x, y
_
I
E
_
x, y
_
. (18)
The second and third images are obtained using the Sobel
Operator and the bright part of the FRST [32] over the
dierence image, respectively; these may be seen in Figures
28(a), 28(b), and 28(c)
I
G
=
_
I S
X
2
+ I S
Y
2
, (19)
where S
x
and S
y
are given by (5).
Most researchers only make use of the dierence image
for pupil detection; however, in real driving conditions, this
Figure 25: Dierent stages of the proposed algorithm during daytime conditions and for several time instants, driving conditions and
dierent drivers.
Synchronization system
(split up in odd and even elds, LM1881N)
Vision system
(CCD camera)
Illumination system
(outer and inner rings of leds)
Figure 26: Perception system schema.
image deteriorates due to external illumination, vibrations,
and so forth and is also very sensitive to lighting conditions.
In such circumstances, it is necessary to incorporate more
robust information to improve the detection step. There-
fore, in this paper, the edge and FRST images have been
implemented to obtain enhanced results considering the
aforementioned drawbacks.
Once all the images used to detect the eyes have been
specied, the next step is to compute a binary threshold
for the dierence, edge, and FRST images. In the rst of
these, the threshold is obtained from a systematic analysis
of its histogram, where two groups are formed. In the
second case, the histogram is modelled using a Gamma
distribution function where the 90% cumulative interval
provides the threshold. Finally, in the third image, the
maximum histogram value produces the required threshold
level. This yields three binary images consisting of binary
blobs that may contain a pupil.
The pupils are detected by searching within the entire
image for the location of two blobs that satisfy a particular
size, shape, and distance constraints. To remove false blobs
an unsupervised classier has been implemented, in this
case, the SMV classier, which is based on statistical learning
theory and pattern classiers. It uses a training set, S =
{(x
i
, y
i
) : i = 1, . . . , m}, where x
i
is the characteristic
vector in R
n
, y
i
{1, 2} represents the class, in this case
1 for open eyes and 2 for noneyes, and m is the number
of elements of S [28]. Before training the SVM, it is crucial
to preprocess each image, where this procedure involves
histogram equalization, ltering using the median lter,
followed by the sharpen lter. The median lter is used to
reduce noise in the image, and the sharpen lter is used
to enhance the borders. After this procedure a Gabor lter
[31] is computed on each image of the eye database. Some
combinations of scales and orientations are more robust
for the classication between eyes and noneyes. One scale
and four orientations have been used in this research; these
are {0} and {0, /4, /2, 3/4} which have been obtained
experimentally from an image size of 30 by 20 (in pixels). To
perform this task, a training set has been built which consists
of open eyes and noneyes, where an example is shown in
Figure 29. Once the response of the Gabor lter is obtained,
(a) (b) (c)
Figure 27: Perception system using infrared illumination (a) vision system, (b) drivers image, and (c) odd (top) and even (bottom) image
elds.
(a) (b) (c)
(d) (e) (f)
Figure 28: Dierence (a), edge (b), and FRST (c) and their binary images (d), (e) and (f), respectively.
the eye characteristic vector is extracted and denoted by d
R
600
.
After several SVM training experiments, it was decided
to use the RBF kernel, that is, K(x
i
, x
j
) is exp(x
i

x
j
2
), C = 100, and = 0.128; these parameters achieve a
high training classication rate of about 93%.
This exhaustive search detects the pair of eyes; this is
followed by an ellipse tting which is applied to each pupil,
and the center of the resulting ellipse is the position of the
detected pupil. This process is presented in Figure 30(b).
These will be used later to initialize the eye tracker.
5.3. Face Detection. Once the eyes have been located, the
system continues with driver face detection. To perform this
task, a human face model has been developed considering the
pupils position and face anthropometric properties [19]. Let
p
1
(x, y) and p
2
(x, y) be the center position of the right and
left eye, respectively, and d
RL
their distance (in pixels); the
area of the face is obtained from the following equations:
d
RL
=
_
(x
1
x
2
)
2
+
_
y
1
y
2
_
2
,
= tan
1
_
y
2
y
1
x
2
x
1
_
,
R
1
= 1.5d
RL
, (20)
R
2
= d
RL
,
p
c
_
x, y
_
=
_
x
1
+ 0.5d
RL
, y
1
+ 0.3d
RL
_
,
where p
c
(x, y) is the centre of the face, and R
1
and R
2
are the
axes of the face ellipse. Figure 30 depicts this model and its
result.
5.4. Tracking. The tracking process has been developed using
the Condensation algorithm for face and eye tracking.
5.4.1. Face Tracking. This system uses the neural networks
[33] and models and parameters proposed previously for
face tracking. Now that the elements for tracking have been
(1) (2)
(3)
(a)
(1) (2)
(3)
(b)
Figure 29: Gabor lter response for = {0} and = {0, /4, /2, 3/4}.
d
RL
p
1
(x, y)
p
c
(x, y)
p
2
(x, y)
R
2
R
1
Minor radius
Major radius
(a) (b)
Figure 30: (a) Human face model using the eye position, (b) face and eye detection.
identied, only the backpropagation neural network training
is developed in this section. Before training, a preprocessing
step that consists of using a Gabor lter [18] with one scale
and two orientations has been implemented. After this, the
characteristic vector which consists of gray-level values of
pixels coming from the face image is extracted. The rate of
classication subsequent to the training is more than 92%.
Figure 31 shows several face examples, and Table 4 presents
experimental results. In this table, the true position has been
retrieved manually.
The density function of the initial state is p(x
0
) =
N(z
0
,
0
), where z
0
is computed using the previous face
detection method, and
0
is given in [4]. Figure 32 shows the
a posterior density function of the face center tracking.
5.4.2. Eye Tracking. For this task, the state of the eye is
characterized by its position and velocity. These parameters
are also described in the diurnal system. To evaluate the
probability observation density, a triangular density function
based on the value of the dierence image has been used
(Figure 33). This function takes into account the gray-level
value for the intensity of the illumination system.
CA is initialized when the eyes are detected from the
method described in the previous section plus a white noise.
In Table 4 the eye tracking results are presented, which have
been carried out in several sequences of images. Again,
the true position has been retrieved manually. Additionally,
Figure 34 shows the a posterior density function of both eyes.
To evaluate the probability observation density, a trian-
gular density function based on a value from the dierence
image has been used. CA is initialized when the eyes are
detected using the method described in the previous section
plus a white noise. Table 4 shows the eye tracking results
which have been obtained from several image sequences.
5.5. Eye State Detection and Drowsiness Index. To identify
drowsiness from an eye analysis, knowledge of the eyes state
is required, that is, open or closed, in time and to develop
an analysis over large periods of time, that is, to measure the
time spent in each state. Classication of the open and closed
state is complex due to changes in the shape of the eye, the
changing position, and face rotations, as well as variations
in twinkling and illumination, and so forth. All of these
factors make it dicult to reliably analyze the eyes. However,
when using the edge and FRST images, the eye state may be
computed satisfactorily.
The PERCLOS [7] has been implemented in this system.
Figure 35 presents an instantaneous result of this system
obtained from a drivers image, and in Figure 36 the
evolution of the drowsiness index graph for a sequence of
driver drowsiness is presented.
5.6. Distraction. This method is similar to the previous case,
once the face is continuously located in time; a neural
network is used to determine its orientation and to verify the
drivers level of distraction. If the system detects that the face
(a) (b) (c)
(d) (e) (f)
Figure 31: Examples of the face database and its Gabor lter response.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0 50 100 150 200 250 300 350
Time
P
r
o
b
a
b
i
l
i
t
y
Figure 32: Estimated value of the a posteriori density of the face-
center in a 350-frame sequence using the proposed tracker; the face
is detected in the fourth frame.
0 255
1
(255,1)
Figure 33: Triangular density function.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0 50 100 150 200 250 300 350
Time
Right eye
Left eye
P
r
o
b
a
b
i
l
i
t
y
Figure 34: Estimated value of the a posteriori density of the eye
center in a 350-frame for right and left eyes, the eyes are detected in
the fourth frame.
position is not facing forward, an alarm cue is issued to alert
the driver of a danger situation.
6. Conclusions
In this paper, a research project to develop a nonintrusive
and autonomous driver drowsiness system based on Com-
puter Vision and Articial Intelligence has been presented.
This system uses advanced technologies which analyze and
Close
Awake
Drowsiness
Time: 9.011 (seg)
Closed percentage: 39.286
Open percentage: 60.714
Perclos: 30
Closed frames: 33
Open frames: 51
Total frames: 64
Figure 35: System instantaneous result.
Table 4: Results of face and eye tracking and eye state analysis.
Driver Total frames
Face tracking Eye tracking Eye state
Tracking failure Correct rate Tracking failure Correct rate Eyes open Eyes closed Correct rate
D1 800 40 95.00% 14 98.25% 690/700 97/100 97.78%
D2 646 29 95.51% 29 95.51% 500/530 100/116 90.27%
D3 600 20 96.67% 24 96.00% 345/244 206/226 91.69%
0
20
40
60
80
1 43 85 127 169 211 253 295 337 379 421
Time
P
e
r
c
e
n
t
a
g
e
Figure 36: Drowsiness index graph in a 341-frame sequence of a
drowsy driver.
monitor the state of the drivers eye in real-time and for
real driving conditions; this is driving conditions for both
daytime and nocturnal situations.
In the rst case, based on the results presented in Tables
1, 2, and 3, the algorithm proposed for eye detection,
face tracking, and eye tracking is shown to be robust and
accurate for varying light, external illumination interference,
vibrations, changing backgrounds, and facial orientations. In
the second case, and as presented in the results of Table 4, the
system is also observed to provide agreeable results.
To acquire the data required to develop and test the
algorithms presented in this paper, several drivers have been
recruited and were exposed to a wide variety of dicult
situations commonly encountered on roadways, for both
daytime and nocturnal conditions. This guarantees and
conrms that the experiments presented here are proven to
be robust and ecient for real trac scenarios. The images
were taken using two cameras within the IVVI (Intelligent
Vehicle based on Visual Information) vehicle (Figure 27(a)):
a pin-hole analog camera connected to a frame-grabber for
the nocturnal illumination and a re-wire camera for the
diurnal use. Besides that, the hardware processes 4-5 frames
(a)
PC1 PC2
PC3
(b)
Figure 37: (a) IVVI vehicle, (b) processing system.
per second using an Intel Pentium D, with 3.2 GHz, 2 GB.
RAM memory and MS Windows XP.
IVVI is an experimental platform used to develop the
driver assistance systems for real-life driving conditions. The
most up to date version is a Nissan-Note car; see Figure 37(a).
This vehicle is equipped with a processing system, which
processes the information coming from the cameras. The
processing system is composed of three personal computers
(Figure 37(b)).
For future work, the objective will be to reduce the
percentage error, that is, reduce the amount of false alarms; to
achieve this, additional experiments will be developed, using
additional drivers and incorporating new analysis modules,
for example, facial expressions.
Acknowledgments
This paper was supported in part by the Spanish Government
through the CICYT projects VISVIA (Grant TRA2007-
67786-C02-02) and POCIMA (Grant TRA2007-67374-C02-
01).
References
[1] T. Brandt, R. Stemmer, B. Mertsching, and A. Rakotonirainy,
Aordable visual driver monitoring system for fatigue and
monotony, in Proceedings of the IEEE International Conference
on Systems, Man and Cybernetics (SMC 04), vol. 7, pp. 6451
6456, October 2004.
[2] Z. Tian and H. Qin, Real-time drivers eye state detection, in
Proceedings of the IEEE International Conference on Vehicular
Electronics and Safety, pp. 285289, October 2005.
[3] W. Dong and X. Wu, Driver fatigue detection based on the
distance of eyelid, in Proceedings of the IEEE International
Workshop on VLSI Design and Video Technology (IWVDVT
05), pp. 397400, Suzhou-China, May 2005.
[4] Q. Ji and X. Yang, Real-time eye, gaze, and face pose tracking
for monitoring driver vigilance, Real-Time Imaging, vol. 8, no.
5, pp. 357377, 2002.
[5] L. M. Bergasa, J. Nuevo, M. A. Sotelo, and M. V azquez, Real-
time system for monitoring driver vigilance, in Proceedings of
the IEEE Intelligent Vehicles Symposium, pp. 7883, June 2004.
[6] L. Fletcher, L. Petersson, and A. Zelinsky, Driver assistance
systems based on vision in and out of vehicles, in Proceedings
of the IEEE Symposium on Intelligent Vehicles, pp. 322327,
2003.
[7] NHTSA, Evaluation of techniques for ocular measurement as
an index of fatigue and the basis for alertness management,
Tech. Rep. DOT HS 808762, National Highway Trac Safety
Administration, Washington, DC, USA, 1998.
[8] L. Hagenmeyer, Development of a multimodal, univer-
sal human-machine-interface for hypovigilance-management-
systems, Ph.D. thesis, University of Stuttgart, Stuttgart, Ger-
many, 2007.
[9] G. Longhurst, Understanding Driver Visual Behaviour, Seeing
Machine, Canberra, Australia.
[10] J. M. Armingol, A. de la Escalera, C. Hilario et al., IVVI:
intelligent vehicle based on visual information, Robotics and
Autonomous Systems, vol. 55, no. 12, pp. 904916, 2007.
[11] T. DOrazio, M. Leo, and A. Distante, Eye detection in faces
images for a driver vigilante system, in Proceedings of the IEEE
Intelligent Vehicles Symposium, pp. 1417, University of Parma,
Parma, Italy, June 2004.
[12] W.-B. Horng, C.-Y. Chen, Y. Chang, and C.-H. Fan, Driver
fatigue detection based on eye tracking and dynamic template
matching, in Proceedings of the IEEE International Conference
on Networking, Sensing and Control, pp. 712, March 2004.
[13] P. Viola and M. Jones, Rapid object detection using a
boosted cascade of simple features, in Proceedings of the IEEE
Computer Society Conference on Computer Vision and Pattern
Recognition, pp. 511518, December 2001.
[14] A. Branzan, B. Widsten, T. Wang, J. Lan, and J. Mah, A
computer vision-based system for real-time detection of sleep
onset in fatigued drivers, in Proceedings of the IEEE Intelligent
Vehicles Symposium (IV 08), pp. 2530, June 2008.
[15] Q. Ji, Z. Zhu, and P. Lan, Real-time nonintrusive monitoring
and prediction of driver fatigue, IEEE Transactions on Vehicu-
lar Technology, vol. 53, no. 4, pp. 10521068, 2004.
[16] J. G. Brookshear, Theory of Computation: Formal Languages,
Automata and Complexity, Addison Wesley Iberoamericana,
1993.
[17] R. Grace, Drowsy driver monitor and warning system,
in Proceedings of Driving Assessment International Driving
Symposium on Human Factors in Driver Assessment, Training,
and Vehicle Design, 2001.
[18] J. G. Daugman, Uncertainty relation for resolution in
space, spatial frequency, and orientation optimized by two-
dimensional visual cortical lters, Journal of the Optical
Society of America, vol. 2, no. 7, pp. 11601169, 1985.
[19] P. Gejgus and M. Sparka, Face Tracking in Color Video
Sequences, The Association for Computing Machinery, 2003.
[20] N. Otsu, A threshold selection method from gray-level his-
tograms, IEEE Transactions on Systems, Man and Cybernetics,
vol. 9, pp. 6266, 1979.
[21] I. Jafar and H. Ying, A new method for image contrast
enhancement based on automatic specication of local
histograms, International Journal of Computer Science and
Network Security, vol. 7, no. 7, 2007.
[22] Y. Wu, H. Liu, and H. Zha, Anewmethod of detecting human
eyelids based on deformable templates, in Proceedings of the
IEEE International Conference on Systems, Man and Cybernetics
(SMC 04), pp. 604609, October 2004.
[23] G. J. McLachlan, The EMAlgorithmand Extensions, John Wiley
& Sons, New York, NY, USA, 1997.
[24] M. Isard and A. Blake, Condensation: conditional density
propagation for visual tracking, International Journal of
Computer Vision, vol. 29, no. 1, pp. 528, 1998.
[25] M. A. Isard, Visual motion analysis by probabilistic propagation
of conditional density, Ph.D. thesis, Oxford University, Oxford,
UK, 1998.
[26] J. R. Parker, Practical Computer Vision Using C, John Wiley &
Sons, New York, NY, USA, 1994.
[27] J. Satake and T. Shakunaga, Multiple target tracking by
appearance-based condensation tracker using structure infor-
mation, in Proceedings of the 17th International Conference on
Pattern Recognition (ICPR 04), vol. 3, pp. 294297, August
2004.
[28] N. Cristianini and J. Shawe-Taylor, An Introduction to Support
Vector Machines and Other Kernel-Based Learning Methods,
Cambridge University Press, Cambridge, UK, 2000.
[29] C. Chang and C. Lin, LIBSVM: a library for support vector
machine, 2001, http://www.csie.ntu.edu.tw/cjlin/libsvm.
[30] I. Guyon, S. Gunn, M. Nikravesh, and L. A. Zadeh, Feature
Extraction: Foundations and Applications, Springer, Berlin,
Germany, 2006.
[31] Y.-W. Chen and K. Kubo, A robust eye detection and
tracking technique using gabor lters, in Proceedings of the 3rd
International Conference on Intelligent Information Hiding and
Multimedia Signal Processing (IIHMSP 07), vol. 1, pp. 109
112, November 2007.
[32] G. Loy and A. Zelinsky, Fast radial symmetry for detecting
points of interest, IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 25, no. 8, pp. 959973, 2003.
[33] C. G. Looney, Pattern Recognition Using Neural Networks,
Theory and Algorithms for Engineers and Scientists, Oxford
University Press, Oxford, UK, 1997.
doi:10.1155/2010/724035
Research Article
Multiobjective Reinforcement Learning for
Trafc Signal Control Using Vehicular Ad Hoc Network
Duan Houli, Li Zhiheng, and Zhang Yi
Department of Automation, Tsinghua University, Beijing 100084, China
Correspondence should be addressed to Duan Houli, duanhouli@gmail.com
Received 1 December 2009; Accepted 5 September 2010
Copyright 2010 Duan Houli et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
We propose a new multiobjective control algorithm based on reinforcement learning for urban trac signal control, named multi-
RL. A multiagent structure is used to describe the trac system. A vehicular ad hoc network is used for the data exchange among
agents. Areinforcement learning algorithmis applied to predict the overall value of the optimization objective given vehicles states.
The policy which minimizes the cumulative value of the optimization objective is regarded as the optimal one. In order to make
the method adaptive to various trac conditions, we also introduce a multiobjective control scheme in which the optimization
objective is selected adaptively to real-time trac states. The optimization objectives include the vehicle stops, the average waiting
time, and the maximum queue length of the next intersection. In addition, we also accommodate a priority control to the buses
and the emergency vehicles through our model. The simulation results indicated that our algorithmcould performmore eciently
than traditional trac light control methods.
1. Introduction
Increasing trac congestion over the road networks makes
the development of more intelligent and ecient trac
control systems an urgent and important requirement. How-
ever, trac systems are typically complex large-scale systems
consisting of a great number of interacting participants. It
is very dicult to use traditional control algorithms to get
satised control eect. Thus, various intelligent algorithms
have been used in attempts to build an ecient trac control
system, such as fuzzy control technologies [1, 2], articial
neural networks [3, 4], and genetic algorithms [5, 6], which
greatly improve the eciency of urban trac signal control
systems.
Reinforcement learning is a category of machine learning
algorithms including Q learning, temporal dierence, and
SARSA algorithm [79]. Reinforcement learning is to learn
the optimal policy by a trial-and-error process including
perceiving states from the environment, choosing an action
according to current states and receiving rewards from the
environment. The policy which maximizes the expected
long-term cumulative reward is considered as the optimal
one. Reinforcement learning is a self-learning algorithm
which does not need an explicit model of the environment.
Thus, it can be applied in trac signal control eectively
to respond to the constant changes of trac ow and
outperform traditional trac control algorithms. Thorpe
studied reinforcement learning for trac light control in
1997. He used a neural network to predict the waiting
time for all cars standing at the intersection and selected
the best control policy using the SARSA algorithm [10].
Abdulhai et al. presented a basic framework of applying
Q-learning to trac signal control and got encouraging
results while applying it to an isolated intersection [11].
Mikami and Kakazu combined the evolutionary algorithm
and reinforcement learning for coordination trac signal
control [12]. However, the above methods use trac-light-
based value functions, which means that the state space is
too large to handle. Therefore, these methods suer from
the dimension curse and achieve limited success when
applied to large-scale road networks. Wiering et al. utilized
a car-based value function to solve this problem [13, 14].
They predicted each cars total expected waiting time until
it arrived its destination given possible choices of related
trac lights using reinforcement learning, and chose the
action which minimized the summed waiting time of all
cars in the network. This method eectively reduces the
state space and thus can be applied to large-network control.
Experiments in a network with 12 edge nodes and 16
junctions proved the eectiveness of this method.
However, Wierings method uses the total waiting time
as the optimization goal which is mainly suitable for the
medium trac condition. In practical trac systems, we
should consider dierent optimization objectives adaptive to
dierent trac situations, called the multiobjective control
scheme in this paper. Under the free trac condition, the
average vehicle speed is high and the average waiting time
is short, so the waiting time is not the focal point, while
the vehicle stops will increase the vehicle emission and oil
consumption. Therefore, we should try to minimize the
overall vehicle stops in the network. Under the medium
trac condition, the overall waiting time is regarded as the
optimization goal because most drivers want to arrive at
their destinations as soon as possible. Under the congested
trac situation, queue spillovers must be avoided to keep
the network from large-scale congestion, thus, the queue
length must be regarded as the control goal [15]. Since the
multiobjective control scheme can adapt to various trac
conditions and make a more intelligent control system, we
propose a multiobjective control strategy based on Wierings
model. In our model, data exchanges among vehicles and
roadside equipments are necessary. Thus, a vehicular ad hoc
network is utilized to build a wireless trac information
system.
This paper is organized as follows: in Section 2, we will
introduce how to model the road network with an agent-
based structure; Section 3 describes how to exchange trac
data using the ad hoc network; in Section 4, a multiagent
trac control strategy using reinforcement learning is pro-
posed; in Section 5, the proposed method is applied to a road
network with 7 intersections to prove its eectiveness; nally,
in Section 6, we draw the conclusion of this paper.
2. Agent-Based Model of Trafc System
We use an agent-based model to describe the practical trac
system. Vehicles and trac signal controllers in the road
network are regarded as two types of agents. Data will be
exchanged among these agents. A typical road network is
built based on Wierings model [14] as shown in Figure 1.
There are six possible settings for each trac controller
to prevent accidents: two trac lights from opposing
directions allow cars to go straight ahead or to turn right
(2 possibilities), two trac lights in the same direction of
the intersection allow the cars from there to go straight
ahead, turn right, or turn left (4 possibilities). Road lanes
are discretized into a number of cells at each trac light.
The capacity of each road lane is determined according
to its practical length. At each time step, new cars with
particular destinations are generated and enter the network
from outside. After new cars have been added, trac light
decisions are made and each car moves to the subsequent
cell if it is not occupied or the cars predecessor is moved
forward. All vehicles are assumed to have the same speed
in this system. Thus, each car is at a specic trac node
(node), a direction at the node (dir), a position in the queue
(place), and has a particular destination (des). Thus, we
can use [node, dir, place, des] ([n, d, p, des] for short) to
denote the state of each vehicle [13]. Vehicles follow the
shortest path through the road network to their destinations.
As mentioned before, a multiobjective control scheme is
adopted in this method. The optimization objectives include
the total waiting time, vehicle stops, and the queue length,
which will be chosen adaptively to the trac condition. We
use Q([n, d, p, des], action) to denote the total expected value
of the optimization objective for each car until it arrives at
the destination given its current node, direction, place and
the decision of the light. The optimal action of a node j is
determined by the following formulation:
A
opt
j
= arg max
Aj
_
i Aj
_
(n,d,p,des) queuei
Q
__
n, d, p, des
_
, red
_
Q
__
n, d, p, des
_
, green
_
.
(1)
It should be noticed that Q([n, d, p, des], action) here does
not only refer to the total waiting time but also refer to
vehicle stops or queue lengths, according to the real-time
trac states. This is the most important dierence between
our model and Wierings model, which will be explained in
detail in Section 4.
3. Trafc Information Exchange SystemUsing
Vehicular Ad Hoc Network
We need to exchange a lot of information during the signal
control process. Thus, a wireless trac information exchange
system based on a vehicular ad hoc network is built to
exchange data among the vehicles and signal controllers.
An illustration of such information exchange system is
showed in Figure 2. It is assumed that all vehicles in the
network are intelligent ones equipped with Vehicular Ad
Hoc Network communication devices, so that they have
the ability of communicating with other vehicles and the
roadside controllers. Thus, all necessary information can be
collected through the intercommunication of vehicles and
controllers. The data to be collected include the followings:
(a) trac ow through each intersection within each
time step;
(b) queue length at each trac light within each time
step;
(c) type of each vehicle (car, bus, or emergent vehicle);
(d) destination of each vehicle;
(e) node where each vehicle stands at;
(f) direction each vehicle moving towards;
Figure 1: Agent-based trac model illustration.
Wireless
network
Controller
Trac control
center
Figure 2: Illustration of trac information exchange system.
(g) position in the queue where each vehicle stands at;
(h) total waiting time each vehicle used to pass through
the network;
(i) total number of stops each vehicle used to pass
through the network.
4. Multiobjective Control Algorithm Based on
Reinforcement Learning (Multi-RL)
We extend Wierings algorithm to a multiobjective scheme
by selecting the optimization objective according to the real-
time trac condition. In addition, it is assumed that some
special vehicles such as buses and ambulances need a priority
control, and thus they should be considered separately.
The multiobjective control algorithm considers three
types of trac conditions as follows. The method to estimate
trac conditions should be dened carefully according to the
actual situation of the road network.
4.1. Free Trac Condition. Under this condition, we aim to
minimize the number of stops, in other words, we expect to
have the vehicles pass through the network with the fewest
stops. Thus, the cumulative number of stops is selected as
the optimization objective.
The number of stops will increase when a vehicle
moving to a green light at current time step meets a red
light at the next time step. Therefore, we denote Q([node,
dir, pos, des], L) as the expected cumulative number of stops
while V([node, dir, pos, des]) denotes the number of stops
(without knowing the trac light decision) for a car at
[node, dir, pos] until it reaches its destination. The iterative
formulation of Q([node, dir, pos, des], L) is shown as follows:
Q
__
node, dir, pos, des
_
, L
_
=
_
(node
, dir
, pos
, L, L
)
P
_
L
|
_
node, dir, pos, des
_
, L,
_
node
, dir
, pos
, des
__
_
R
__
node, dir, pos, des
_
,
_
node
, dir
, pos
, des
__
+V
__
node
, dir
, pos
, des
___
,
V
__
node, dir, pos, des
__
=
_
L
P
_
L |
_
node, dir, pos, des
__
Q
__
node, dir, pos, des
_
, L
_
,
(2)
where [node
, dir
, pos
, des] means the state of a vehicle at

the next time step; L is the action of the trac light at
the current time step, while L
is the action of the trac

light at the next time step. P(L
| [node, dir, pos, des], L,

[node
, dir
, pos
, des]) gives the probability that the trac

light turns L
at the next time step given the current state

and the next state of this vehicle; R([node, dir, pos, des],
[node
, dir
, pos
, des]) is a reward function as follows: if L =

Green, L
= Red, which means the vehicle moving to a green

light at the current time step meets a red light at the next time
step, then the number of vehicle stops will increase, R = 1;
otherwise, R = 0; is the discount factor (0 < < 1) which
ensures that the Q-values are bounded. The probability that
a trac light turns red is calculated as follows:
P
_
L
|
_
node, dir, pos, des
_
, L,
_
node
, dir
, pos
, des
__
=
C
__
node, dir, pos, des
_
, L,
_
node
, dir
, pos
, des
_
, L
_
C
__
node, dir, pos, des
_
, L,
_
node
, dir
, pos
, des
__ ,
(3)
where C([node, dir, pos, des], L, [node
, dir
, pos
, des])
means the number of times a car in the state of [node, dir,
pos, des] transiting to the state of [node
, dir
, pos
, des]
and the transiting light is L, C([node, dir, pos, des], L,
[node
, dir
, pos
, des], L
) is the number of times the light

turns L
after such a transiting procedure.

4.2. Medium Trac Condition. Under this medium trac
condition, we focus on the overall waiting time of vehi-
cles, which is the same as in Wierings model [13, 14].
Q([node, dir, pos, des], action) is used to denote the total
waiting time before all trac lights for each car until it
arrives at the destination given its current state and the
action of the light. V([node, dir, pos, des]) denotes the total
waiting time (without knowing the trac light decision)
for a car at [node, dir, pos]until it reaches its destination.
Q([node, dir, pos, des], action) and V([node, dir, pos, des])
are iteratively updated as follows:
V
__
node, dir, pos, des
__
=
_
L
P
_
L |
_
node, dir, pos, des
__
Q
__
node, dir, pos, des
_
, L
_
,
(4)
Q
__
node, dir, pos, des
_
, L
_
=
_
(node
, dir
, pos
)
P
__
node, dir, pos, des
_
, L,
_
node
, dir
, pos
, des
__
_
R
__
node, dir, pos, des
_
,
_
node
, dir
, pos
, des
__
+V
__
node
, dir
, pos
, des
___
,
(5)
where L is the trac light state (red or green), P(L |
[node, dir, pos, des]) is calculated in the same way as (3),
R([node, dir, pos, des], [node
, dir
, pos
, des]) is dened as
follows: if a car stays at the same place, then R = 1, otherwise,
R = 0 (the car can move forward).
4.3. Congested Trac Condition. Under the congested trac
condition, we must do our best to avoid the queue spillovers,
which will seriously degrade the trac control eect and
probably cause large-scale trac congestion [15]. Therefore,
the queue length is taken into consideration when we design
the Q learning procedure. Denote the maximum queue
length at the next trac light tl
as K
tl
, shortly written as
K. When the trac light is red, no vehicle can pass through
to the next light. Thus, the equations at a red light do not
change, we focus on the function when light is green. Then
(5) can be rewritten as follows:
Q
__
node, dir, pos, des
_
, Green
_
=
_
(node
, dir
, pos
)
P
__
node, dir, pos, des
_
, Green,
_
node
, dir
, pos
, des
__
_
R
__
node, dir, pos, des
_
,
_
node
, dir
, pos
, des
__
+ R
__
node, dir, pos, des
_
,
_
node
, dir
, pos
, des
__
+V
__
node
, dir
, pos
, des
___
,
(6)
Q
__
node, dir, pos, des
_
, Red
_
=
_
(node
, dir
, pos
)
P
__
node, dir, pos, des
_
, Red,
_
node
, dir
, pos
, des
__
_
R
__
node, dir, pos, des
_
,
_
node
, dir
, pos
, des
__
+V
__
node
, dir
, pos
, des
___
,
(7)
where Q([node, dir, pos, des], L) and V([node, dir, pos, des])
have the same meanings as under the medium trac
condition. Compared (6) with (5), another reward function
R
([node, dir, pos, des], [node
, dir
, pos
, des]) is added to
indicate the inuence from trac condition at the next light.
, dir
, pos
, des]) is the reward

of vehicles waiting time while R
([node, dir, pos, des],

[node
, dir
, pos
, des]) indicates the reward from the queue

length increasing at the next trac light. The parameter is
an adjusting factor.
, dir
, pos
, des]) is dened
as follows: if a car stays at the same place, then R = 1,
otherwise, R = 0 (the car can move forward).
R
([node, dir, pos, des], [node
, dir
, pos
, des]) is dened
as follows: if a car passes through the current intersection to
the next trac light, which means that the queue length at
the next trac light will increase by 1 in a short time, then
R = 1, otherwise, R = 0.
Given the capacity of the lane of next trac light is L,
then the adjusting factor is determined by the queue length
K
tl
as follows. Note when queue spillovers happen, K
tl
will
be larger than L [15]
=
0, if K
tl
0.8L,
10
_
K
tl
L
0.8
_
, if 0.8L < K
tl
L,
2, if K
tl
> L.
(8)
Through the denition we can nd that will increase
sharply when the queue length approaches the capacity of
the lane, which means that queue spillovers would like
to happen. Thus, under such a situation, Q([node, dir,
pos, des], Green) will increase sharply and make the gain
of this policy decrease. Therefore, the green phase length
and the number of vehicles allowed to pass through will be
decreased until the queue at the next light has been dispersed.
The largest value of is set to 2 in this paper, but you can
adjust its value according to the practical trac condition.
4.4. Priority Control for Buses and Emergency Vehicles. When
buses or emergency vehicles (re trucks or ambulances)
enter the road network, they should have a priority to pass
through. It is necessary to realize the priority control of these
special vehicles with least disturbance to the regular trac
order. Thus, we revise (5) as follows. A priority factor
is added to describe the emergency degree of these special
vehicles, which needs to be determined separately by the
trac management department
Q
__
node, dir, pos, des
_
, L
_
=
_
(node
, dir
, pos
)
P
__
node, dir, pos, des
_
, L,
_
node
, dir
, pos
, des
__
_
R
__
node, dir, pos, des
_
,
_
node
, dir
, pos
, des
__
+V
__
node
, dir
, pos
, des
___
.
(9)
5. Case Studies
We have done some case studies to prove the eectiveness
of our model. Since it is very hard to apply a model to
the real trac system management, trac simulation is
chosen to do the case studies. Paramics V6.3 was selected
as the simulation platform because it is a professional trac
simulation tool which is recognized by trac engineers all
over the world. A practical road network within Beijing
Second Ring Road was modeled in Paramics as shown
in Figure 3. This is a network with 7 intersections (N1
N7) and 8 OD zones (Zone1Zone8). Intersections N1N7
correspond to the real intersections Xiaoweihutong, Dong-
dansantiao, Jingyuhutong, Dengshidongkou, Dengshikou,
Wangfujingbeikou, and Taiwanfandian.
N5
N4
N6
N3 N2
N1 N7
Zone1
Zone2
Zone3
Zone4
Zone5
Zone6 Zone7
Zone8
Figure 3: Sketch diagram of a practical road network in Beijing.
The simulation ran for 10000 time steps, the rst 4000
steps made up the learning process, and the latter 6000 steps
was used to collect the simulation results. Factor is set to
be 0.9 and is set to be 3. The lanes in the network are
divided into cells with length of 7.5 m. The capacity of the
lanes equals to the number of the cells.
We compared our method with the xed control, the
actuated control and also Wierings method. The setting of
xed control is as follows, the cycle is 2 minutes and the green
time is equally assigned to all phases. In the actuated control
strategy, the minimum green time is 10 s, the maximum
green time is 50 s, and the extension of green time is set to 4 s.
Parameters of Wierings method are the same as our model
under the medium trac condition.
We wanted to estimate the eectiveness of the mul-
tiobjective scheme, thus, we estimated the control eects
of these four algorithms under dierent trac conditions.
We changed the trac volume entering the network every
minute from 30 to 270 and estimated the average waiting
time, the number of stops, and maximum queue length of
these four methods.
In our model, when the trac volume entering the
network in a minute is less than 90, it is regarded as the
free trac; when the volume is larger than 90 but less than
180, it is regarded as the medium trac; when the trac
volume is larger than 180, it is regarded as the congested
trac condition.
5.1. Comparison of the Number of Stops. The comparison of
the number of stops with respect to the increasing of trac
volume is shown in Figure 4. Fixed means the xed control
strategy, actuated means the vehicle actuated method, RL
means the algorithm proposed by Wiering [13, 14], and
multi-RL means the model proposed in this paper.
It is obvious that when the trac volume is less than
90, which means that the trac state is free. The number
of stops under the multi-RL control is less than those under
other control strategies. This is because the multi-RL is
0 50 100 150 200 250 300
1
2
3
4
5
6
7
Trac volume
A
v
e
r
a
g
e
s
t
o
p
s
Fixed
Actuated
RL
Multi-RL
Figure 4: Control eects comparison estimated by average stops.
the only one that aims to minimize the number of stops.
However, with the increase of trac volume, the multi-RL
method changes its objective, and the actuated control gets
the minimum stops.
5.2. Comparison of the Average Waiting Time. The com-
parison of the average waiting time with respect to the
increasing of trac volume is shown in Figure 5. Since
the multi-RL is the same as the RL method under the
medium trac condition, they have almost the same average
waiting time in the middle. Under the free trac state,
the RL gets the minimum waiting time because this is its
optimization objective. It should be noticed the multi-RL
gets the minimum waiting time when the trac is congested.
This indicates that although the RL aims to minimize the
waiting time, the queue spillover which is not considered will
decrease the trac eciency and increase the waiting time.
5.3. Comparison of Maximum Queue Length. The compari-
son of the average waiting time with respect to the increasing
of trac volume is shown in Figure 6. The maximum queue
length exceeds 40 under the xed control, which indicates
that there must be some queue spillovers. This is taken into
consideration in the multi-RL, thus, we get a short queue
under the congested trac condition.
6. Conclusion
In this paper, a multiobjective control algorithm based on
reinforcement learning is proposed. The simulation results
indicate that the multi-RL gets the minimum stops under
the free trac, though not the minimum waiting time;
the multi-RL has almost the same performance with the
0 50 100 150 200 250 300
Trac volume
Fixed
Actuated
RL
Multi-RL
150
200
250
300
350
400
A
v
e
r
a
g
e
w
a
i
t
i
n
g
t
i
m
e
Figure 5: Control eects comparison estimated by average waiting
time.
0 50 100 150 200 250 300
Trac volume
Fixed
Actuated
RL
Multi-RL
M
a
x
i
m
u
m
q
u
e
u
e
l
e
n
g
t
h
0
5
10
15
20
25
30
35
40
45
Figure 6: Control eects comparison estimated by maximum
queue length.
RL method under the medium trac, which is better than
the xed control and the actuated control; under congested
condition, the multi-RL can eectively prevent the queue
spillovers to avoid large-scale trac jams. It should be also
noticed that multi-RL is a car-based algorithm. Therefore,
it is less time consuming than the light-based reinforcement
learning algorithms [13].
However, there are still some system parameters that
should be carefully determined by hand, for example, the
adjusting factor indicating the inuence of the queue at
next trac light to the waiting time of vehicles at current
light under the congested trac condition. This is a very
important parameter, which we should further research its
determining way based on the trac owtheory. In addition,
some phenomena in real trac system such as the lane
changing and overtaking of cars will inuence their travel
time. The assumption that all vehicles run at the same
speed is also not so reasonable. We would take these into
consideration and build a model closer to the real trac
system in future work. Besides, the communications between
trac signal controllers will help to observe the network-
wide trac states and predict future trac conditions, which
will improve the trac control eect and should be further
researched in the future.
Acknowledgments
This work is supported by the National High Technology
Research and Development Program (863 Program) of
China, Contract no.s 2006AA11Z229, 2007AA11Z215; by the
Key Project of Chinese National Programs for Fundamental
Research and Development (973 program), Contract no.
2006CB705506; by Chinese National Natural Science Foun-
dation, Contract nos. 60834001, 60774034.
References
[1] C. P. Pappis and E. H. Mamdani, Fuzzy logic controller for
a trac junction, IEEE Transactions on Systems, Man and
Cybernetics, vol. 7, no. 10, pp. 707717, 1977.
[2] M. B. Trabia, M. S. Kaseko, and M. Ande, A two-stage fuzzy
logic controller for trac signals, Transportation Research
Part C, vol. 7, no. 6, pp. 353367, 1999.
[3] J. C. Spall and D. C. Chin, Trac-responsive signal timing
for system-wide trac control, Transportation Research Part
C, vol. 5, no. 3-4, pp. 153163, 1997.
[4] Z. Liu, Hierarchical fuzzy neural network control for large
scale urban trac systems, Information and Control, vol. 26,
no. 6, pp. 441448, 1997.
[5] M. D. Foy et al., Signal timing determination using genetic
algorithms, Transportation Research Record 1365, National
Research Council, Washington, DC, USA, 1992.
[6] B. Park et al., Enhanced genetic algorithm for signal timing
optimization of oversaturated intersections, Transportation
Research Record 1727, National Research Council, Washing-
ton, DC, USA, 2000.
[7] R. S. Sutton, Learning to predict by the methods of temporal
dierences, Machine Learning, vol. 3, no. 1, pp. 944, 1988.
[8] C. Watkins, Learning from delayed rewards, Ph.D. thesis, Kings
College, Cambridge, UK, 1989.
[9] L. P. Kaelbling, M. L. Littman, and A. W. Moore, Rein-
forcement learning: a survey, Journal of Articial Intelligence
Research, vol. 4, pp. 237285, 1996.
[10] T. Thorpe, Vehicle trac light control using SARSA, M.S. thesis,
Colorado State University, 1997.
[11] B. Abdulhai, R. Pringle, and G. J. Karakoulas, Reinforcement
learning for true adaptive trac signal control, Journal of
Transportation Engineering, vol. 129, no. 3, pp. 278285, 2003.
[12] S. Mikami and Y. Kakazu, Genetic reinforcement learning for
cooperative trac signal control, in Proceedings of the 1st IEEE
Conference on Evolutionary Computation, vol. 1, pp. 223228,
Orlando, Fla, USA, June 1994.
[13] M. Wiering et al., Intelligent trac light control, Tech. Rep.
UU-CS-2004-029, University Utrecht, 2004.
[14] M. Wiering, Multi-agent reinforcement learning for trac
light control, in Proceedings of the 17th International Confer-
ence on Machine Learning (ICML 2000), pp. 11511158, 2000.
[15] C. F. Daganzo, Queue spillovers in transportation networks
with a route choice, Transportation Science, vol. 32, no. 1, pp.
311, 1998.
doi:10.1155/2010/656407
Research Article
Design and Experimental Evaluation of a Vehicular Network
Based on NEMOand MANET
Manabu Tsukada,
1
Jos e Santa,
2
Olivier Mehani,
1
Yacine Khaled,
1
and Thierry Ernst
1
1
INRIA Paris, Rocquencourt Domaine de Voluceau Rocquencourt, B.P. 105, 78153 Le Chesnay Cedex, France
2
Department of Information and Communications Engineering, University of Murcia, Campus de Espinardo, 30100 Murcia, Spain
Correspondence should be addressed to Manabu Tsukada, manabu.tsukada@inria.fr
Received 1 December 2009; Revised 19 July 2010; Accepted 5 September 2010
Copyright 2010 Manabu Tsukada et al. This is an open access article distributed under the Creative Commons Attribution
cited.
Mobile Ad hoc Network (MANET) routing protocols and Network Mobility (NEMO) Basic Support are considered key
technologies for vehicular networks. MANEMO, that is, the combination of MANET (for infrastructureless communications) and
NEMO (for infrastructure-based communications) oers a number of benets, such as route optimization or multihoming. With
the aim of assessing the benets of this synergy, this paper presents a policy-based solution to distribute trac among multiple
paths to improve the overall performance of a vehicular network. An integral vehicular communication testbed has been developed
to carry out eld trials. First, the performance of the Optimized Link State Routing protocol (OLSR) is evaluated in a vehicular
network with up to four vehicles. To analyze the impact of the vehicles position and movement on network performances, an
integrated evaluation environment called AnaVANET has been developed. Performance results have been geolocated using GPS
information. Second, by switching from NEMO to MANET, routes between vehicles are optimized, and the nal performance is
improved in terms of latency and bandwidth. Our experimental results show that the network operation is further improved with
simultaneous usage of NEMO and MANET.
1. Introduction
Terrestrial transportation is one of the most important
services that support humans daily life. Intelligent Trans-
portation Systems (ITS) aim at enhancing road trac
safety and eciency as well as optimizing social costs and
improving drivers comfort by providing services such as
eet management, route guidance, billing, or infotainment.
These days, communication technologies are more and more
considered as a key factor for ITS deployment however, new
approaches are needed to integrate mobile networks in the
vehicle eld.
IPv6 can be a good base technology to fulll several
ITS communication requirements, thanks to its extended
addressing space, embedded security, enhanced mobility
support, and autoconguration advances. Moreover, future
vehicles will embed a number of sensors and other IPv6-
enabled devices [1]. A number of ITS applications can be
conceived when sensors deployed in vehicles are connected
to the Internet and data collected from them is shared among
vehicles and infrastructure. Since the speed and position of
vehicles can be shared in real time, valuable information
about trac conditions can be inferred. For example, by
reporting brake events, vehicles driving towards the aected
road segment can be warned in advance and authorities can
be ready for possible fatalities.
In order to deal with communication requirements of
ITS applications [2], on-the-move and uninterrupted Inter-
net connectivity is necessary. Network Mobility (NEMO)
Basic Support has been specied by the IETF (Internet
Engineering Task Force) NEMO Working Group [3] to pro-
vide on-the-move IP connectivity maintaining addressing
conguration. NEMO is an essential part of the Commu-
nication Architecture for Communications Access for Land
Mobiles (CALM)) (http://www.calm.hu/), currently being
standardized at ISO [4]. The European ITS Communication
Architecture dened by COMeSafety [5] and ETSI [6] also
integrates NEMO and IPv6 to provide and maintain Internet
connectivity to vehicles.
Additionally, Mobile Ad hoc Networks (MANETs) can
be used for vehicular communications without depending
on any third-party infrastructure. Several MANET protocols
have been specied by the IETF MANET Working Group.
These routing protocols are classied as reactive or proac-
tive [7], depending on whether communication routes are
created when needed or they are continuously maintained.
The Optimized Link State Routing (OLSR) protocol has
been specied at IETF as a proactive protocol [8]. This
protocol has been chosen in the present research to create
a Vehicular Ad hoc Network (VANET), since it is a well-
know implemented, tested, and standardized protocol in the
MANET literature.
This paper describes the work done to combine NEMO
and MANET/VANET in a design that distributes trac
among multiple paths to improve the overall performance
of the vehicular network. A complete testbed has been
developed and used to experimentally evaluate the system.
The rest of the paper is organized as follows. Network
technologies related to vehicle communications are summa-
rized in Section 2. Section 3 outlines scenarios and objec-
tives of our network platform. Our integrated evaluation
environment for vehicular networks and the Linux-based
implementation are described in Section 4. Experimental
results are covered in the following two parts: Section 5
deals with the performance of the VANET subsystem, while
Section 6 evaluates the integrated MANEMO performance,
both indoor and outdoor, considering eld trials of the
IPv6 mobility testbed of the Anemone project [9]. Finally,
Section 7 concludes the paper summarizing main results and
addressing future works.
2. Network Technologies in
Vehicular Communications
This section presents a brief overview of relevant networking
technologies in vehicular communications. Several research
elds highly related to the work described in this paper,
regarding NEMO and MANET, are also introduced, such as
Multihoming, Route Optimization, and MANEMO.
2.1. VANET. Vehicular Ad hoc Networks (VANET) are a
particular case of MANET, but they are characterized by
battery constraints free, high speed, GPS-equipped nodes,
and regular distribution and movement. First, vehicles have
batteries better than the ones integrated in mobile terminals
or sensor devices. Moreover, they are recharged while the
vehicles engine is on. Hence, it is not necessary to take
specic measures to save energy resources (e.g., avoid signal-
ing trac). Second, mobility conditions of road vehicles are
dierent from the ones given in common portable terminals.
The relative speed between two vehicles driving in opposite
direction can reach 300 km/h. Thus, in some scenarios, the
lifetime of routing entries can be extremely short. Third, GPS
information can be assumed to be available in many cases,
since an increasing number of vehicles are equipped with
navigation systems. Position and movement information can
be used to improve network performances. Additionally, the
movement and density of vehicular nodes are not random,
since vehicles drive along roads. This makes the position of
nodes somehow predictable.
Although there are many works related to VANET
applications, as well as basic research at the physical link
and network layers in vehicular communications, there is
an important lack of real evaluation analysis. Many VANET
solutions and protocols could be considered as nonpractical
designs if they were tested in real scenarios, as it has
been proved for MANETs [10]. Performance of VANET
protocols based on a pure broadcast approach can be more
or less predictable in simple congurations, even if not
experimentally evaluated. However, the number of issues
concerning real performances of multhop designs is much
larger. There are works related to real evaluations of VANET
designs [11, 12], and a limited literature for concrete cases
of multi-hop transmissions [13], but there is an important
lack on works evaluating routing protocols on VANETs.
This paper details the works carried out towards easing
the experimental evaluation of a multi-hop and IPv6-based
vehicular network. The design covers the integration of
various communication technologies to overcome common
problems in VANETs, such as penetration rate or the need of
Internet connectivity.
OLSR is a well-known protocol in the MANET literature.
Since the application of MANET concepts in the particular
VANET case is a common procedure, the results given in
this paper assess how a common ad hoc proactive protocol
operates under vehicular conditions. Because vehicles are
not constrained by battery restrictions, one may think that
a proactive protocol tuned for highly dynamic topologies
could be suitable in the vehicular domain. Evaluating
this idea is an interesting point in the work. Moreover,
the existence of stable implementations of OLSR and its
popularity among real ad hoc deployments have encouraged
us to create a reference point in the VANET literature with
real multi-hop experiments based on this protocol. The
testbed platform presented in next sections is prepared to
change the routing protocol, thus it will be extended with
future implementations of pure-VANET protocols in the
frame of our research on georouting [14].
2.2. NEMO. The NEMO Basic Support functionalities
involve a router on the Internet to allow mobile computers
to communicate with mobile or static remote nodes. The
application of NEMOin ITS is direct and it is done as follows.
A Mobile Router (MR) located in the vehicle acts as a gateway
for the in-vehicle Mobile Network and manages mobility on
behalf of its attached nodes (Mobile Network Nodes, or MNNs
for short). MR and a xed router in the Internet, its Home
Agent (HA), establish a bidirectional tunnel to each other
which is used to transmit packets between the MNNs and
their Correspondent Nodes (CN).
The possible congurations oered by NEMO have been
classied in [15], according to three parameters: the number
Internet
Mobile network nodes
MR1
GPRS/UMTS WiMAX IEEE802.11x
HA1
HA2
MANET
NEMO
MR2
A
B
1) Integrated evaluation environment
2) Simultaneous usage of NEMO and MANET
Figure 1: Generic intervehicle communication scenario. Network nodes inside vehicles communicate with their peers via the VANET or
through the Internet using NEMO.
Figure 2: Prototype vehicles used in the eld experiments.
x of MRs in the mobile network, the number y of HAs
serving the mobile network and the number z of MNPs
(Mobile Network Prexes) advertised in the mobile network.
In this paper, we focus on the single MR, single HA and
single MNP conguration, commonly called (x, y, z) =
(1, 1, 1).
2.3. Multihoming. Mobile Routers can be shipped with
multiple network interfaces such as Wi-Fi (IEEE 802.11 a/b/g
and more recently 802.11 p), WiMAX (IEEE 802.16-2004/e-
2005) or GPRS/UMTS. When an MR simultaneously main-
tains several of these interfaces up and thus has multiple
paths to the Internet, it is said to be multihomed. In mobile
environments, MRs often suer from scarce bandwidth,
frequent link failures and limited coverage. Multihoming
brings the benets of alleviating these issues.
NEMO Basic Support establishes a tunnel between the
Home Agents address and one Care-of Address (CoA) of
the MR, even if the MR is equipped with several interfaces.
In [16], it is proposed the Multiple Care-of Addresses
Registration (MCoA), an extension of both Mobile IPv6 and
NEMO Basic Support, to establish multiple tunnels between
MRs and HAs. Each tunnel is identied by its Binding
Identication Number (BID). In other words, MCoA deals
with simultaneous usage of multiple interfaces.
2.4. Route Optimization. Route Optimization allows to sort
the communication path between a mobile router (or a host)
and a correspondent node that is not connected to the Home
Agent at a concrete moment. In NEMO, all the packets to and
fromMNNs must be encapsulated within the tunnel between
MR and HA. Thus, all packets to and from CNs must go
through HA. This causes various problems and performance
degradations. One could imagine the delay of using the HA
tunnel when both nodes could (in the worst case) be in the
same transiting network. A standardized solution for Route
Optimization is still missing for NEMO Basic Support, while
there exists one for Mobile IPv6 [17]. Main drawbacks of
such NEMO behavior can be classied as follows.
(1) Suboptimal routes are caused by packets being forced
to pass by HA. This leads to an increased delay
which is undesirable for applications such as real-
time multimedia streaming.
(2) Encapsulation with an additional 40-bytes header
increases the size of packets and the risk of frag-
mentation. This results in a longer processing time
for every packet being encapsulated and decapsulated
both at MR and HA.
(3) Bottlenecks in HA is an important problem, since a
signicant amount of trac for MNNs is aggregated
at HA, particularly when it supports several MRs
acting as gateways for several MNNs. This may cause
congestion which would in turn lead to additional
packet delays or even packet losses.
(4) Nested Mobility which occurs when a Mobile Router
get attached to other Mobile Routers. This could
arise, for example, when passengers carry a Per-
sonal Area Network or in scenarios where the same
outbound MR is used by several vehicles. Nested
Mobility further amplies the aforementioned route
suboptimality.
Vehicle network
MNN2
MR1
3G
Ethernet
MNN1
NEMO1
IPv4 internet
IPv6 over IPv4 tunnel
NEMO2
MR2 MR4
Ethernet
HA1
HA2
IEEE 802.11b
managed mode
Inria IPv6
network (France)
Infrastructure network
MR3
SFR 3G
IPv4 network
IPv6 over IPv4 tunnel
Irisa IPv6
network (France)
IEEE 802.11b ad hoc mode
Figure 3: Topology of the vehicular network and Internet connectivity.
The previous route optimization issues of NEMO are
identied in [18] by the IETF whereas the solution space is
analyzed in [19]. Requirements for Route Optimization in
various scenarios have been described for vehicle networks
in [20] and for aeronautic environments in [21].
2.5. MANEMO. Both MANET and NEMO are layer-three
technologies. NEMO is designed to provide global con-
nectivity, while MANET provides direct routes in wireless
local area networks. MANEMO combines both concepts to
provide several benets related to route optimization.
Since direct routes are available in MANETs, they can
provide direct paths between vehicles. These paths can be
optimal and free from NEMO tunnel overhead [22, 23].
Possible topology congurations with MANEMO have been
described in [24], while issues and requirements have been
summarized in [25]. In addition, MANEMO has already
been suggested for vehicular communications. For example,
VARON [26] focuses on NEMO route optimization using
MANET. It also provides the same level of security as the
current Internet, even if communications are done via the
MANET route.
3. Scenario and Objectives
This paper focuses on the scenario of intervehicle communi-
cation shown in Figure 1. Sensors installed in the vehicle are
connected to the Internet to share real-time information, and
on-board computers or mobile terminals (i.e., MNNs) are
connected to the mobile network within the vehicle. Vehicles
are connected to the Internet everywhere and anytime with
multiple interfaces using NEMO. Each MR, acting as a
gateway for the mobile network, supports both NEMO and
MANET connectivity.
In this paper, the focus is on investigating the operation
and performance of the simultaneous usage of VANET and
NEMO routes. An initial set up of a eld testbed based
on four-wheeled electric vehicles was carried out, called
CyCabs [27], to identify issues and requirements of real
environments. This testbed helped us to prepare a feasible
study considering issues such as wireless links features, con-
nectivity changes or vehicles movement. The experiments
presented in the following sections were conducted using
up to four common commercial vehicles (Citro en C3s) as
depicted in Figure 2.
Among the dierent advantages of the developed testbed,
three main capabilities can be remarked. First, apart from
studying trac ows sent through the xed network, it is
possible to evaluate VANET performances in detail using an
integrated testing environment. Second, the testbed is open
to develop and validate any ITS application. Third, a number
of dierent scenarios can be tested to analyze the operation
of all network layers working together.
In order to measure the network performance of a
VANET, various metrics should be considered. The band-
width, round-trip time (RTT), jitter, packet delivery ratio
(PDR), and number of hops are measured for various com-
munication types (e.g., UDP, TCP, or ICMPv6). Geographic
metrics, such as speed, position and distance between cars
are also collected and linked with the previous network mea-
suraments. As far as authors know, there are no integrated
tools that perform all this tasks at once.
Several issues arise when the previous performance mea-
surements are collected and linked. These can be grouped in
the next three classes.
(1) Path awareness. This comprises the problem of deter-
mining the route followed by packets from source to
destination in a dynamic topology.
(2) Performance measurements hop-by-hop. Perfor-
mance data is usually collected in an aggregate end-
to-end manner by classical network analysis tools
(e.g., ping6 or IPerf), but is not accessible on a
per-hop basis. Hence, it is not easy to identify where
packets are lost, for instance.
6
4
2
0
2
4
6
N
u
m
b
e
r
o
f
h
o
p
s
90
80
70
60
50
40
30
20
10
0
0
D
i
s
t
a
n
c
e
(
m
)
50 100 150 200 250 300 350 400
Time (seconds)
Car 3
Sender
(UDP, TCP, ICMPv6
trac generation)
Ethernet
(in-vehicle
network)
Ethernet
(in-vehicle
network)
MR3
Sender
script
Iperf and
ping6 logs
MR script MR script
MR script
Tcpdump
log
Tcpdump
log
Tcpdump
log
GPS log GPS log
GPS log
Receiver
script
Iperf log
(Optional) (Optional)
Car 4/2/
Car 1
MR4/2/
Receiver
Wi-Fi
(VANET)
Wi-Fi
(VANET)
Processing
XML
statistics
Packet
traces
Graphic
generator
Analysis
Web front-end (google maps)
AnaVANET
Experiments
Distance between MR3 and MR2
Distance between MR2 and MR1
Hops
Figure 4: Experimental setup and data processing units.
(3) Movement awareness. The route followed by vehicles
in the physical world is also an important issue to
further identify the cause of network problems due
to real mobility conditions.
Moreover, in preceding works [28], switching from a
NEMO to a MANET route gave benets regarding route
optimization in terms of bandwidth and delay. In this paper,
we also propose to distribute trac into multiple paths to
improve the global network performance. This simultaneous
usage of NEMO and MANET has been experimentally
evaluated within our testbed.
4. Vehicular Network Design and Testbed
Architecture
Our network architecture setup is detailed in this section.
First, the global architecture is introduced in Section 4.1.
Sections 4.2 and 4.3 focus on describing the evaluation
environment used to analyze the VANET performance and
the general MANEMO architecture, respectively.
4.1. Vehicular Network Architecture. The testbed comprises a
combination of vehicle-to-vehicle and infrastructure-based
::0 (NEMO route)
::/64 (MANET route)
::/64 (other route)
::128 (other route)
Packet
IF
IF
Routing table
Figure 5: Classic routing. A single routing table is used, and packets are forwarded along the route with the longest matching prex.
Packet IF
IF
Packet mark
Src address
Dst address
Src port
Dst port
Flow type
Routing tables
NEMO route (BID1)
Routing
policy
database
NEMO route (BID2)
MANET route
Other routes IF
Figure 6: Policy routing. Depending on several criteria, each packet is routed according to one of several routing tables.
networks, as Figure 3 depicts. Each vehicle is equipped with a
mobile router, with at least two interfaces: an Ethernet link
and an 802.11b adapter in ad hoc mode. MNNs connect
to the in-vehicle network via its Ethernet interface (an
internal managed Wi-Fi network could also be used for this
purpose), while the ad hoc Wi-Fi interface is used for the
inter-vehicle connections. In Figure 3, MR1 and MR2 are
also connected to an infrastructure network using another
802.11 interface in managed mode. MR1 has an additional
3G modem to establish a second link to the Internet (PPP
link provided by SFR (SFR is a french mobile telephony
operator partially owned by Vodafone) ). MR1 is supported
by HA1 at INRIA Rocquencourt and MR2 is supported by
HA2 inside Irisas network. Both networks are located in
France and interconnected via Renater (French backbone for
education and research) using a direct 6in4 tunnel to work
around some IPv6 rewalling problems (the testbed sites are
12 IPv4 hops apart).
4.2. VANET Experimentation Subsystem. An experimenta-
tion tool has been designed to overcome the issues related
to VANET evaluation described in Section 3. This software
covers the VANET part of the testbed architecture (i.e.,
bottom part of Figure 3).
4.2.1. Data Acquisition and Postprocessing Fusion with Ana-
VANET. An overview of the experimental evaluation process
is presented in Figure 4. The four vehicles previously
described are considered here, although the system can sup-
port any number of vehicles. A sender terminal (MNN), con-
nected to one of the in-vehicle networks, is in charge of gen-
erating data trac towards a receiver terminal (MNN) inside
another vehicle. Both sender and receiver save a high level
performance log according to the applications used to gener-
ate network trac. All MRs keep track of sent or forwarded
data packets using tcpdum (http://www.tcpdump.org/) and
log the vehicles position. All these data are then postpro-
cessed by the AnaVANET software.
AnaVANET is a Java application which traces all data
packets transmitted or forwarded by mobile routers. It thus
detects packet losses and can generate both end-to-end and
per-hop statistics, as well as join these measurements with
transport level statistics from the trac generation tool.
AnaVANET generates XML les with statistics at a one
second granularity, and packet trace les listing the path
followed by each data packet.
The XML statistics le is uploaded to a Web server, which
uses the Google Maps API to graphically replay the tests and
show performance measurements in a friendly way, as can
be seen in Figure 4. A screenshot of this web application
is available on Figure 10 in Section 5. All experiments
which have been performed up to now can be replayed and
main performance metrics can be monitored at any time,
by using the control buttons on the left side of the web
page. The replay speed can be adjusted and a step-by-step
mode has been implemented. On the map, the positions
and movements of the vehicles are depicted along with
their speed and distance to the rest of cars. The amount
of transferred data, throughput, packet loss rate, round-
trip time, and jitter, both per-hop and end-to-end, are also
displayed. Main network performances can be graphically
checked looking at the width and color of the link lines
among vehicles.
The Graphic Generator module gives another view of the
network performance. It processes both the XML statistics
and packet traces to generate several types of graphs using
GNUPlot (http://www.gnuplot.info/).
4.2.2. Trac Analysis and Performance Metrics. Three dif-
ferent types of trac have been considered over the IPv6
VANET in the tests.
UDP: A unidirectional transmission of UDP packets, from
the sender to the receiver terminal has been generated
using IPerf (http://iperf.sourceforge.net/). The packet
length is 1450 bytes to avoid IP fragmentation, and
they are sent at a rate of 1 Mbps.
TCP: A TCP connection is established between the sender
and receiver terminals without any bandwidth limit.
IF
MR
MNN
Packet
mark
NEMO routing table (BID2)
MAIN routing table
MANET routing table
NEMOD
OLSRD
OLSR node
OLSR node
IF
IF
IF
HA
BID1
BID2
Egress interface
User
policy
Ingress interface
Packets transmission
Entries addtion
IP tunnel
NEMO routing table (BID1)
Rule add/del
Route add/del
Route add/del
Routing
policy
database
Rule add/del
Ad hoc
Figure 7: Internal route updating and selection mechanisms. NEMO and OLSR routes are stored in completely independent routing tables.
Web server
3G
MNN1
HTTP request
(2 seconds)
IPerf server
Web server
MR1
IEEE 802.11b
infrastructure mode
MNN2
IPerf client
MR2
HA2
IEEE 802.11b
infrastructure mode
HA1
XML
IEEE 802.11b ad hoc mode
Figure 8: Network topology of the MANEMO demonstration system.
< markers >
<
markers >
<? xml version =
1.0
encoding =
utf-8
? >
marker interval =
2.0
transfer =
649
bandwidth =
2660
lat =
48.8375
lng=
2.1010
oset lat=
6.59
oset lng =
7.21
distance =
9.77
time =
1195225195
/ >
< /
Figure 9: XML data le generated based on IPerf measurements.
IPerf is again used as the trac generator. The
segment size is 1440 bytes.
ICMP: The Windows XP ping6 utility is used to generate
IPv6 ICMP echo request packets at the sender
node and to receive echo reply packets from the
remote note.
These three trac types have been used to analyze hop-
by-hop and end-to-end network performances, considering
the most extended metrics for MANET evaluations [7].
In the TCP case, statistics from IPerf with a 0.5 second
granularity, such as the current throughput, are directly used
by AnaVANET for performance analysis. Each ICMPv6 and
UDP packet is, however, traced across nodes. Since there is
Figure 10: AnaVANET replaying a VANET experiment. Buildings
avoid a direct line-of-sight communication, thus forcing the usage
of multihop routes.
no fragmentation for UDP packets, a direct correspondence
exists between MAC and IP packets. At this level, the
packet delivery ratio (PDR), number of hops and jitter are
calculated. For ICMPv6 tests, the RTT is logged to analyze
the network delay.
20
60
100
20
60
100
20
60
100
20
60
100
100 200 300 400 500 600
PDR from MR3 to MR1
0
20
40
60
80
100
PDR from MR3 to MR2
PDR from MR2 to MR1
0
100 200 300 400 500 600 0
100 200 300 400 500 600 0
100 200 300 400 500 600 0
PDR
Jitter
Hops
100 200 300 400 500 600 0
Time (seconds)
Time (seconds)
Time (seconds)
Time (seconds)
Time (seconds)
P
a
c
k
e
t
d
e
l
i
v
e
r
y
r
a
t
i
o
(
%
)
P
a
c
k
e
t
d
e
l
i
v
e
r
y
r
a
t
i
o
(
%
)
P
a
c
k
e
t
d
e
l
i
v
e
r
y
r
a
t
i
o
(
%
)
P
a
c
k
e
t
d
e
l
i
v
e
r
y
r
a
t
i
o
(
%
)
J
i
t
t
e
r
(
m
s
)
N
u
m
b
e
r
o
f
h
o
p
s
5
4
3
2
1
0
1
2
3
4
5
Figure 11: UDP performances over a multihop VANET of three cars moving in an urban environment.
4.3. MANEMO Implementation. A policy routing algorithm
has been developed and integrated in the architecture to
allow simultaneous usage of NEMO and MANET. This
subsystem allows vehicles to communicate with each other
over both the xed and VANET networks at the same time,
as was illustrated in Figure 3.
4.3.1. Policy Routing. The system has been implemented
on GNU/Linux (kernel 2.6.21.3). To distribute packets to
multiple paths simultaneously from a MR, a policy routing
scheme has been designed. Classic routing mechanisms are
not suitable because of the longest match principle. As
shown in Figure 5, packets arriving to the MR are forwarded
to the routing table entry which has the longest prex in
common with the destination address. In the MANEMO
case, MANET routes typically have longer prex lengths than
NEMOones. The formers are thus used in priority when they
are available in the routing table. NEMO routes then have
the least preference and are used as default routes. A single
routing table can be used for switching between routes but
not for simultaneous usage of NEMO and MANET.
To solve the previous problem, we propose multiple
routing tables using a Route Policy Database (RPDB), as
shown in Figure 6. To achieve this goal, the Netlter
(http://www.netlter.org/) framework is used. The RPDB
allows to maintain several independent routing tables in
the kernel. Each packet can then be routed according to
any of these tables. The determination of which routing
table that should be used in a particular case is up to
the implementation. It is usual to route depending on the
type of ow that is being treated. This mechanism allows
distributing packets to multiple concurrent routes at the
same time.
10 5 0 5 10 10 5 0 5 10 10 5 0 5 10
10 5 0 5 10 10 5 0 5 10
5 0 5 10 10 5 0 5 10 10 5 0 5 10
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
B
a
n
d
w
i
d
t
h
(
K
b
i
t
s
/
s
e
c
)
Average
1st
2nd
Average
1st
2nd
Average
1st
2nd Average
1st
2nd
Average
1st
2nd
Average
1st
2nd
Average
1st
2nd
Average
1st
2nd
NW
176, 344,
477, 594
W
192, 361,
489,
E
111, 272,
432, 551
SW
216, 388,
504
S
231, 403,
519
SE
97, 257,
419, 537
Time
N
146, 322,
459, 580
NE
127, 304,
447, 566
4th
3rd
4th
3rd
4th
3rd
4th
3rd
4th
3rd
3rd 3rd
3rd
Time (seconds) Time (seconds) Time (seconds)
Time (seconds)
Time (seconds) Time (seconds) Time (seconds)
Time (seconds)
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
0
200
400
600
800
1000
1200
Figure 12: Network throughput at corners and straight roads for the UDP multihop test.
0
100
200
300
400
500
600
0 50 100 150 200 250 300
Always 3G
3G or managed
Any interface
11b managed is available
11b ad-hoc
is available
Loss
Time (seconds)
R
o
u
n
d
t
r
i
p
t
i
m
e
(
m
s
)
Figure 13: Impact of route changes on the RTT, measured using ICMPv6 packets in the absence of background trac.
4.3.2. Implementation Details. NEPL (NEMO Platform
for Linux) (http://software.nautilus6.org/NEPL-UMIP/) ver-
sion 20070716 has been installed on MRs along with
olsrd (OLSR Daemon) (http://www.olsr.org/) version
0.5.3. NEPL is developed and distributed freely by Nau-
tilus6 (http://www.nautilus6.org/) within the WIDE project
(http://www.wide.ad.jp/). NEPL is based on MIPL (Mobile
IPv6 for Linux) (http://www.mobile-ipv6.org), developed
by the Go-Core (Helsinki University of Technology) and
Nautilus6 projects.
The OLSR daemon has been adapted to the routing
scheme proposed in Section 4.3.1. In this way, OLSR routing
entries are maintained in one of the multiple routing tables
of the kernel. The NEMO daemon already handles policy
routing when patched for MCoA support (http://software
.nautilus6.org/MCoA/).
NEMO and OLSR daemons operate independently in
MRs. The NEMO one maintains its binding update list
and NEMO routes, while the OLSR daemon takes care of
MANET routes. As shown in Figure 7, both NEMO and
MANET routing entries are kept up-to-date in separate
tables.
When started, both daemons add rule entries that specify
which packets should be routed according to which routing
table (these are removed at the execution end). MRs have
multiple routing tables, which save NEMO and MANET
routes, and the default one (depicted as MAIN in Figure 7),
which saves the rest of routes. There is the same number
of NEMO routing tables than egress interfaces on the MR.
Each of these routing tables has a specic BID. The MANET
routing table is used for trac that should be routed directly
to neighboring vehicles, and the MAIN table is mostly used
to route OLSR signaling.
Packets from MNNs arrive at the MR containing the
source and destination addresses and ports, as well as the
ow type information. Packets are distributed according to
the latter mark to either the NEMO or MANET routing
tables. Packets matched with a NEMO routing table are
transmitted to the tunnel bound to the HA, while packets
matched with the MANET table are transferred to other
OLSR nodes directly.
4.3.3. Demonstration Platform. As a demonstration of the
policy-based MANEMO system, the performance measured
in a communication between two vehicles is shown on
a website (http://fylvestre.inria.fr/tsukada/experiments/),
mapped to their geographical positions. The data have
been collected during eld trials on the Promotion Days
of the Anemone Project (1214th December 2007). This
project aims at developing a large-scale testbed for mobile
communication technologies. Our demonstration was an
example of a third party system using the mobility testbed.
Measurements were made with a GPS-enabled IPerf
(http://gforge.inria.fr/frs/?group id=620&release id=915) in
a topology as shown in Figure 8. This diagram illustrates in
detail the MANEMO part of the general vehicular network
described at the beginning of this section in Figure 3. MNN1
works as an IPerf server and MNN2 is the client. IPerf
reports the amount of transferred data and used bandwidth.
Additionally, the GPS patch appends location information
(latitude and longitude) as well as the oset and distance
from the starting point. Only a regular GPS receiver is
needed.
The demonstration can be performed either in real-
time or log mode. The former shows network performances
mapped with position in real time on the website, while the
latter saves themon the MNNs local disk to be displayed later
(see Figure 18).
In real-time mode, XML les are generated from mea-
sured metrics and positions every two seconds by MNN1.
An example XML output is shown in Figure 9. The remote
web server periodically gets the data le from MNN1 using
0
10
20
30
40
50
60
110 115 120 125 130
Loss
Up
(122.5, 21.27)
170 175 180 185 190
3G or managed
Any interface
Down
R
o
u
n
d
t
r
i
p
t
i
m
e
(
m
s
)
Time (seconds) Time (seconds)
Figure 14: Closer look at the RTT values collected when the adhoc interface is turned on and o.
0
1000
2000
3000
4000
5000
0 50 100 150 200 250 300
Any interface
3G or managed
Always 3G
11b ad-hoc
is available
Time (seconds)
T
h
r
o
u
g
h
p
u
t
(
k
b
p
s
)
Figure 15: Evolution of the throughput of three TCP ows between
MNNs using routing policies.
wget. The real-time mode has the advantage of everyone can
immediately check the network operation. Measurements
can, however, be slightly aected by the XML le transfers, as
they are carried over the NEMO route. By contrast, this eect
is not present in log mode. Main results of these experiments
are later analyzed in the paper using the log mode.
5. VANET-Only Performance Evaluation
This section presents an experimental evaluation of our
VANET testbed. Here, we only consider the lower part of
the architecture shown in Figure 3. Then, eld trials have
been performed using the integrated evaluation environment
described in Section 4.1. Seven dierent scenarios with
0
200
400
600
800
1000
0 50 100 150 200 250 300
Always 3G
3G or managed
Any interface
11b ad-hoc
is available
Loss
R
o
u
n
d
t
r
i
p
t
i
m
e
(
m
s
)
Time (seconds)
Figure 16: RTT between MNNs with three background TCP ows.
various mobility patterns have been evaluated, using the
three types of trac previously described (UDP, TCP, and
ICMPv6). This section analyzes one of our urban scenarios
as reference. See [29] for a complete description of all the
experiments.
5.1. Experimental Setup Details. In the VANET evaluation
shown in Figure 4, mobile routers use only the routing
table given by OLSR to forward data packets. MNN1 is
a Mac OS X 10.4 laptop and MNN2 is a Windows XP
Professional PC. An embedded computer is used as MR
in each car. It consists of a Soekris net4521 board with
a Texas Instruments ACX111 Mini-PCI 802.11 b/g wireless
transceiver and a compact ash memory card. The wireless
interface has been setup for 11 Mbps operation, emulating an
0
500
1000
1500
2000
2500
3000
3500
0 20 40 60 80 100 120 140
Any interface
3G or managed
Always 3G
11b managed
is available
11b managed
is available
Time (seconds)
T
h
r
o
u
g
h
p
u
t
(
k
b
p
s
)
11b ad hoc is available
Figure 17: Throughputs of three TCP streams in a eld experiment.
Performances are less stable than in the indoor case.
Table 1: OLSR Parameters for the VANET-only experiments.
Parameter Value (default)
HELLO interval 0.5 sec (2.0 sec)
HELLO validity 6.0 sec (6.0 sec)
HNA interval 3.0 sec (5.0 sec)
HNA validity 9.0 sec (15.0 sec)
802.11 b device. This computer is also connected via a serial
port to a Trimble AgGPS 323 GPS receiver. Both the Wi-Fi
and the GPS antennas are axed on the roof of the car.
The timing parameters of the OLSR daemon installed in
MRs have been modied as showed in Table 1, to accom-
modate mobility conditions of a vehicular network. These
modications enable MRs to discover topology changes
more quickly.
5.2. Non-Line-of-Sight Multihop Communication. This sce-
nario considers a typical urban environment where buildings
prevent a direct line of sight between the source and
destination cars. A multi-hop network is better suited to
provide a robust connectivity under these conditions.
During 600 seconds of test, a unidirectional transmission
of UDP packets is generated from MNN1, in vehicle 3, to
MNN2, in vehicle 1. As was explained in Section 4.2, the
packet size is 1450 Bytes to avoid IP fragmentation and they
are sent at a rate of 1 Mbps.
The results of this experiment (along with the rest
of performed trials) is available on a public website
(http://fylvestre.inria.fr/tsukada/experiments/vanet-jose/),
and can be replayed to graphically show the performance of
the network during the tests (see Figure 10).
The experiment was performed in the Rocquencourt
campus of INRIA. This area contains a set of small buildings
surrounded by streets, as can be seen in Figure 10. The
four streets showed in the image, which round three of the
buildings, have been chosen for this scenario. They stand
in a 100 100 m square area. Three vehicles have been
driven around the buildings, trying to block the direct link
between cars one and three. The speed of the vehicles was
kept between 15 and 30 km/h. The right and left roads visible
in Figure 10 are very narrow and some communication
problems were experienced when approaching the corners.
The results collected in the UDP tests are plotted in
Figure 11. The several graphs show the results collected
during four tests around the buildings. The upper plot shows
the number of hops used in the paths followed by UDP
packets whereas the lower graphs show the end-to-end and
per-link PDR. PDR is computed every second, while the
number of hops is plotted for each packet transmitted from
the sender node. When no hops are drawn, the route to the
destination vehicle is not available. Zero hops means that the
packet was sent by the rst MR, but was not received by any
other. Negative values represent those packets that did not
arrive to their destination, but reached some intermediate
hops.
As can be seen, a direct relation exists between PDR and
number of hops. When the number of hops is equal to or
lower than zero, PDR decreases. When the vehicles drive
along the same street, some direct paths (one hop) appear.
On the contrary, when the distance between the sender and
the receiver cars is large enough, the two-hop routes are
used. These dierent types of paths can also be observed
with the per-link PDR. Whereas the direct link (MR3-MR1)
gives intermediate PDR values, the PDR between consecutive
vehicles is almost identical and close to 100% when the two-
hop link is used, due to the lower distances between nodes.
The performance obtained in the scenario has been
analyzed according to the location of vehicles: corner and
straight road. As can be seen in Figure 12, each corner is
called SE, NE, NW, SW, and according to its position (i.e.,
South-East for SE). In the same way, straight roads have
been assigned the names E, N, W, and S. Numbers below
corner and road names indicate the time in which car 2
(vehicle in the middle) passes these points. For example, for
SE, numbers 97, 257, 419, and 537 mean that car 2 passes
the SE corner at times 97s, 257s, 419s, and 537s. At these
times, the sender vehicle is at the next road (E in this case)
and the receiver vehicle is at the previous one (S). On the
other hand, when the middle vehicle is at a straight road, the
sender vehicle reaches the next corner and the receiver vehicle
is at the previous corner. The driving order is SE - E - NE -
N - NW - W - SW - S, and three and a half complete rounds
have been considered. The analysis starts at time 97s at SE
corner, and it ends at 594 s at NW.
As can be seen in Figure 12, the throughput obtained has
been mapped with corner and straight road segments, and it
has been analyzed for each round at periods of 10 seconds.
A dotted line shows the result of each trial considered in the
segment, and a bold line shows the average bandwidth. One
can notice the two dierent bandwidth patterns obtained
at corners and straight roads. Communication performance
increases in corner scenarios, while it decreases at straight
roads. When the intermediate vehicle reaches a corner, the
direct path between the source and the destination vehicle is
blocked, thus a multi-hop route is established in the network.
At this moment, a good bandwidth is noticeable. When the
heading vehicle turns again in the next corner, a multi-hop
route still maintains connectivity but, as soon as car 3 and
2 cannot maintain the communication link, the network
performance falls. This can be seen in the last ten seconds
of straight road graphs in Figure 12. The eect of loosing
the link between vehicles 2 and 1 is also present when the
intermediate vehicle left the corner (last seconds of corner
graphs and rst seconds of straight road graphs), but it is less
noticeable, since these two vehicles were arbitrarily driven
more closely during tests.
Results obtained for segment W shows a dierent behav-
ior than for the rest of straight roads. This is explained by
the special physical conditions of the environment. First, this
stretch comprises a narrow street surrounded by buildings
on both sides. As can be seen in Figure 10, these conditions
are only present in this segment, since the rest of roads have
a clear space on one side. This fact enables the reection of
signals on the various walls. Moreover, the second interesting
condition identied in road W is the greater altitude of the
sender car with regard to the receiver car, when these are
located near corners NW and SW, respectively. About ve
meters of altitude dierence increases the packet reception
probability, and a direct path between cars 1 and 3 is even
noticeable at this segment. This can be checked at time
367s in Figure 11. It is interesting to note that buildings
on this INRIA area are quite low, about 3.5 meters, what
complements the altitude eect. The rest of direct paths
collected in the trials belong to segments S and E, which do
have open areas on one of the sides.
6. MANEMO Performance Evaluation
For the case of the MANEMO subsystem described in
Section 4.2, measurements of latency and throughput have
been collected using both the VANET and the infrastructure
segments of the testbed (Figure 3). A set of indoor and
oudoor experiments have been conducted also at the Roc-
quencourt campus of INRIA and this section presents and
analyzes most interesting results.
6.1. Experimental Setup Details and Initial Tests. Attending
to the global testbed setup in Figure 3, tests have been
carried out generating trac from MNN1 towards MNN2
using the best available communication route (i.e., NEMO
or MANET). MNN1 is a Mac OSX 10.4 laptop and MNN2
is a Windows XP tablet PC. As was done for VANET-
only experiments, OLSR settings have been adjusted with
the values shown in Table 2. These have been chosen
to maintain a tradeo between the delay experimented
when a topology change occurs and the network overload
that implies control messages. Signaling trac, apart from
reducing the eective bandwidth of the network, it consumes
computation resources on the nodes. For these experiments,
OLSR settings have been adjusted to be aware of topology
changes faster than in VANET-only tests in Section 5, with
Figure 18: Website screen shot of the MANEMO experiment.
0
500
1000
1500
2000
2500
3000
Throughput
T
h
r
o
u
g
h
p
u
t
(
k
b
p
s
)
Distance (meters)
Average throughput
1
0
1
5
2
0
2
5
3
0
3
5
4
0
4
5
5
0
5
5
6
0
6
0
Figure 19: Relation between distance and bandwidth using the

MANEMO system.
the aim of performing fast route changes between NEMO
and MANET.
Some initial tests were performed to check the oper-
ation of the network. One issue that had to be solved
was radio interferences between 802.11b managed and ad
hoc networks. Even when channels were chosen with a
good distribution the problem persisted. To overcome this
drawback, the bandwidth of MR1 interfaces were limited
to 2 Mbps using a Linux QoS system based on tc (Trac
Control) (http://www.linux-foundation.org/en/Net:Iputils).
Network performance measurements under static condi-
tions, and without any policy, between MNN1 and MNN2
are summarized in Table 3, including three dierent routes.
RTT results is the average of 100 packets of ICMPv6 between
MNNs and throughput results have been obtained averaging
results obtained during a total of ten minutes of TCP tests.
Table 2: OLSR parameters for the MANEMO experiments.
Parameter Value
HELLO interval 0.5 sec
HELLO validity 1.5 sec
HNA interval 1.0 sec
HNA validity 3.0 sec
Table 3: Performance of the MANEMO system under static
conditions and depending on the route type.
Route & interface RTT Throughput
NEMO over 3G 279.43 ms 416 kbps
NEMO over 802.11b managed 32.74 ms 1977 kbps
OLSR over 802.11b ad-hoc 8.58 ms 1987 kbps
As can be seen, the results of the UMTS link oer the worse
results, due to the delay of the operators network and the
bandwidth limited by the radio coverage and the available
resources in the cell. The 802.11b link improves these results,
oering bandwidth capabilities equivalent to the ad hoc case.
However, the delay induced by the managed mode and the
relay access point impact on the RTT results.
6.2. Indoor Test Scenario. The policy-based MANEMO sys-
tem has been rstly evaluated in an indoor testbed, to avoid
interferences of other equipments and diculties to trace the
movement of MRs. The following experiments have been
performed without any vehicle. Neither MRs nor MNNs
have moved during a reference experiment of 300 seconds.
It clearly demonstrates the performance expected for longer
times or subsequent trials.
MNN1 has three addresses (A, B, and C) in the MNP,
and MR1 distributes trac from the mobile network via
multiple paths depending on the source address. Packets
from source address A or to port number 5102 are always
forwarded via the 3G interface. Those from source address
B or to destination port 5101 are routed via the Wi-Fi
managed interface when it is available. Otherwise, they are
forwarded over the 3G interface. Trac from source address
C or to destination port number 5009 is transmitted via
whatever available interface, prioritizing the most ecient
(i.e., prefer ad hoc to managed Wi-Fi and only use 3 G if
no other link is available). Table 4 summarizes these policies
and indicates priorities in case several routes can be chosen.
MR2 distributes returning ows into its managed and ad hoc
interfaces according to the third policy, since it does not have
any 3 Ginterface. The Home Agents distribute ows to match
these policies and avoid asymmetric routes.
For this indoor experiment, connection and discon-
nection events have been created using a shell script and
common system tools. From t = 0 to t = 60, both Wi-Fi
managed and ad hoc interfaces of MR1 are down. At t = 60,
the managed interface comes up. At t = 120, the ad-hoc one
is made available. From t = 120 to t = 180, all the interfaces
are up and running. At t = 180, the ad-hoc link is turned o.
At t = 240, the managed one is also switched o. The 3 G
interface is always available throughout the test.
6.2.1. Latency Measurements. To measure the RTT between
MNNs, MNN1 sends 56 Bytes ICMPv6 echo request
packets from all addresses (A, B, and C) to MNN2 once every
0.5 sec. There is no other trac. These packets are distributed
according to the policies described above. Results are showed
in Figure 13. The average RTT on the NEMO route over
3 G has been 261.9 ms. Changing paths to the NEMO route
over the managed Wi-Fi interface, has reduced the RTT to an
average of 34.72 ms, which represents an 87% improvement.
During the time the ad hoc mode has been available, the
average RTT collected on the OLSR route (ad hoc link)
has been 7.93 ms. In this way, route optimization using
MANEMOhas further reduced the latency by 26.79 ms, what
represents an extra improvement of 77%.
For the two periods where the three ICMPv6 ows are
carried over the 3G network (from t = 0 to t = 60 and from
t = 240 to t = 300) an oset of 20 ms of delay between them
is noticeable. It has been checked that the transmission of the
three echo request packets in a consecutive way results in
a rst-in rst-out problemdue to transmission and reception
times needed by the 3G driver. The extra overload incurred
by the NEMO and tunnel and the L2TP (Layer-2 Tunneling
Protocol) tunnel, necessary to support IPv6 trac in the 3G
network, increases the impact of this eect.
Figure 14 gives a closer look at the RTT results when the
ad hoc interface goes up/down and routes thus change. At
t = 120, the ad hoc interface comes up, and then direct route
information of both MNPs are exchanged. At t = 122.5,
the RTT obtained for the marked packet is 21.27 ms, which
comprises an intermediate value between NEMO and OLSR
modes. This is because the ICMPv6 echo request has
used the NEMO route, while the echo reply has returned
through the ad hoc one. It takes 2.5 seconds for OLSR
routing entries to be added to MR1s table after the ad hoc
link has been connected. By contrast, the route is changed
back from OLSR to NEMO 1.5 sec after the ad hoc link
is disconnected. During this switching phase, three packets
have been lost (Fromt = 180 to t = 181.5), due to the sudden
disconnection of the ad hoc interface.
6.2.2. Throughput Measurements. To measure the through-
put between MNNs, MNN1 sends three TCP streams to
MNN2 by means of IPerf, with destination port numbers
5102, 5101 and 5009, in the same routing scenario used
above. At the same time, MNN1 also sends 56 Bytes ICMPv6
echo request packets as in the previous section. IPerf gives
a report once every two seconds and ping6 gives it every 0.5
seconds. A reference test has been chosen among the set of
performed tests. Figure 15 shows the achieved throughput
with stacked area graph and Figure 16 shows the observed
RTT when the TCP ows are active.
A summary of throughput results is given in Table 5.
The average total throughput on the NEMO route over 3G is
455 kbps from t = 0 to t = 60. Since an 802.11b managed
network is available from t = 60 to t = 120, the ows
Table 4: Flow distribution policies for MR1. Smaller numbers reect higher priorities.
Policy Targets 3G Managed Wi-Fi Adhoc Wi-Fi
Always 3G
Source address A or
1
destination port 5102
3G or managed
Source address B or
2 1
Any interface
Source address C or
3 2 1
are distributed in two paths and the average throughput
increases up to 1913 kbps, which represents an improvement
of 76% (1458 kbps). From t = 120 to t = 180, ad hoc
connectivity is also available. The average total throughput
increases again up to 3752 kbps when the three TCP ows
are distributed through the three paths, which represents a
new improvement of 49% (1837 kbps).
The average RTT between MNNs is also listed in Table 5.
The RTT on the NEMO route is about 400 ms when the three
TCP streams are transmitted using the 3G link. When two
TCP streams are diverted to the 802.11b managed interface,
from t = 60 to t = 120, the RTT over the 3G link decreases
by about 280 ms, which represents an improvement of 30%.
The RTT also decreases from 400 ms to about 130 ms for
policies 3G or managed and Any interface when Wi-Fi
managed is available, which comprises an improvement of
68%. In addition, a further 50% (approx.) of improvement is
observed for policies 3G or managed and Any interface
when all the interfaces on MR1 are available, since each
communication technology is used by only one ow.
6.3. Field Experiment. The system has been evaluated with a
set of eld trials performed on the T el ecom Bretagne/INRIA
Rennes campus. 40 access points have been installed in
this area. The test has been performed in a straight road
surrounded by buildings, where two access points have been
installed at two far away locations. The source vehicle starts
moving at a speed of 10 km/h from a position before the
rst access point, while the destination vehicle with MR2
has been parked next the two access points. Both MRs
were mounted inside the vehicles. Three TCP ows were
transmitted from MNN1 to MNN2 as in the previous tests.
The ow distribution policies of MR1, MR2, HA1 and
HA2 are also identical to those of the indoor testbed but,
obviously, periods of 802.11 connectivity are not simulated
now.
The switch between access media and/or networks has
a clear impact on the available bandwidth, as can be seen
in Figure 17. From t = 0 to t = 60, the path between
MNNs is only via the NEMO route over 3G. The average
total throughput of the TCP ows is 344 kbps during this
period. The throughput in this eld experiment is 111 kbps
less than in the indoor experiment. This is mostly due to
obstacles and movements of the vehicle equipped with MR1.
From t = 62 to t = 86 and from t = 106 to t = 116,
the NEMO route through managed Wi-Fi is available, since
the moving vehicle is near one access point each time. The
average total throughput of the TCP stream at these two
periods is 1430.83 kbps and 957.34 kbps, respectively. From
t = 124 to t = 130, the OLSR route over the VANET is
available. The average throughput increases until 2408.4 kbps
during this period.
In the evaluation, the NEMO route on the 802.11b
managed interface has been used for 24 seconds for the
rst access point, and then an additional 10 seconds for the
second one. As the speed of the vehicle was 10 km/h, the
coverage of the access points can be estimated to be between
30 and 65 meters. The ad hoc interface has been available
during six seconds, thus the VANET range can be estimated
to be 17 meters. In this case, the antennas of both MRs
were located inside the vehicles. 802.11 performance could
therefore be improved by mounting external and/or more
powerful antennas.
6.4. Impact of Geographical Location on Network Performance.
In the previous section, the range of the available access
points and VANET links are estimated considering a sim-
plication of the driving speed and the time of connection.
This section presents more thorough range measurements.
These have been collected by maintaining MR1 (and thus
MNN1) moving in a 65 meters radius around the position
of MR2, and reporting the achievable throughputs. None of
the MRs have gone out of the access points coverage and a
building sometimes block the VANET route. As the wireless
access points are quite close to the test site, the managed
interface has been forcibly limited to 1 Mbps to account for
more distant APs and highlight which network path MR1
uses. By contrast, the ad hoc interface was not limited and
the average throughput between MRs using this interface was
2685 kbps.
The position-mapped throughputs were measured at
INRIA Rocquencourt in France, using the GPS-patched
version of IPerf. To obtain a high density of throughput data
around MR2, the evaluation was actually performed without
vehicles. MR1 has been carried by a human at an average
speed of 4 km/h. It starts moving from the position of MR2
and comes back to the same position within 250 seconds. The
experiments were run eight times. All the results are publicly
available (http://fylvestre.inria.fr/tsukada/experiments/).
A screenshot of the Web application can be seen in
Figure 18. The website displays the throughput between
MNNs by varying the size of the blue circles at each measure
point. All tests can be displayed by selecting the Log option.
Clicking on one of the circles reveals additional information,
Table 5: Total throughput of three TCP ows and RTT between MNNs.
Policy
Available interfaces
3G only 3G and managed All interfaces
Throughput
Always 3G 156 kbps 262 kbps 276 kbps
3G or managed 184 kbps 733 kbps 1612 kbps
Any interface 114 kbps 918 kbps 1863 kbps
Total 455 kbps 1913 kbps 3752 kbps
Round-Trip Time
Always 3G 389 ms 277 ms 275 ms
3G or managed 411 ms 127 ms 64 ms
Any interface 432 ms 130 ms 64 ms
including the test time, location and oset from the starting
point, transferred data size and bandwidth in the last two
seconds. All the data can be shown at once with the Show Log
button. Users can also analyze the results by changing data
density and see the trajectory of MR1 (and thus MNN1).
Throughput results depending on the distance between
MNNs can be seen in Figure 19. Data points show the
throughput obtained at the current distance, while the bar
graph represents an average for ve meters. Values over
1 Mbps (the arbitrary limitation on the managed Wi-Fi link)
are those recorded when the VANET route was available. One
can see that it is available up to 40 meters. Between (approx.)
20 and 40 meters, throughput measurements spread over
a wide range from 100 kbps to 2700 kbps, because media
handovers between the managed and ad hoc interfaces are
performed in these zones.
An asymmetrical tendency of the ad hoc link ranges has
been observed. From the collected results, it turns out that
the OLSR route is usable over a longer distance when two
vehicles are getting further from one another than when they
are getting closer. This hysteresis behavior is due to OLSRs
initial delay caused by the period of sharing HELLO packets.
This fact is further analyzed in [29].
7. Conclusions and Future Works
A proposal to distribute data trac in vehicular communi-
cations combining NEMO and VANET has been presented.
This comprises an integral communication platform for
the ITS frame, which has been experimentally evaluated by
means of a complete and open testbed. In a rst stage, an
integrated evaluation environment for VANET enabled us to
analyze the network performance in detail in both per-hop
and end-to-end manners, also considering the movement
of vehicles. The evaluation environment provides novel
performance metrics for VANET, according to the current
literature, such as the number of hops used to deliver a packet
or the per-link PDR, in addition to typical statistics, such as
end-to-end packet delivery ratio, round-trip delay time, jitter
or bandwidth. Although it has been tuned to dynamic con-
ditions, the OLSR protocol shows limitations to eciently
update routing tables under stressful conditions, as it has
been seen above all in the MANEMO evaluation. A more
VANET-oriented protocol developed at INRIA, in frames of
the GeoNet Project (http://www.geonet-project.eu/) will be
evaluated through new eld trials, using the presented test-
bed. This is located inside the geographic-based routing pro-
posals, which are demonstrating to be the correct direction
in vehicular network research.
The MANEMO proposal has been evaluated using
common vehicles in real environments. Up to four vehicles
have been setup to carry out a number of experiments. In
our system, mobile routers use multiple egress interfaces
simultaneously with NEMO and OLSR. The latter could
thus mitigate the sub-optimality caused by NEMO routes.
Previous experiments results showed that MANEMO with
route switching from NEMO to MANET improved network
performance in terms of latency and bandwidth. It can
now be stated that MANEMO with simultaneous usage of
NEMO and MANET can achieve further improvements on a
integrated vehicular network. Experimental results show that
the achievable throughput and delay are improved when a set
of interfaces (3G, 802.11 b managed and 802.11 b ad hoc) are
available.
Among the dierent research lines that are now active
regarding this work, we plan to extend the MANEMO system
as follows. First, evaluation results show that network perfor-
mances such as latency and bandwidth dynamically change
according to available interfaces, mobility or obstacles. Adap-
tive applications are thus desirable in these environments.
Second, trac ows have to be allocated to appropriate
paths depending on the application demands and network
performances. Since real-time applications are sensitive to
handovers, an intelligent path allocation is required. Third,
trac between MNNs has been distributed according to
policies manually specied by the administrator. As an MR
can only control its outbound trac, policy changes on an
MR may create asymmetric routes. By introducing common
lter rules and a exchanging procedure among MRs and HAs
[30, 31], policies on each entity can be synchronized. Fourth,
currently, the position-mapped reports for the MANEMO
case are focused on bandwidth statistics. Network metrics
such as latency, packet loss rate or layer-two information will
be considered in further analysis for the whole MANEMO
system, in the line of the work carried out for the VANET
segment.
Acknowledgments
The authors would like to thank the partners of European
project Anemone, and more specically T el ecom Bretagne
and INRIA Rennes who built the IPv6 mobility testbed from
which the MANEMOexperiments exposed in this paper have
largely beneted.
References
[1] T. Ernst, The information technology era of the vehic-
ular industry, ACM SIGCOMM Computer Communication
Review, vol. 36, no. 2, pp. 4952, 2006.
[2] Y. Khaled, M. Tsukada, J. Santa, J. Choi, and T. Ernst, A usage
oriented analysis of vehicular networks: from technologies to
applications, Journal of Communications, vol. 4, no. 5, pp.
357368, 2009.
[3] V. Devarapalli, R. Wakikawa, A. Petrescu, and P. Thubert,
Network mobility (NEMO) basic support protocol, RFC
3963 (Proposed Standard), January 2005.
[4] ISO/TC204 WG16. ISO/WD 21210, CALMmedium and
Long Range, High Speed, Air Interfaces parameters and
protocols for broadcast, point to point, vehicle to vehicle,
and vehicle to point communication in the ITS sectorIPv6
Networking, Working draft, February 2009.
[5] R. Bossom, R. Brignolo, T. Ernst et al., European ITS commu-
nication architectureoverall frameworkproof of concept
implementation, Tech. Rep. Version 2.0, EC Information
Society Technologies Programme, March 2009.
[6] T. Kosch, I. Kulp, M. Bechler, M. Strassberger, B. Weyl, and
R. Lasowski, Communication architecture for cooperative
systems in Europe, IEEE Communications Magazine, vol. 47,
no. 5, pp. 116125, 2009.
[7] Z. Chang, G. Gaydadjiev, and S. Vassiliadis, Routing proto-
cols for mobile ad-hoc networks: current development and
evaluations, in Proceedings of the Annual Workshop on Cir-
cuits, Systems and Signal Processing, pp. 489494, Veldhoven,
The Netherlands, November 2005.
[8] T. Clausen and P. Jacquet, Optimized link state routing
protocol (OLSR), RFC 3626 (Experimental), October 2003.
[9] L. Bokor, N. Montavont, P. Di Francesco, T. Ernst, T. Hof,
and J. Korva, ANEMONE: a pan-European testbed to validate
IPv6 mobility technologies, in Proceedings of the International
Symposium on Applications and the Internet, Workshop on
Network Mobility (SAINT WONEMO 07), Hiroshima, Japan,
2007.
[10] C. Tschudin, H. Lundgren, and E. Nordstr om, Embedding
MANETs in the real world, in Proceedings of the 8th IFIP
Conference on Personal Wireless Communications, vol. 2775 of
Lecture Notes in Computer Science, pp. 578589, September
2003.
[11] V. Gonz alez, A. Los Santos, C. Pinart, and F. Milagro,
Experimental demonstration of the viability of IEEE 802.11b
based inter-vehicle communications, in Proceedings of the
4th International Conference on Testbeds and Research Infras-
tructures for the Development of Networks & Communities
(TridentCom 08), pp. 17, Innsbruck, Austria, 2008.
[12] J. P. Singh, N. Bambos, B. Srinivasan, and D. Clawin,
Wireless LAN performance under varied stress conditions
in vehicular trac scenarios, in Proceedings of the 56th IEEE
Vehicular Technology Conference (VTC 02), vol. 2, pp. 743
747, Vancouver, Canada, September 2002.
[13] M. Jerbi, S.-M. Senouci, and M. Al Haj, Extensive experi-
mental characterization of communications in vehicular ad
hoc networks within dierent environments, in Proceedings
of the 65th IEEE Vehicular Technology Conference (VTC 07),
pp. 25902594, Dublin, Ireland, 2007.
[14] M. Tsukada, I. Ben-Jemaa, H. Menouar, W. Zhang, M. Goleva,
and T. Ernst, Experimental evaluation for IPv6 over VANET
geographical routing, in Proceedings of the 6th International
Wireless Communications and Mobile Computing Conference,
pp. 736741, Caen, France, June 2010.
[15] C. Ng, T. Ernst, E. Paik, and M. Bagnulo, Analysis of
multihoming in network mobility support, RFC 4980 (Infor-
mational), October 2007.
[16] R. Wakikawa, V. Devarapalli, G. Tsirtsis, T. Ernst, and K.
Nagami, Multiple Care-of Addresses Registration, RFC 5648
(Proposed Standard), October 2009.
[17] D. Johnson, C. Perkins, and J. Arkko, Mobility support in
IPv6, RFC 3775 (Proposed Standard), June 2004.
[18] C. Ng, P. Thubert, M. Watari, and F. Zhao, Network
mobility route optimization problem statement, RFC 4888
(Informational), July 2007.
[19] C. Ng, F. Zhao, M. Watari, and P. Thubert, Network
mobility route optimization solution space analysis, RFC
4889 (Informational), July 2007.
[20] R. Baldessari, T. Ernst, A. Festag, and M. Lenardi, Automotive
industry requirements for NEMO route optimization, Jan-
uary 2009.
[21] W. Eddy, W. Ivancic, and T. Davis, Network Mobility
Route Optimization Requirements for Operational Use in
Aeronautics and Space Exploration Mobile Networks, RFC
5522 (Informational), October 2009.
[22] R. Wakikawa, K. Okada, R. Koodli, A. Nilsson, and J. Murai,
Design of vehicle network: mobile gateway for MANET and
NEMO converged communication, in Proceedings of the 2nd
ACM International Workshop on Vehicular Ad Hoc Networks
(VANET 05), pp. 8182, September 2005.
[23] J. Lorchat and K. Uehara, Optimized inter-vehicle communi-
cations using NEMO and MANET, in Proceedings of the 2nd
International Workshop on Vehicle-to-Vehicle Communications
(V2VCOM 06), July 2006.
[24] R. Wakikawa, T. Clausen, B. McCarthy, and A. Petrescu,
MANEMO topology and addressing architecture, July 2007.
[25] R. Wakikawa, P. Thubert, T. Boot, J. Bound, and B. McCarthy,
Problem statement and requirements for MANEMO, July
2007.
[26] C. J. Bernardos, I. Soto, M. Calder on, F. Boavida, and A.
Azcorra, VARON: vehicular ad hoc route optimisation for
NEMO, Computer Communications, vol. 30, no. 8, pp. 1765
1784, 2007.
[27] C. Pradalier, J. Hermosillo, C. Koike, C. Braillon, P. Bessi` ere,
and C. Laugier, The CyCab: a car-like robot navigating
autonomously and safely among pedestrians, Robotics and
Autonomous Systems, vol. 50, no. 1, pp. 5167, 2005.
[28] M. Tsukada and T. Ernst, Vehicle communication experiment
environment with MANET and NEMO, in Proceedings of
the International Symposium on Applications and the Internet,
Workshop on Network Mobility (SAINT WONEMO 07), p. 45,
Hiroshima, Japan, January 2007.
[29] J. Santa, M. Tsukadat, T. Emstt, and A. F. Gomez-Skarmeta,
Experimental analysis of multi-hop routing in vehicular ad-
hoc networks, in Proceedings of the 2nd Workshop on Exper-
imental Evaluation and Deployment Experiences on Vehicular
Networks in Conjunction with the International Conference on
Testbeds and Research Infrastructures for the Development of
Networks and Communities and Workshops (TridentCom 09),
Washington, DC, USA, April 2009.
[30] C. Larsson, M. Eriksson, K. Mitsuya, K. Tasaka, and R. Kuntz,
Flow Distribution Rule Language for Multi-Access Nodes,
IETF, Work in progress, draft-larsson-mext-ow-distribution-
rules-02, February 2009.
[31] H. Soliman, G. Tsirtsis, N. Montavont, G. Giaretta, and
K. Kuladinithi, Flow Bindings in Mobile IPv6 and Nemo
Basic Support, IETF, Work in progress, draft-ietf-mext-ow-
binding-04, November 2009.
doi:10.1155/2010/719294
Research Article
Trafc Data Collection for Floating Car Data Enhancement in
V2I Networks
D. F. Llorca, M. A. Sotelo, S. S anchez, M. Oca na, J. M. Rodr guez-Ascariz,
and M. A. Garc a-Garrido
Department of Electronics, University of Alcal a, Ctra. N-II Km. 33, Alcal a de Henares, C.P. 28871 Madrid, Spain
Correspondence should be addressed to D. F. Llorca, llorca@depeca.uah.es
Received 19 November 2009; Accepted 5 July 2010
Academic Editor: Shahrokh Valaee
Copyright 2010 D. F. Llorca et al. This is an open access article distributed under the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.
This paper presents a complete vision-based vehicle detection system for oating car data (FCD) enhancement in the context of
vehicular ad hoc networks (VANETs). Three cameras (side-, forward- and rear-looking cameras) are installed onboard a vehicle
in a eet of public buses. Thus, a more representative local description of the trac conditions (extended FCD) can be obtained.
Specically, the vision modules detect the number of vehicles contained in the local area of the host vehicle (trac load) and
their relative velocities. Absolute velocities (average road speed) and global positioning are obtained after combining the outputs
provided by the vision modules with the data supplied by the CAN Bus and the GPS sensor. This information is transmitted by
means of a GPRS/UMTS data connection to a central unit which merges the extended FCD in order to maintain an updated
map of the trac conditions (trac load and average road speed). The presented experiments are promising in terms of detection
performance and computational costs. However, signicant eort is further necessary before deploying a system for large-scale
real applications.
1. Introduction
Floating car data (FCD) refers to technology that collects
trac state information from a set of individual vehicles
which oat in the current trac. Each vehicle, which can
be seen as a moving sensor that operates in a distributed
network, is equipped with positioning (GPS) and commu-
nication (GSM, GPRS, UMTS, etc.) systems, transmitting
its location, speed, and direction to a central control unit
that integrates the information provided by each one of the
vehicles.
FCD systems are being increasingly used in a variety of
important applications since they overcome the limitations
of xed trac monitoring technologies (installation and
maintenance costs, lack of exibility, static nature of the
information, etc.). We refer to [1] for general background
concerning the most representative FCD activities in Japan,
Europe, and the United States.
FCD can be used by the public sector to collect road
trac statistics and to carry out real-time road trac control.
The information provided by FCDsystems can be supplied to
individual drivers via dynamic message signs, PDA devices,
satellite navigation systems, or mobile phones, including
dynamic rerouting information. Thus, drivers would be
able to make more informed choices, spending less time in
congested trac. In addition, the knowledge of the current
trac situation can be also used to estimate time of arrival
of a eet of public transport vehicles and, furthermore,
to plan and coordinate the movements of the eet (eet
management) so that driving assignments can be carried out
more eciently. Besides previous applications, the use of
FCD entails environmental benets since it can be used to
reduce fuel consumption and emissions.
The basic data provided by FCD systems (vehicle loca-
tion, speed, and direction) can be enriched using new
onboard sensors (ambient temperature, humidity and light,
windshield wiper status, fog light status, fuel consump-
tion, emissions, tire pressure, suspension, emergency brake,
etc.) which are centralized by means of the controller-
area-network (CAN) bus. Such data can be exploited to
extend the information horizon including trac, weather,
road management, and safety applications [1]. In addition,
computer vision systems can be included in order to
improve the automatic detection of potentially interesting
events and to document them by sending extended data
[2].
In order to provide ubiquitous coverage of the entire road
network, a minimum representation of the total passenger
car eet has to be used, since each moving sensor (each
vehicle) only supplies information about its status. The
fact that everyday road users have to be asked to share
information regarding their movements and speeds arises
privacy issues that have to be addressed. Many potential road
travellers may be reluctant to join FCD projects because of
violations of their privacy due to permanent traceability or
possible liability in case of speed limit violations. Thus, the
fundamental concept for FCD systems calls for no identi-
cation information to be sent with the basic data, which
can be easily implemented from a technical perspective. For
example, in [3] a general method for anonymization of FCD
by deriving pseudonyms for trips is presented.
Another approach consists of using the information sup-
plied by a specic eet of vehicles, rather than information
coming from individual road users. Taxis or public transport
buses can be used due to the extended periods of time
they spend on the urban road network. Although taxis and
buses provide a major source of innercity trac information
because of the time they spend mobile, they have limitations.
Problems arise if the taxi drivers, through detailed knowledge
of the local road network, take steps to avoid congested areas
which will not be reported [4]. Trac load perception may
be lower than the actual one if reserved taxis or buses lanes
are used. On the contrary, privacy issues are not as critical
as before, especially when using a eet of public transport
buses.
This paper presents a complete vision-based vehicle
detection system onboard a eet of public transport buses
with the aim of improving the data collected in FCD
applications. The proposed system has been developed in the
framework of the GUIADE project. Three cameras covering
the local environment of the vehicle are used: forward-rear-
and side-looking cameras. The system obtains under certain
constraints, such as good weather and daytime conditions,
the number of vehicles in the local range of the bus as well
as their relative position and velocity. This information is
combined with the data provided by regular FCD systems
(global location, speed, and direction), obtaining a more
detailed description of the local trac load and the average
speed. The communication system between the vehicles and
the central control unit is based on wireless technology
via GPRS/UMTS cellular protocols. Finally, the central unit
integrates the data collected by the eet in order to generate
updated trac status maps.
The remainder of this paper is organized as follows: the
description of the system including the wireless communica-
tion scheme is summarized in Section 2. Section 3 describes
the vision-based vehicle detection systemas well as the spatial
and temporal integration of the collected data. Experimental
results that validate the proposed approach are presented in
Section 4. Finally, conclusions and future works are discussed
in Section 5.
2. SystemDescripction
The proposed FCD architecture can be seen in Figure 1.
Floating car data is supplied by a eet of public transport
buses which corresponds to an inner-city bus line. Each vehi-
cle is equipped with a global positioning system (GPS), wire-
less communication interfaces (GPRS/UMTS and WLAN
IEEE 802.11) and a complete vision-based vehicle detection
system.
The vehicle-to-infrastructure (V2I) communication sys-
tem is based on the geographic coverage provided by cellular
networks. General packet radio service (GPRS) and universal
mobile telecommunications system (UMTS) are used to
connect each vehicle with the central control unit. Each
vehicle provides information that can be divided in three
main groups.
(1) Standard FCD information: vehicle identier (2
bytes), timestamp (11 bytes), GPS position (8 bytes),
speed (2 bytes), and direction (2 bytes).
(2) Vehicle status information: ambient temperature (2
bytes), humidity (2 bytes), light (2 bytes), windshield
wiper status (1 byte), fog light status (1 byte), fuel
consumption (4 bytes), and emissions (4 bytes).
(3) Extended FCD information: globally referenced aver-
age trac load (2 bytes) and average road speed for a
measured segment travel time (2 bytes).
As can be observed, the total message size per vehicle is
45 bytes. The extended FCD information is supplied to the
central unit at a frequency of 1 Hz. Accordingly, the band-
width currently demanded by vehicular communication in
the communication channel, that is, the vehicle throughput,
is 360 bps without overheads. This value can be considered
negligible taking into account the available bandwidth and
the proposed FCD architecture.
The central control unit integrates the information
provided by each one of the vehicles in order to compute
updated trac and weather maps which will be used for eet
management tasks as well as to estimate the time of arrival.
The vehicle-to-vehicle (V2V) communication system
is dened as a backup communication system based on
a wireless-delity (WiFi) IEEE 802.11a/b/g interface. In
situations where the cellular network is not working, in-
range vehicles will exchange the most updated information
available.
One of the main advantages of the proposed approach
is that it does not need to deal with privacy issues since the
oating vehicles correspond to a eet of public transport
buses.
3. Vision-Based Trafc Detection System
In this section, we present the main contribution of this
work: a complete vision-based trac detection system which
enhances the data supplied by standard FCD systems.
The benets of using computer vision instead of other
technologies such as radar-based systems can be summarized
as follows. Computer vision systems can compensate for
FLEET
V2V WLAN
backup
communication
Global
positioning
system
3G/GPRS/UMTS
cellular
communications
Central control unit
Figure 1: Overview of the proposed FCD architecture.
the lower angular resolution of the low-cost radar and the
increased appearance of ghost radar targets (guard-rails,
railings, lamp posts, reections, etc.). These false positives
are relevant and they cannot simply be ignored. The camera
has very good angular resolution and can be used to
determine height, width, and lateral speed of the target.
Pattern recognition can be used to classify the object and
even weakly reective targets such as pedestrians can be
detected. Moreover, the cost of a vision system is signicantly
lower than the cost saved by using the simpler radar. A vision
system, in addition to overcoming cost reduction problems,
can contribute to the system features such as road analysis
and scene understanding.
Each individual vehicle is equipped with three FireWire
cameras (forward-, rear- and side-looking cameras) that
cover the local environment of the bus (see Figure 2). A com-
mon hardware trigger synchronizes the image acquisition of
the three cameras and an onboard PC houses the computer
vision software.
Each individual vehicle detection system provides infor-
mation about the number of detected vehicles and both
their relative position and speed. These results are combined
with the GPS measurements and the data provided by the
CAN bus in order to provide globally referenced trac
information. This scheme is described in Figure 3.
The layers of the proposed architecture of the three vision
modules are conceptually the same: lane detection, vehicle
candidates selection, vehicle recognition, and tracking. The rst
step of each one of the vision systems consists of reducing the
searching space in the image plane in an intelligent manner
in order to increase the performance of the vehicle detection
module. Accordingly, road lane markings are detected and
used as the guidelines that drive the vehicle searching process
(see Figure 4). The area contained by the limits of the lanes is
scanned in order to nd vehicle candidates that are passed
on to the vehicle recognition modules. Thus, the rate of
false positives is reduced. In case that no lane markings are
detected, a basic region of interest is used instead covering the
front, rear, and side parts of the vehicle. Finally, a tracking
stage is implemented using Kalman ltering techniques.
3.1. Lane Detection. An attention mechanism is necessary in
order to lter out inappropriate candidate windows based
on the lack of distinctive features, such as horizontal edges
and vertical symmetrical structures, which are essential
characteristics of road vehicles. This has the positive eect
of decreasing both the total computation time and the
rate of false positive detections. Lane markings are detected
using gradient information in combination with a local
thresholding method which is adapted to the width of the
projected lane markings. Then, clothoid curves are tted
to the detected markings. The algorithm scans up to 25
lines in the candidates searching area, from 2 meters in
front of the camera position to the maximum range in
order to nd the lane marking measurements. The proposed
method implements a nonuniform spacing search that
GPS
Forward
looking
camera
Rear
looking
camera
Side looking camera
Figure 2: Main vehicle sensors: three cameras (forward, rear, and side looking cameras) and a global positioning system.
Forward
vehicle
detection
Rear
vehicle
detection
Side
vehicle
detection
GPS
Hardware
triggering
CAN
bus
Data fusion
(spatial and temporal integration)
Figure 3: Stages of the vision-based trac detection system.
Forward lane
detection
Forward lane
detection
Side lane
detection
Rear lane
detection
Rear lane
detection
Figure 4: Rear, side, and forward lane detection.
reduces certain instabilities in the tted curve. The nal state
vector is composed of 6 variables [5] for each lane on the
road
x = [c
oh
, c
1h
, c
ov
, c
1v
, x
o
,
o
, w
o
]
T
, (1)
where c
oh
and c
1h
represent the clothoid horizontal curvature
parameters, c
ov
and c
1v
stand for the clothoid vertical curva-
ture parameters, while x
o
,
o
, and w
o
are the lateral error and
orientation error with regard to the centre of the lane and
the width of the lane, respectively. The clothoid curves are
then estimated based on lane marking measurements using a
Kalman lter for each lane.
Apart from the detected road lanes additional virtual
lanes have been considered so as to cope with situations
in which a vehicle is located between two lanes (e.g., if
it is performing a change lane manoeuvre). Virtual lanes
provide the necessary overlap between lanes, avoiding both
misdetections and double detections caused by the two
halves of a vehicle being separately detected as two potential
vehicles. A virtual lane is located to provide overlap between
two adjoining lanes. Figure 5 provides some examples of lane
markings detection in real outdoor scenarios. Detected lanes
determine the vehicle searching area and help reduce false
positive detections. In case no lane markings are detected
by the system, xed lanes corresponding to a straight road
model are assumed instead.
3.2. Side Vehicle Detection. Side vehicle detection module [6]
relies on the computation of optical ow. In order to reduce
computational time, optical ow is computed only on Canny
points in the image. Canny edge pixels are consequently
matched and grouped together in order to detect clusters
of pixels that can be considered as candidate vehicles in the
image. Classical clustering techniques are used to determine
groups of pixels, as well as their likelihood to form a single
object. Even after pixels clustering, some clusters can still
be clearly regarded as belonging to the same real object.
A second grouping stage (double-stage) is then carried out
among dierent clusters in order to determine which of
them can be further merged into a single blob. For this
purpose, simple distance criteria are considered. Two objects
that are very close to each other are nally grouped together
in the same cluster. The reason for computing a two-stage
clustering process relies on the fact that by selecting a small
distance parameter in the rst stage, interesting information
about clusters in the scene can be obtained. Otherwise, using
a large distance parameter in the single clustering process,
highly gross clusters would have been achieved, losing all
information about the granular content of the points that
provide optical ow in the image.
The selected clusters constitute the starting point for
locating candidate vehicles in the image. For that purpose,
the detected positions of clusters are used as a seed point
to search for a collection of horizontal edges that could
potentially represent the lower part of a car. The candidate
is located on the detected horizontal edges that meet
certain conditions of entropy and vertical symmetry. Some
of the most critical aspects in side vehicle detection are
(a) (b) (c)
Figure 5: Vehicle searching area as a result of the lane markings analysis for forward, rear and side modules.
HOG/SVM
classication
Yes No
Warning
Yes No No
No
No
Yes
Yes
Yes
Vehicle
behaviour
Pre-detector
Frontal
part
detection
Mass
center
detection
Tracking
(relative position
and speed)
Figure 6: Side vehicle detection ow diagram.
subsequently listed: (1) shadows on the asphalt due to
lampposts, other artefacts or a large vehicle overtaking the
ego-vehicle on the right lane; (2) self-shadow reected on the
asphalt (especially problematic in sharp turns like in round-
about points), or self-shadow reected on road protection
fences; (3) robust performance in tunnels; and (4) avoiding
false alarms due to vehicles on the third lane.
The ow diagram of the two-stage detection algorithm
is depicted in Figure 6. As can be observed, there is a pre-
detector that discriminates whether the detected object is
behaving like a vehicle or not. If so, the frontal part of the
vehicle is located in the region of interest. In addition, the
vehicle mass centre is computed. In case the frontal part of
the vehicle is properly detected and its mass centre can also
be computed, a nal warning message is issued. After being
located, vehicle candidates are classied by using a linear
SVM classier [7] with HOG features [8] previously trained
with the samples obtained from real road images, and at that
point vehicle tracking starts. Tracking is stopped when the
vehicle gets out of the image. Sometimes, the shadow of the
vehicle remains in the image for a while after the vehicle
disappears from the scene, provoking the warning alarm to
hold on for 1 or 2 seconds. This is not a problem, however,
since the overtaking car is running in parallel with the ego-
vehicle during that time although it is out of the image scene.
Thus, maintaining the alarm in such cases turns out to be a
desirable side eect.
Figure 7 shows an example of blind spot detection in a
sequence of images. The indicator depicted in the upper-
right part of the gure toggles from green to blue when
a vehicle enters the blind spot area (indicated by a green
polygon). A blue bounding box depicts the position of the
detected vehicle.
3.3. Forward and Rear Vehicle Detection. Forward- and rear-
looking vehicle detection systems share the same algorithmic
core. The attention mechanism sequentially scans each road
lane from the bottom to the maximum range looking for
a set of features that might represent a potential vehicle.
Firstly, the vehicle contact point is searched by means
of the top-hat transformation. This operator allows the
detection of contrasted objects on nonuniform backgrounds
[9]. There are two dierent types of top-hat transformations:
white hat and black hat. The white hat transformation
is dened as the residue between the original image and
its opening ( operator). The black hat transformation is
dened as the residue between the closing ( operator) and
the original image. The white and black hat transformations
are analytically dened as follows:
WH
T
_
x, y
_
=
_
f f b
__
x, y
_
White Hat, (2)
BH
T
_
x, y
_
=
_
f b f
__
x, y
_
Black Hat. (3)
The opening operator () is dened as the dilation of the
erosion and the closing operator () is dened as the erosion
of the dilation (for more details see [10]). In our case we use
the white hat operator (2) since it enhances the boundary
between the vehicles and the road [11]. Horizontal contact
points are preselected if the number of white top-hat features
is greater than a congurable threshold. Then, candidates are
preselected if the entropy of Canny points is high enough
(a) (b) (c) (d)
Figure 7: Example of the side-vehicle detection module (also called blind spot detection) in a sequence of images. The indicator in the
upper-right part of the gure toggles from green to blue when a car is detected in the blind spot.
(a) (b) (c)
Figure 8: From left to right: original image; contact point detection on white top-hat image; candidate preselected with high entropy of
Canny points.
(a)
(b)
Figure 9: Canny images after adaptive thresholding.
(a)
(b)
(c)
Figure 10: Upper row: gray level symmetry; Middle row: vertical edges symmetry; Lower row: horizontal edges symmetry.
(a) (b)
Figure 11: (a) Overlapped candidates. (b) Nonmaximum suppression results.
for a region dened by means of perspective constraints and
prior knowledge of target objects (see Figure 8).
Before computing the Canny features, an adaptive
thresholding method is applied. This process is based on
an iterative algorithm that gradually increases the contrast
of the image, and compares the number of Canny points
obtained in the contrast increased image with the number
of edges obtained in the current image. If the number of
Canny features in the actual image is higher than in the
contrast increased image the algorithm stops. Otherwise,
the contrast is gradually increased and the process resumed.
This adaptive thresholding method permits to obtain ro-
bust image edges, as depicted in the examples provided in
Figure 9.
(a)
(b)
Figure 12: Forward data set. (a) positive samples (vehicles). (b) negative samples.
(a)
(b)
Figure 13: Rear data set. (a) positive samples (vehicles). (b) negative samples.
(a) (b)
Figure 14: Linear SVM with HOG features classication examples: nonvehicle (red) and vehicle (green).
(a) (b)
(c) (d)
Figure 15: Data association by features matching. (a, b) Harris features on image t. (c, d) matched Harris features on image t + 1.
0
1
2
3
4
5
6
N
u
m
b
e
r
o
f
v
e
h
i
c
l
e
s
i
n
r
a
n
g
e
0 100 200 300 400 500 600 700 800
Frames
Ground truth
Detected vechicles (Side/forward/rear)
Figure 16: Number of vehicles detected by the three vision modules compared with the manually labeled ground truth in a real sequence.
In a second step, vertical edges (S
v
), horizontal edges
(S
h
), and grey level (S
g
) symmetries are obtained, so that,
candidates will only pass to the next stage if their symmetries
values are greater than a threshold. The vertical and hori-
zontal edges symmetries are computed as listed inAlgorithm
1. The grey level symmetry computation procedure is
shown inAlgorithm 2. Some examples of the three types of
symmetries are depicted in Figure 10.
Symmetry axes are linearly combined to obtain the
nal position of the candidate. Finally, a weighted vari-
able is dened as a function of the entropy of Canny
points, the three symmetry values and the distance to the
host vehicle. We use this variable to apply a nonmaxi-
mum suppression process per lane which removes over-
lapped candidates. An example of this process is shown in
Figure 11.
(a) (b)
Figure 17: Examples with strong illumination changes after passing beneath a bridge.
(1) Initialize Acc
0,...,ROI
WIDTH
= 0
(2) For i = 0, ..., ROI
HEIGHT
(3) For each pair of vertical/horizontal edge pixels (x
1
, i) and (x
2
, i)
(4) Acc
(x1+x2)/2
+ +
(5) S
v,h
= arg
i
max(Acc
i
/S
v,h,MAX
)
Algorithm 1: Vertical and horizontal edges symmetries computation procedure.
The selected candidates are classied by means of a
linear SVM classier [7], in combination with histograms of
oriented gradients features [8]. We have developed and tested
two dierent classiers depending on the module (forward
and rear classiers). All candidates are resized to a xed
size of 64 64 pixels to facilitate the features extraction
process. The rear-SVM classier is trained with 2000 samples
and tested with 1000 samples (1/1 positive/negative ratio)
whereas the forward-SVM classier is trained with 3000
samples and tested with 2000 samples (1/1 positive/negative
ratio). Figures 12 and 13 depict some positive and negative
samples of the forward and rear training and test data sets,
respectively. Figure 14 shows a couple of examples of vehicle
detection after linear SVM classication with HOG features.
After detecting consecutively an object classied as
vehicle a predened number of times (empirically set to
3 in this work), data association and tracking stages are
triggered. The data association problemis addressed by using
feature matching techniques. Harris features are detected
and matched between two consecutive frames as depicted in
Figure 15.
Tracking is implemented using Kalman ltering tech-
niques [12]. For this purpose, a dynamic state model and
a measurement model must be dened. The proposed
dynamic state model is simple. Let us consider the state
vector x
n
, dened as follows:
x
n
=
_
u, v, w, h, u, v, w,

h
_
T
. (4)
In the state vector x and y are the respective horizontal and
vertical image coordinates for the top left corner of every
object, and w and h are the respective width and height in
the image plane, a dynamical model equation can be written
like this
x
n+1
= A x
n
+
n
=
1 0 0 0 t 0 0 0
0 1 0 0 0 t 0 0
0 0 1 0 0 0 t 0
0 0 0 1 0 0 0 t
0 0 0 0 1 0 0 0
0 0 0 0 0 1 0 0
0 0 0 0 0 0 1 0
0 0 0 0 0 0 0 1
u
v
w
h
u
v
w
n
+
n
.
(5)
In the model, t is the simple time, A represents the system
dynamics matrix and
n
is the noise associated to the model.
Although the denition of A is simple, it proves to be
highly eective in practice since the real time operation
of the system permits to assure that there will not be
great dierences in distance for the same vehicle between
consecutive frames. The model noise has been modelled as a
function of distance and camera resolution. The state model
equation is used for prediction in the rst step of the Kalman
lter. The next step is to dene the measurement model. The
(1) For each possible symmetry axis x
i
initializes S
i
= 0
(2) For j = 0, ..., ROI
HEIGHT
(3) For k = 0, ..., ROI
WIDTH
/2
(4) If abs(image
[ j][x
i
+k]
image
[ j][x
i
k]
) <
(5) S
i
+ +
(6) S
g
= arg
i
max(S
i
/(area
ROI
/2))
Algorithm 2: Gray level symmetry computation procedure.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
T
r
a
c
l
o
a
d
(
L
i
)
0 100 200 300 400 500 600 700 800
Frame
Figure 18: Trac load L
i
at every frame in a real sequence.
0
20
40
60
80
100
120
140
A
v
e
r
a
g
e
r
o
a
d
s
p
e
e
d
a
t
f
r
a
m
e
i
(
k
m
/
h
)
0 100 200 300 400 500 600 700 800
Frame
Figure 19: Average road speed v
i
at every frame in a real sequence.
measurement vector is dened as z
n
= [u, v, w, h]
T
. Then,
the measurement model equation is established as follows:
z
n+1
= H x
n
+ v
n
=
1 0 0 0 0 0 0 0
0 1 0 0 0 0 0 0
0 0 1 0 0 0 0 0
0 0 0 1 0 0 0 0
u
v
w
h
u
v
w
n
+ v
n
.
(6)
In last equation, H represents the measurement matrix and
v
n
is the noise associated to the measurement process. The
purpose of the Kalman ltering is to obtain a more stable
position of the detected vehicles. Besides, oscillations in
vehicles position due to the unevenness of the road makes
v coordinate of the detected vehicles change several pixels up
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
A
v
e
r
a
g
e
t
r
a
c
l
o
a
d
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31
Time (seconds)
Figure 20: Average trac load at every second in a real sequence.
0
20
40
60
80
100
120
140
A
v
e
r
a
g
e
r
o
a
d
s
p
e
e
d
(
k
m
/
h
)
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31
Time (secconds)
Figure 21: Average road speed at every second in a real sequence.
or down. This eect makes the distance detection unstable,
so a Kalman lter is necessary for minimizing these kinds of
oscillations.
3.4. FCD Integration. As depicted in Figure 3, the FCD
integration or Data Fusion module uses three sources of data:
the measurements provided by the GPS, the data supplied
by the CAN bus, and the output obtained from the three
vision-based vehicle detection modules. Whereas the GPS
and the CAN bus sample frequency is 1 Hz, the vision-based
system operates in real-time at 25 frames per second (25 Hz).
The proposed data fusion scheme provides information at
the lowest sample frequency (1 Hz) covering two consecutive
GPS measurements, the vehicle speed v
h
i
(via CAN bus) and
the outputs of the vision module.
The outputs of the side, forward, and rear vehicle
detection systems at frame i are the number of detected
vehicles N
i
and their corresponding distances to the host
vehicle d
(k)
i
(note that d is used here as a distance/range
Figure 22: GPS trajectory and the corresponding trac load computed at the central unit (the aerial image has been obtained from Google
Earth).
measurement). These outputs are combined to cover the
whole local environment of the vehicle. The trac load at
frame i is given by next expression
L
i
=
N
i
+ 1
N
MAX
, (7)
where N
MAX
is the maximum number of vehicles in range
that can be detected by the three systems (in our case N
MAX
is
dened as 9 or 13 for two lanes and three lanes roads, resp.).
The average road speed at frame i is computed as follows:
v
i
=
1
N
i
+ 1
Ni 1
_
k=0
_
d
(k)
i
d
(k)
i1
_
t
+ v
h
i
+ v
h
i
=
1
(N
i
+ 1)t
Ni 1
_
k=0
_
d
(k)
i
d
(k)
i1
_
+ v
h
i
,
(8)
where d
(k)
i
and d
(k)
i1
represent the distance between the host
vehicle and vehicle k at frames i and i 1, respectively, t
corresponds to the sample time, v
h
i
is the host vehicle speed
provided by the CAN bus, and N
i
is the number of detected
vehicles. Note that the distance values correspond to ltered
measurements since they are obtained from the rst two
elements of the Kalman lter state vector (u and v) using
known camera geometry and ground-plane constraints.
Two consecutive GPS measurements dene both a spatial
and a temporal segment. The temporal segment corresponds
to the GPS sample time (1 second), and the spatial segment
will be dened as the globally referenced trajectory between
the two GPS measurements. In order to obtain the extended
FCD information (i.e., the road trac load and the road
speed) for this spatio/temporal segment we integrate the
values supplied by the vision modules during 25 consecutive
frames. With this approach a dense coverage of the road
trac load and the road speed can be assured for host vehicle
speeds up to 180 km/h since the total range of the vision
module covers more than 50 m (25 meters for both the rear
and the forward looking modules; the side range covers up
to two third parts of the bus length in the adjacent lane).
Obviously this maximum speed will never be exceeded by a
public bus. This approach facilitates further map-matching
tasks since the extended FCD information between two
consecutive points will always be globally referenced.
4. Experimental Results
The system was implemented on a PC Core 2 Duo at 3.0 GHz
and tested in real trac conditions using CMOS cameras
with low-resolution images (320 240). After training and
test, a tradeo point has been chosen at detection rate (DR)
of 95% and false positive rate (FPR) of 5% for the rear-
SVM classier and at DR of 90% and FPR of 6% for the
forward-SVM classier. We have to note that these numbers
are obtained in an oine single-frame fashion, so that, they
will be improved in subsequently stages. In addition, the lane
detection system reduces the searching area and the number
of false candidates passed to further stages.
In order to validate the proposed vision-based vehicle
detection system as an extended source for FCD applications
we have recorded several video sequences in real trac
conditions, and we have manually labeled the number of
vehicles in range at every frame (a total of 800 frames). The
speed of the host vehicle was around 90 km/h so the length
of the traveled route was 1 km approximately. Both the trac
load L
i
and the average road speed v
i
are computed at every
frame using (7) and (8). Figure 16 shows the ground truth
and the number of vehicles detected in range. Most of the
errors take places in cases where the host vehicle is passing
beneath a bridge due to strong illumination changes (see
Figure 17) and in curves or cases where there are strong
changes in the vehicle pitch, roll or camera height.
The trac load L
i
and the average road speed v
i
at
every frame are depicted in Figures 18 and 19, respectively.
These values are provided by the vision modules at a
frequency of 25 Hz. As the extended FCD information is
supplied to the central unit at a frequency of 1 Hz the
trac load and the average road speed are nally integrated
during 25 consecutive frames. These results are shown in
Figures 20 and 21. We use a colour code to describe the
level of trac load and the road speed: green indicates
that there is good ow/high speeds; yellow indicates that
there is semi-dense trac/medium speeds, and red shows
dense trac/slow speeds (trac jams). After combining the
results with the GPS measurements we can obtain the trac
load in universal transverse mercator (UTM) coordinates, as
depicted in Figure 22 (note that map-matching is not carried
out; the aerial image has been obtained from Google Earth).
5. Concluding Remarks
This paper presented a complete vision-based vehicle detec-
tion system that enhances the data supplied by FCD systems
in the context of vehicular ad hoc networks. The system
is composed of three vision subsystems (side, forward and
rear subsystems) that detect the trac load and the relative
velocities of the vehicles contained in the local area of
the host vehicle. Under certain constraints, such as good
weather and daytime conditions, absolute velocities, and
global positioning are obtained after combining the outputs
provided by the vision modules with the outputs supplied
by the CAN Bus and the GPS sensor. Standard FCD systems
provide the vehicle position, speed, and direction. The
proposed approach extends this information by including
more representative measurements corresponding to the
trac load and the average road speed.
In order to cover the entire road network, the proposed
vision-based system is dened for being installed onboard
a eet of public buses where privacy is a minor issue.
The extended packets collected by each moving vehicle are
transmitted to the central unit by means of a GPRS/UMTS
data connection. The central unit merges the extended FCD
in order to maintain an updated map of the trac conditions
(trac load and average road speed).
The presented experiments are promising in terms of
detection performance and computational costs. However,
signicant eort is further necessary before deploying a
systemfor large-scale real applications. For this purpose, new
experiments will be carried out merging the data collected by
more than one vehicle, including map-matching techniques
and further analysis on V2I and V2V communications (e.g.,
using repetition based MAC protocols [13]). In addition,
the proposed vision-based vehicle detection system will be
extended to deal with complex weather conditions (e.g., wet
or snowy roads) as well as night-time conditions.
Acknowledgments
This work has been supported by the Spanish Ministry
of Science and Innovation by means of Research Grant
TRANSITO TRA2008-06602-C03 and Spanish Ministry of
Development by means of Research Grant GUIADE P9/08.
References
[1] R. Bishop, Intelligent Vehicle Technologies and Trends, Artech
House, Boston, Mass, USA, 2005.
[2] S. Messelodi, C. M. Modena, M. Zanin et al., Intelligent
extended oating car data collection, Expert Systems with
Applications, vol. 36, no. 3, part 1, pp. 42134227, 2009.
[3] S. Rass, S. Fuchs, M. Schaer, and K. Kyamakya, How to
protect privacy in oating car data systems, in Proceedings
of the 5th ACM International Workshop on Vehicular Inter-
Networking (VANET 08), pp. 1722, September 2008.
[4] P. Day, J. Wu, and N. Poulton, Brieng Note on Floating Car
Data, ITS Internacional, 2006.
[5] M. A. Sotelo, J. Nuevo, L. M. Bergasa, M. Oca na, I. Parra,
and D. Fern andez, Road vehicle recognition in monocular
images, in Proceedings of the IEEE International Symposium
on Industrial Electronics (ISIE 05), pp. 14711476, June 2005.
[6] M.

A. Sotelo and J. Barriga, Blind spot detection using vision
for automotive applications, Journal of Zhejiang University:
Science A, vol. 9, no. 10, pp. 13691372, 2008.
[7] C. J. Christopher Burges, A tutorial on support vector
machines for pattern recognition, Data Mining and Knowl-
edge Discovery, vol. 2, no. 2, pp. 121167, 1998.
[8] N. Dalal and B. Triggs, Histograms of oriented gradients
for human detection, in Proceedings of the IEEE Computer
Society Conference on Computer Vision and Pattern Recognition
(CVPR 05), pp. 886893, June 2005.
[9] Y. Derong, Z. Yuanyuan, and L. Dongguo, Fast computation
of multiscale morphological operations for local contrast
enhancement, in Proceedings of the 27th Annual International
Conference of the Engineering in Medicine and Biology Society
(IEEE-EMBS 05), pp. 30903092, September 2005.
[10] E. R. Dougherty, An Introduction to Morphological Image
Processing, SPIE Optical Engineering Press, 1992.
[11] D. Balcones, D. F. Llorca, M. A. Sotelo, et al., Real-
time vision-based vehicle detection for rear-end collision
mitigation systems, in Proceedings of the 12th International
Conference on Computer Aided Systems Theory (EUROCAST
09), vol. 5717 of Lecture Notes in Computer Science, pp. 320
325, 2009.
[12] R. E. Kalman, A new approach to linear ltering and predic-
tion problems, Journal of Basic Engineering, Series D, vol. 82,
pp. 3545, 1960.
[13] B. Hassanabadi, L. Zhang, and S. Valaee, Index coded
repetition-based MAC in vehicular ad-hoc networks, in
Proceedings of the 6th IEEE Conference on Consumer Commu-
nications and Networking Conference, 2009.
doi:10.1155/2010/295016
Research Article
Improvement of Adaptive Cruise Control Performance
Shigeharu Miyata,
1
Takashi Nakagami,
2
Sei Kobayashi,
2
Tomoji Izumi,
2
Hisayoshi Naito,
2
Akira Yanou,
3
Hitomi Nakamura,
1
and Shin Takehara
1
1
Department of Intelligent Mechanical Engineering, Faculty of Engineering, Kinki University, 1, Takaya-Umenobe,
Higashihiroshima, Hiroshima 739-2116, Japan
2
Vehicle Development Division, Vehicle System Development Department, Mazda Motor Corporation, 3-1,
Shinchi, Aki-Fuchu, Hiroshima 730-8670, Japan
3
Department of Intelligent Mechanical Systems, Graduate School of Natural Science and Technology,
Okayama University, 3-1-1, Tsushima-naka, Kita-ku, Okayama 700-8530, Japan
Correspondence should be addressed to Shigeharu Miyata, miyata@hiro.kindai.ac.jp
Received 14 October 2009; Revised 7 June 2010; Accepted 24 August 2010
Copyright 2010 Shigeharu Miyata et al. This is an open access article distributed under the Creative Commons Attribution
cited.
This paper describes the Adaptive Cruise Control system (ACC), a system which reduces the driving burden on the driver. The
ACC system primarily supports four driving modes on the road and controls the acceleration and deceleration of the vehicle in
order to maintain a set speed or to avoid a crash. This paper proposes more accurate methods of detecting the preceding vehicle by
radar while cornering, with consideration for the vehicle sideslip angle, and also of controlling the distance between vehicles. By
making full use of the proposed identication logic for preceding vehicles and path estimation logic, an improvement in driving
stability was achieved.
1. Introduction
The number of trac accidents and injuries continues to
increase year by year, and annual trac fatalities in Japan
remain at over 7000. Under these conditions, there is an
urgent need for technologies which can mitigate the serious
damage caused by car accidents, as well as prevent the
accidents themselves. Some members of this study have
played active roles in a study group for the Advanced Safety
Vehicle (ASV), a project conducted by the Japanese Ministry
of Land, Infrastructure, Transport and Tourism, and are
involved with ASV research and development. This study
group has proposed and studied safety technologies such as
Adaptive Cruise Control (ACC) and Precrash Safety Systems
[17]. Some of these technologies have already been put into
practical use.
An ACC system maintains the vehicle at the speed set by
the driver, and when it detects a preceding vehicle travelling
at a slower speed than the drivers vehicle, it controls the
vehicle speed to match the speed of the preceding vehicle.
It also performs following control to maintain the level of
distance between vehicles which was set by the driver (a
distance proportional to the vehicle speed). There has been
past research for ACC systems aimed at designing a vehicle
following distance control systemusing linear approximation
and linear control logic [8]. However this research has not
succeeded in producing natural vehicle behavior that meets
the drivers expectations. In addition, while a method for
following a preceding vehicle on curves has been proposed
[9] based on the following distance measured by stereo image
processing and on the relative vehicle speed, this method is
limited to following distances in the range of 4 m22 m, and
it is not suitable for maintaining the following distance of
approximately 100 m that is needed for expressway driving.
When a vehicle is driving at high speed such as on an
expressway, sideslip occurs when the vehicle corners. It is
known that this sideslip operates toward the outside of the
turn when the vehicle is at slow speed, and toward the
inside of the turn when the vehicle is at high speed [10
12]. Research and development of 4 WS systems that steer
the rear wheels in order to reduce this sideslip and improve
driving performance have been carried out. However, no
attempts other than this paper have been found at creating
an ACC system with following control that incorporates
consideration of the vehicle sideslip. Our research team has
already submitted a patent application for this concept [13].
This paper describes the Adaptive Cruise Control system
(ACC), a system which reduces the driving burden on the
driver. The ACC system primarily supports the four driving
modes on the road that are described in Section 2.1, and
controls the acceleration and deceleration of the vehicle in
order to maintain a set speed or to avoid a collision. The key
to achieving intelligent ACC control is the method used to
detect and follow the preceding vehicle.
The use of obstacle detection equipment, which detects
the preceding vehicle by means of millimeter wave radar, is
already well established [1418]. When the obstacle detection
equipment detects a vehicle (obstacle) in the path ahead of
the drivers vehicle, the turning radius of the drivers vehicle
(calculated based on the values from the yaw rate sensor
and steering angle sensor) is used to estimate the position of
the preceding vehicle after the next radar scan. The system
then judges whether or not the estimated position of the
preceding vehicle matches the position detected after the
radar scan. As long as the turning radius of the drivers
vehicle does not change signicantly between the radar
scans, then calculating the amount of oset of the detected
preceding vehicle from its path center line allows the future
position of the preceding vehicle to be easily estimated. In
particular when the speeds of the drivers vehicle and the
preceding vehicle are approximately the same, the amount
of this oset can be assumed to be unchanging between
radar scans, making it possible to easily predict the position
of the preceding vehicle. However as described above, on
expressways where the vehicles are travelling at high speed,
if there is a curve where the turning radius of the drivers
vehicle changes between radar scans, the sideslip that occurs
when the vehicle corners causes the amount of oset from
the preceding vehicle path center line to change. As a result,
it becomes impossible to accurately estimate the position of
the preceding vehicle, and the drivers vehicle accelerates,
producing a potentially dangerous situation.
This ACC system that incorporates consideration of the
sideslip angle makes it possible for one vehicle to reliably
follow the preceding vehicle by performing the following two
operations. First, the system estimates the path (the radius
of curvature of the path center line) by determining the
cornering radius of the vehicle based on the detected yaw
rate, steering wheel angle, and vehicle velocity. Second, the
system judges whether or not the current detected vehicle
is the same as the previous detected vehicle by comparing
the position of the detected vehicle with a position that
is estimated based on the calculated oset from the center
line of the path each time a radar scan occurs. Usually, the
radius of curvature of the path center line is assumed to be
unchanged during the period between radar scans. However,
in cases of curves near an entrance or exit, the radius of
curvature of the path actually does change during the period
between radar scans. In these cases, if the oset is assumed
to be unchanged, the radar fails to lock onto the preceding
vehicle. Therefore, because the position of the preceding
vehicle relative to the drivers own vehicle is inuenced by
the sideslip of vehicle, the oset needs to be corrected at
each radar scan based on the most recent detected cornering
radius.
For the following operation, ACC performs control in
order to maintain a constant distance. However, for example,
when the preceding vehicle is traveling close to the second
vehicle ahead, and if the preceding vehicle brakes and
accelerates suddenly and repeatedly (jerky motions), the
drivers own vehicle is forced to perform the same jerky
motions. As a result, such a system cannot be expected to
provide good ride quality and a feeling of safety. Therefore, it
is necessary to change from velocity control which maintains
a constant distance between the drivers vehicle and the
preceding vehicle to velocity control which maintains a
constant distance between the drivers vehicle and the second
vehicle ahead. In this way, even if the preceding vehicle
performs repeated jerky motions, the drivers own vehicle is
not forced to perform sudden braking and acceleration. This
results in a denite improvement to the ride quality and the
feeling of safety.
From the above points of view, this paper proposes a
more accurate method of detecting the preceding vehicle by
radar, and of controlling the distance between vehicles. This
proposal also for the rst time includes consideration of the
vehicle sideslip angle. These methods are expected to result
in improved driving stability.
Section 2 of this paper describes the ACC system
conguration. Section 3 explains the functions related to
velocity control. Evaluation for the ACC system based on
the experiment results and conclusions are presented in
Section 4.
2. ACC SystemConguration
2.1. Primary ACC Modes. The ACC system supports four
control modes, which are described below and shown in
Figure 1.
(1) Constant velocity control: when there are no vehicles
straight ahead, or when there is a large distance
between the drivers vehicle and the preceding vehi-
cle, the system maintains a constant vehicle velocity.
(2) Deceleration control: when a vehicle traveling ahead at
a slower speed is detected, the system uses the throttle
to decelerate the drivers vehicle. If this deceleration
is insucient, the system uses the brake to decelerate
the vehicle.
(3) Following control: when the drivers vehicle is fol-
lowing behind the preceding vehicle, the system
controls the throttle and brake so that the time
interval between the vehicles (which corresponds to a
distance between the vehicles that is proportional to
the velocity of the drivers vehicle) is the time which
was set by the driver.
(4) Acceleration control: when, due to a lane change, there
is no longer a vehicle ahead of the drivers vehicle, the
system accelerates the vehicle up to the velocity set
60 km/h

80 km/h
80 km/h
60 km/h
60 km/h 60 km/h
50 km/h
Set velocity
80 km/h
60 km/h
50 km/h
(
1
)
C
o
n
s
t
a
n
t
v
e
l
o
c
i
t
y
o
p
e
r
a
t
i
o
n
(
2
)
D
e
c
e
l
e
r
a
t
i
o
n
c
o
n
t
r
o
l
(
3
)
F
o
l
l
o
w
i
n
g
o
p
e
r
a
t
i
o
n
(
4
)
A
c
c
e
l
e
r
a
t
i
o
n
c
o
n
t
r
o
l
Figure 1: Four modes supported by the ACC system.
by the driver, and then maintains a constant velocity.
Furthermore, when the drivers vehicle approaches a
vehicle ahead of it without slowing down enough, an
alert buzzer and display prompt the driver to apply
the brakes or take other appropriate action.
2.2. System Conguration. Figure 2 shows the ACC system
conguration and unit layout in a vehicle. This conguration
includes the existing Auto Speed Control (ASC), Dynamic
Stability Control (DSC), Millimeter Wave Radar used to
detect objects ahead, ACC Electrical Control Unit (ACC
ECU) used to calculate control values for the engine and
brake, Distance Setting Switch used to set a time interval
for following the preceding vehicle (in order to maintain a
constant distance), and Indicator/Display which informs the
driver of the control mode. Figure 3 shows the part layout for
the sensors, ECU, and other components. For brake control,
the DSC brake actuator function is extended to perform
feedback control, so that the deceleration demanded by the
driver is delivered. For engine control, the vehicle velocity set
by ACC is transmitted to the vehicle velocity control section
Millimeter wave radar
Distance setting switch
Brake lamp
Cruise control switch
Electronic throttle valve
Wheel speed sensor
Brake pressure sensor
Brake actuator
Steering angle sensor
Indicator / display
Meter
BCM ECU
DSC ECU
PCM ECU
DSC
ASC
CAN
CAN
Yaw rate/lateral G sensor
ACC ECU
Figure 2: System conguration.
ACC ECU
Brake pressure
sensor
Wheel speed
sensor
Steering
angle sensor
Yaw rate/
lateral G sensor
DSC
Brake
lamp
PCM
Cruise control switch
Distance setting switch
Millimeter
wave radar
Speedmeter
BCM
Figure 3: ACC unit layout.
of the existing ASC, and the system then controls the vehicle
velocity so as to follow the vehicle preceding it.
Values from the wheel speed sensor and yaw rate/lateral
G sensor are transmitted to the DSC ECU, and values for
the vehicle velocity and other items set from the operation
switches are transmitted to the PCM ECU. Furthermore,
the signals of both ECUs are collected at the ACC ECU via
CAN (Control Area Network). The transmissions between
the Millimeter Wave Radar and the ACC ECU are carried
out via CAN. In addition, when brake control is performed,
ACC controls the relay between the brake pedal switch and
the brake lamp in order to turn on the brake lamps.
As described above, rather than developing a new actu-
ator control for the engine and brake, instead the functions
of the existing system are extended, ensuring that this system
can be eciently developed.
Figure 4: View of the Millimeter Wave Radar.
Table 1: Millimeter Wave Radar performance.
Detection range
Max 150 m
Min 2 m
Resolution 0.1 m
Range rate
Max 200 km/h
Min 200 km/h
Resolution 0.36 km/h
Azimuth angle
Area 7.5 deg
Resolution 0.1 deg
Data rate 100 msec
2.3. Millimeter Wave Radar. Figure 4 shows an external view
of the Millimeter Wave Radar used in this system. The
frequency of the Millimeter Wave Radar is 76 to 77 GHz.
An FMCW system is used, allowing the distance between
vehicles and the relative velocity to be simultaneously
measured with very high precision. In order to detect the
horizontal angle, a mechanical scan system is used, in which
the antenna and millimeter wave transmitter/receiver are
panned back and forth to the right and left sides by a
motor. This system yields a relatively high angle resolution.
In addition, the thickness of this radar has been reduced
to 70 mm, which makes it possible to install it inside the
bumper, despite the mechanical scanning. The specications
of the millimeter wave radar are shown in Table 1.
3. ACC ECU
The main functions of the ACC system are selection of the
preceding vehicle and control of the distance between the
vehicles. Figure 5 shows a block diagram of the control logic.
The ACC ECU selects the preceding vehicle to follow
by utilizing the information (distance and relative velocity)
transmitted by the Millimeter Wave Radar. The system then
controls the acceleration and deceleration of the vehicle
based on control values such as the target vehicle velocity and
the target acceleration/deceleration.
3.1. Selection of the Preceding Vehicle during Cornering. The
Millimeter Wave Radar is used to detect the vehicles traveling
ahead of the drivers vehicle. When the road is straight,
the preceding vehicle can be easily identied as the vehicle
traveling ahead on the same path as the drivers vehicle.
However, it becomes more dicult to identify the preceding
vehicle when there are curves in the road. As shown in
Figure 6, when three vehicles are traveling ahead of the
drivers vehicle at a curve in the road, it is rst necessary
to determine which vehicle is on same path as the drivers
vehicle and which vehicles are not on the same path. Then
the vehicle to be followed can be correctly identied.
The system judges whether the object detected by the
radar is a relative static object or moving object by comparing
the velocities of the drivers vehicle and detected vehicle.
If the preceding vehicle is traveling at a speed that is
approximately the same, that vehicle is considered to be a
relative static object.
At the same time, the path (with radius of curvature R) is
estimated by determining the cornering radius of the drivers
vehicle based on the detected yaw rate, steering wheel angle,
and vehicle velocity.
The system judges whether or not the current detected
vehicle is the same as the previous detected vehicle each
time a radar scan occurs. This judgment is performed by
comparing the position of the detected vehicle with the
estimated position. Therefore the oset from the center line
of the path is recalculated at each radar scan in order to
estimate a new position for the detected vehicle after it has
moved.
Usually, the radius of curvature R of the path center line
is assumed to be unchanged during the period between radar
scans. However, in cases when the vehicle is on a curve near
an entrance or exit, the path radius of curvature actually does
change between radar scans. If the sideslip is assumed to be
unchanged, the system fails to lock onto the target vehicle.
Therefore, because the relative position of the preceding
vehicle to the drivers vehicle is inuenced by the sideslip of
the vehicle, the oset needs to be corrected at each radar scan
based on the most recent detected cornering radius.
3.2. Basic Logic. The systemestimates the radius of curvature
for the path of the drivers vehicle based on the yaw rate
and vehicle velocity. When the vehicle is cornering, the
relationship among the velocity V, the yaw rate r, and the
cornering radius R is obtained as follows:
R =
V
r
. (1)
Figure 7 shows the geometric relationships between the
drivers vehicle and the vehicle immediately preceding it. The
arc of radius R is the trac lane of the drivers vehicle. If the
oset , which is the distance between the trac lane and
the preceding vehicle, is within a certain range, the preceding
vehicle is determined to be traveling on the same path. Here,
if the distance to the vehicle immediately preceding is d,
and the irradiation angle of the radar is , the following
relationship can be derived:
{R ( d)}
2
= R
2
d
2
. (2)
dr, vr
Vt
vr
vL
Target
headway
distance
generator
Selection
leading
vehicle
Target
acceleration/
deceleration
generator
Vehicle model
Plant
Feedback
compensator
Feedback
compensator
Vr
Vr
dL
dt
+
Gc Gt
Gt
Gr
Vc
PCM
DSC
ACC C/M
Figure 5: Control block diagram.
FOV of radar
Predicted lane
Mazda MPV
Detected vehicle
Other objects
out side of road
- Detect several objects by radar
- Predict own lane by Yaw rate
sensor and vehicle velocity.
- Decide target vehicle based on
the Predicted lane and position
moving object
Obstacle detection and path estimation
Leading vehicle
Figure 6: Detection of preceding vehicle and estimation of path.
Assuming that ( d) is suciently smaller than the radius
R, the oset can be derived as follows:
= d +
d
2
2R
= d
_
+
d
2R
_
. (3)
3.3. Introduction of the Sideslip Angle. In actual cases when a
vehicle is traveling at a certain velocity on the road, a sideslip
angle occurs at the vehicle during cornering. When the trac
lane curves, this sideslip result in a deviation in the range
of irradiation. Correcting the range of irradiation can be
expected to improve the oset accuracy.
A dynamic model of the vehicle shown in Figure 8 can be
expressed as follows [12]:
2
_
K
f
+ K
r
_
+
_
mV +
2
V
_
l
f
K
f
l
r
K
r
_
_
r = 2K
f
,
2
_
l
f
K
f
l
r
K
r
_
+
2
V
_
l
2
f
K
f
+ l
2
r
K
r
_
= 2l
f
K
f
.
(4)
R
R
d
d
Figure 7: Position of preceding vehicle during cornering without
considering the sideslip angle.
Using (4), the sideslip angle can be calculated as follows:
=
_
1
m
2l
l
f
l
r
K
r
V
2
_
l
r
Rl
. (5)
Rear tire
r: Yaw rate
Front tire
C.G.
f
l
r
l
f
Figure 8: Two-wheel model of a vehicle.
Table 2: Vehicle system parameters.
m Vehicle mass
V Vehicle speed
C.G. Center of gravity
l
f
Distance between front axle and C.G.
l
r
Distance between rear axle and C.G.
K
f
Front cornering power
K
r
Rear cornering power
r Yaw rate
Sideslip angle
f
Tire angle
From the relationship between the sideslip and the
cornering radius shown in Figure 9, the amount of center
travel L can be obtained as follows:
L = R =
_
1
m
2l
l
f
l
r
K
r
V
2
_
l
r
l
. (6)
Based on the relationships shown in Figure 9 and considering
the sideslip element, the following relationship can be
expressed in nearly the same way as in Figure 7:
_
R
_
d
_
+
___
2
= R
2
d
2
. (7)
Assuming that ( d( + )) is suciently smaller than
the radius R, the oset can be obtained as follows:
= d
_
+
_
+
d
2
2R
= d
_
+
d + 2L
2R
_
.
(8)
The absolute amount of center travel L is proportional to
the square of the vehicle velocity as shown in (6). Therefore,
here L can be expressed as follows:
L = aV
2
+ b. (9)
Parameters a and b are then identied by using the actual
experiment data. Figure 10 shows the relationship between
the vehicle velocity V and the amount of center travel L.
Figure 11 shows the relationship between the amount of
correction d and the cornering radius R. It can be seen that
L
d( + )
R
R
Figure 9: Position of preceding vehicle during cornering, with

consideration for the sideslip angle.
2
1
0
1
2
3
4
0 20 40 60 80 100
Vehicle velocity (km/h)
C
e
n
t
e
r
t
r
a
v
e
l
L
(
m
)
Figure 10: Relationship between amount of center travel L and the
vehicle velocity.
the smaller the cornering radius is, the larger the amount of
correction becomes.
The system judges whether the objects detected by the
millimeter wave radar are relative static objects, which
include a preceding vehicle traveling at the same speed, or
relative moving objects. This judgment is performed by using
the velocity of the objects relative to the velocity of the
drivers vehicle. The driving path for the drivers vehicle is
estimated from the current vehicle conditions such as the
vehicle velocity, steering angle, and yaw rate. The target
preceding vehicle on the estimated path is selected based on
the position of the preceding vehicle relative to the estimated
driving path.
In Figure 12, the green lines indicate the estimated path.
Despite the fact that there are many reectors at the side of
road and a vehicle traveling in the adjacent lane, the system
has reliably locked onto the preceding vehicle. Figure 13
shows an actual scene in which the preceding vehicle can be
selected quickly by using the estimated path and the position
of the target vehicle.
3.4. Control of the Distance between Vehicles. The logic
described in Section 3.3 makes faster lock-on times possible,
allowing the ACC ECU to smoothly control the velocity of
0
0.02
0.04
0.06
0.08
0.10
0.12
0.14
0.16
0.18
100 300 500 700 900
C
o
r
r
e
c
t
i
o
n
d
(
m
)
Cornering radius R (m)
Figure 11: Relationship between amount of correction d and
cornering radius R (vehicle velocity: 7075 km/h).
Preceding vehicle
(m)
(m)
Estimated path
Vehicle in adjacent lane
Scanning area
Reectors
20 15 10 5 0 5 10 15 20
20
40
60
80
100
Figure 12: Estimation of trac lane and recognition of the
preceding vehicle.
the drivers vehicle in order to maintain a constant distance
between the vehicles. Furthermore, it allows the radar to
maintain a continuous lock on the preceding vehicle.
The target vehicle velocity and target acceleration/decel-
eration are calculated using the time interval which was
set by the driver, the current distance between the drivers
vehicle and the preceding vehicle, and the current relative
velocity of the vehicles. Based on the dierence between
the target velocity and the current velocity, the velocity
can be controlled so that the distance between vehicles
gently converges on the target distance. Acceleration and
deceleration are performed in the same way as when the
vehicle is operated by the driver. As a result, the driver does
not experience any discomfort. Figure 14 shows the changes
in the relative velocity when the drivers vehicle is traveling
at a speed of 85 km/h and approaches a preceding vehicle
traveling at a constant speed of 60 km/h. It can be seen that
the control of acceleration and deceleration by the ACC ECU
is almost the same as when the vehicle is operated by the
driver. In fact, the ACC ECU control can been seen to be
smoother than driver control.
View from the millimeter wave radar
Predicted path
Leading
vehicle
Vehicle
Frontal view and detected data overlay
The leading vehicle can be selected quickly using path and
position of the moving objects.
Figure 13: Actual case in which the preceding vehicle can be
selected quickly by using the path and the position of the target
vehicle.
30
22
14
6
2
10
16 8 0 8 16 24 32 40 48
Target headway distance deviation (m)
R
e
l
a
t
i
v
e
v
e
l
o
c
i
t
y
(
k
m
/
h
)
by ACC
by driver
Figure 14: Results of driving test.
4. Conclusions
Improving the preceding vehicle lock-on performance by
improving the Millimeter Wave Radar unit and making full
use of object identication logic and path estimation logic
resulted in improved driving stability. This performance was
achieved even on an expressway with continuous sharp turns
in a mountainous region, or an expressway with multiple
lanes and heavy trac in an urban area. The preceding
vehicle identication performance was approximately the
same as or better than the performance of other such systems
in the industry.
For the acceleration/deceleration performance, which
has a large eect on occupant comfort, when the lane was
changed during tracking and the preceding vehicle sped up,
the acceleration performance satised the need for a smooth
feeling of acceleration with a short response delay.
Conversely, when the drivers vehicle caught up with
the preceding vehicle, a smooth and comfortable feeling of
deceleration was achieved.
This Adaptive Cruise Control system was developed for
the purposes of driving safety and comfort. It reduces the
number of brake and switch operations that are required
of the driver. As a result, the system reduces the driving
burden so that the driver can drive in comfort. The system
demonstrated sucient lock-on, tracking, and accelera-
tion/deceleration performance, and the system was able to
provide a satisfactory driving experience for the driver.
References
[1] Y. Yamamoto, T. Terano, and T. Nakagami, Development of
Mazda radar cruise control system, Tech. Rep. 24, MAZDA,
2006.
[2] Y. Yamamura, Y. Seto, H. Nishira, and T. Kawabe, An ACC
design method for achieving both string stability and ride
comfort, Journal of System Design and Dynamics, vol. 2, no.
4, pp. 979990, 2008.
[3] H. Soma, Y. Shiraishi, T. Watanabe, Y. Takada, and Y. Takae,
Trust in low-speed adaptive cruise control systemsanalysis
of trust structure, Review of Automotive Engineering, vol. 26,
no. 2, pp. 211212, 2005.
[4] Y.-S. Kim and K.-S. Hong, An IMM algorithm for tracking
maneuvering vehicles in an adaptive cruise control envi-
ronment, International Journal of Control, Automation and
Systems, vol. 2, no. 3, pp. 310318, 2004.
[5] H. Fukuoka, Y. Shirai, and K. Kihei, Driving support system
adaptive to the driver state of surrounding vehicles: simulation
study on a rear-end precrash safety system, Transactions of the
society of Automotive Engineering of Japan, vol. 40, no. 3, pp.
933938, 2009.
[6] K. Fujita and S. Tokoro, Pre-crash safety, Journal of Society
of Automotive Engineerings of Japan, vol. 59, no. 12, pp. 8590,
2005.
[7] Z. Sun, G. Bebis, and R. Miller, Monocular precrash vehicle
detection: features and classiers, IEEE Transactions on Image
Processing, vol. 15, no. 7, pp. 20192034, 2006.
[8] K. Adachi, Proposal of a target headway distance method
and applying this method to a car for adaptive cruise
control system, Proceedings of the Japan Society of Mechanical
Engineers, no. 06-52, pp. 2328, 2006.
[9] N. Shimomura, A study on preceding vehicle tracking on
curved roads using stereo vision, Tech. Rep. PRMU97-27,
IEICE, 1997.
[10] T. Hiraoka, H. Kumamoto, and O. Nishihara, Sideslip angle
estimation and active front steering system based on lateral
acceleration data at centers of percussion with respect to
front/rear wheels, JSAE Review, vol. 25, no. 1, pp. 3742, 2004.
[11] X. Gao, Z. Yu, J. Neubeck, and J. Wiedemann, Sideslip angle
estimation based on input-output linearization with tire-road
friction adaptation, Journal of JSAE, vol. 48, no. 2, pp. 217
234, 2010.
[12] B.-C. Chen and F.-C. Hsieh, Sideslip angle estimation using
extended Kalman lter, Vehicle System Dynamics, vol. 46, no.
1, pp. 353364, 2008.
[13] H. Okazaki, H. Omura, T. Seto, and T. Nakagami, Obstacle
detection equipment of vehicle, Patent Application 2006-
106549(P2006-106549).
[14] J. Takizawa, K. Sakagami, M. Masuda, K. Yamada, and T.
Kyuma, A development of pre-crash safety system for a mini
sized vehicle, Transactions of Society of Automotive Engineering
of Japan, vol. 39, no. 2, pp. 1520, 2008.
[15] N. Shimomura, A. Nakamura, T. Goto, K. Fujimoto, and
H. Shitsu, A method of tracking a forward vehicle using a
scanning laser radar and a camera, Transactions of the Institute
of Electrical Engineers of Japan C, vol. 123, no. 8, pp. 1427
1438, 2003.
[16] H. Miyazaki, S. Arita, and W. Ishio, Development of laser
radar for automotive application, Journal of Society of Auto-
motive Engineering of Japan, vol. 60, no. 5, pp. 4548, 2006.
[17] K. Osugi, K. Miyauchi, N. Furui, and H. Miyakoshi, Devel-
opment of the scanning laser radar for ACC system, DENSO
Technical Review, vol. 6, no. 1, pp. 4348, 2001.
[18] M. ABE, Motion and Control of Vehicle, Sankaido, Tokyo,
Japan, 2nd edition, 2007.
doi:10.1155/2010/169503
Research Article
Reducing Congestion in Obstructed Highways with Trafc Data
Dissemination Using Ad hoc Vehicular Networks
Thomas D. Hewer,
1, 2
Maziar Nekovee,
1, 3
and Peter V. Coveney
2
1
Department of Computer Science, UCL, Gower Street, London WC1E 6BT, UK
2
Centre for Computational Science, UCL, 20 Gordon Street, London WC1H 0AJ, UK
3
Mobility Research Group, BT Research, Polaris 134, Adastral Park, Martlesham IP5 3RE, UK
Correspondence should be addressed to Thomas D. Hewer, t.hewer@cs.ucl.ac.uk
Received 1 December 2009; Revised 14 June 2010; Accepted 5 July 2010
Academic Editor: Shahrokh Valaee
Copyright 2010 Thomas D. Hewer et al. This is an open access article distributed under the Creative Commons Attribution
cited.
Vehicle-to-vehicle communications can be used eectively for intelligent transport systems (ITSs) and location-aware services.
The ability to disseminate information in an ad hoc fashion allows pertinent information to propagate faster through a network.
In the realm of ITS, the ability to spread warning information faster and further is of great advantage to receivers. In this paper
we propose and present a message-dissemination procedure that uses vehicular wireless protocols to inuence vehicular ow,
reducing congestion in road networks. The computational experiments we present show how a car-following model and lane-
change algorithm can be adapted to react to the reception of information. This model also illustrates the advantages of coupling
together with vehicular ow modelling tools and network simulation tools.
1. Introduction
In the realm of vehicle-to-vehicle (V2V) communications
there are several methods for the dissemination of data that
are being actively researched. The use of satellite commu-
nication such as those linked to global positioning services
(GPS) oers global communication but requires expensive
equipment, large antennae (for two-ray transmission) and,
due to the large distances the signal must travel, they will
have a high latency. Cellular telephone networks oer a
lower latency over large distances but are still slow when
communicating with nearby vehicles, due to a centralised
approach, and require cellular contracts to use the network.
The scenarios presented in this paper require high-speed
communication disseminated from source, which is dicult
to achieve using either cellular or satellite communications.
Ad hoc networks oer a good method to spread
information outwards from an origin quickly and eciently.
It has been shown by Nekovee [1] that in ad hoc networks
worms spread in an epidemic pattern that can be modelled.
Using such modelling techniques we can develop algorithms
that allow for a change to be made to the speed, position and
route of a vehicle. A further advantage of ad hoc networking
is the unlicensed use of the radio-spectrum and the recent
reduction in cost for the equipment for communication.
A separate strand of the research being undertaken on
wireless delity (under IEEE standard 802.11) has been
developed in the past few years specically for vehicular
ad hoc networks (VANETs). The 802.11p WAVE standard
species network protocols which address the diculties
associated with vehicular networks. These diculties involve
short link times, delay tolerance, and the ineciency of
wired-network paradigms that have been inherited into the
wireless standards.
The following simulation experiments and algorithms
were developed with particular scenarios in mind. These
scenarios operate in a dual-carriageway environment with no
junctions together with an obstacle or danger that is present
at some point in the eld. The simulations use both vehicular
ow modelling and message propagation to advise the
vehicles of the obstacle at a greater distance than line-of-sight
provides. The results show that by spreading information
quickly and eciently through the network we can develop
systems to reduce congestion and limit negative vehicular
ow eects. By coupling the telecommunications and driver
model in one tool we can perform these simulations in the
same runtime. This method of simulation can also be useful
for intelligent transport systems, where simulations run in
parallel can sweep the parameter space for a desired outcome.
The impact of applications such as this is to reduce the
time spent in congested vehicular ow and also to increase
the safety interval and reduce collisions. Many possible
applications of VANET technology are designed to improve
safety of life while driving, and in this work we show that the
application of our novel algorithms will realise this goal.
The paper is organised as follows. Section 2 shows some
related work in VANET and ITS applications. Section 3
presents the simulation system we have used for this work
and the models that have been incorporated. Section 4
presents the algorithms for the intelligent driver model
and the lane-changing decision, and also the epidemic
propagation modelling for dissemination of information. We
then present the simulation scenario in Section 5. Section 6
shows the studies we have explored and the results on
vehicular ow through the simulated system. The paper ends
with conclusions in Section 7.
2. Related Work
There are many examples of work using vehicular network
simulation to test the applications of wireless technology.
Yin et al. [2] use simulation to evaluate the performance of
dedicated short-range communication (DSRC) and Eichler
et al. [3] analyse the impact of V2V messaging on vehicular
ow. Torrent-Moreno [4] has produced a great deal of
work on safety applications for vehicular communications
networks, providing empirical comparison of results under
dierent network settings and providing mechanisms for the
delivery and dissemination of vital data.
There are tools available for performing coupled simu-
lations of vehicular mobility and network communication
including VGrid [5] which has been designed for use
mainly with V2I systems, and MobiReal [6] which produces
scenario data for accurate mobile ad hoc network (MANET)
simulations, similar to the work of Wellnitz et al. [7]. These
systems are useful but we found that the ability to take
the best from both aspects and combine them in a highly
congurable and extensible simulation system, was best met
by the vehicular ow application of Treiber. The work of
Chen et al. [8] is similar to ours for a dierent vehicular ow
and accident scenario. We explore our data deeply so that we
can validate against this established work.
3. Simulation System
The simulation tool we use is adapted from the dynamic
vehicular ow simulator by Treiber et al. [9]. This tool
uses a simple model of a two-lane roadway, but contains
an advanced driver model and lane-changing algorithm,
MOBIL [10]. We have added telecommunications function-
ality to the original simulator for V2V communication. In
this section we show the underlying models that operate
within the simulation system.
3.1. Vehicular Modelling. To accurately simulate vehicular
behaviour, there are several key components: a driver model
to develop how real people will drive under certain circum-
stances, a lane-changing model to make realistic decisions on
when it would be advantageous to change lane and a roadway
with rules (e.g., drive on the left in the UK).
A car following algorithm will contain at least a desired
velocity, a safe time separation when following other vehicles
and acceleration and braking criteria [9]. At each simulation
time step the acceleration is calculated for each vehicle. The
parameters of these models can be changed to emulate more
aggressive and more considerate drivers, as required.
When modelling vehicular networks over large areas
(such as metropolitan areas) the ow of vehicles on a single
road behaves as an incompressible uid according to Q =
V, where is the average density of vehicles (cars/km)
and V is the average velocity on the road (km/h) [11].
At microscopic levels of simulation (across any eld size)
each vehicle is treated as an individual entity, which greatly
increases the computational requirements of the system but
provides a more realistic and component-based approach to
modelling.
The Intelligent Driver Model (IDM) in the simulator
follows the MOBIL model [10] developed by Treiber. MOBIL
operates as a car-following model such that the acceleration
and braking are dened by the distance from the car in front.
The function of such an acceleration dv/dt is shown in
dv
dt
= a
_
1
_
v
v
0
_
_
s
s
_
2
_
, (1)
where
s
= s
0
+
_
vT +
vv
2
ab
_
(2)
for acceleration on an open road a, velocity v, desired
velocity v
0
, distance s to front vehicle, desired dynamic
distance to front vehicle s
, velocity dierence v, a safe

time delay between vehicles T, a comfortable braking value
b, a minimum distance between vehicles s
0
and nally an
exponent which is adjusted in order to mimic real vehicular
ow patterns.
Lane changing algorithms add a necessary level of
complexity to the IDM. In order to decide whether to change
lane or not, the current acceleration must be calculated for
the current lane and the acceleration in the new lane (with
regard to the car behind and in front in the original lane).
If the acceleration in the new lane is greater than that in
the current lane, there is an advantage to be gained by
changing lane. Many models, including those in the original
simulator, include a bias in the model for particular lanes,
which simulates the real scenario of slow lanes and fast lanes.
3.2. MAC Layer Protocol. In simulating wireless delity
networks, the majority of simulations use the IEEE 802.11
protocols [12], as these oer the best simulation of the
MAC layer functionality. The IEEE 802.11 MAC layer uses
a distributed coordination function (DCF), which has been
simulated in [13], to ensure ecient communication on the
medium, and implements controls to reduce collisions. More
recently the IEEE 802.11p standard has been tested in [2]
specically for intervehicular communications. This allows
the foundation of underlying strengths in the 802.11 suite to
be enhanced for vehicular networks.
The MAC layer in the simulator operates using an
adapted version of IEEE 802.11 which removes the inter-
frame spacing (IFS) model, enabling equal priority for all
network trac. Due to implementation, and the need for
simplicity in the model, our implementation of 802.11 does
not suspend the backo timer when the medium is busy
during that frame, as 802.11 does.
The network backo when the medium is busy X
operates as shown in [11]
X 2
n
[B
min
, B
max
], (3)
where n is the number of times it has previously had to
back o in succession. B
min
and B
max
are the minimum and
maximum possible back o time, respectively. B
min
is often
set at 0. The medium is dened as busy if any car within the
transmitters interference range, R
i
, is currently broadcasting.
Every car within the transmission range, represented by
R
c
, (which is usually twice as small as the interference range)
will receive the message with probability .
3.3. Radiowave Propagation. In the modied simulator the
reception of messages is performed by a basic algorithm
controlled by the simulation engine. In advanced network
simulators the realistic reception of messages depends on
the signal strength at the receiver. The basic propagation
model is the Friis free-space calculation, which extends the
ideal free-space propagation formula (Pr 1/d
2
) [14]
where Pr is received power and d is the distance from
transmitter) to incorporate the antenna gain(both transmit
and receive). The Friis model, however, will only hold true
with a clear line-of-sight (LOS) between transmitter and
receiver, and assumes no level of multipath or shadowing of
the radiowave (which becomes very apparent in urban and
highway scenarios [15]). Friis operates as shown in
Pr(d) =
P
t
G
t
G
r
2
(4)
2
d
2
L
, (4)
where Pr is the received power at distance d with respect to
the antenna gain and height and the system loss L.
One method of altering the propagation of messages
through the simulated network, is to change the transmitted
power and therefore the transmission range. By transmitting
information further the message has a greater probability
of retransmission in sparse networks and also a faster dis-
semination through the system [16]. The main disadvantage
of having a large transmission range is that information
propagates further, eventually reaching a point where the
information usefulness is low, and so takes up bandwidth
and time for almost redundant data. In a system with limited
time and bandwidth this can cause localised problems, where
more pertinent (i.e., geographically closer) information is
lost to redundant data.
4. Algorithms
This section examines the algorithms used in the simulation.
These algorithms form the basis for the work we present
here and have been designed specically for vehicle-to-
vehicle and VANET scenarios. We have added to the original
simulator, by Treiber et al., including telecommunications
functionality and adapting both the IDM and lane-changing
model, based on reactions to information received from an
obstacle or danger. From this, we are able to show that
more ecient information dissemination through a network
can increase vehicular ow and also reduce stop-and-go
formations (presented in our results in Section 6).
4.1. Epidemic Message Passing Algorithm. The propagation of
messages through a systemrequires an ecient delivery algo-
rithm. In our simulation we use a probabilistic information
dissemination protocol, which is fully dened in Nekovee
and Bogasons work [11]. We assume that all vehicles know
their location (via GPS technology) and that each message
contains information about location and time of creation.
To ensure that propagation does not extend to unnecessary
distances from the source each message has a Time-To-Live
(TTL) setting.
The algorithm allows for a reasonable amount of retrans-
mission and dissemination through the network and bal-
ances the relevance of the information with the distance from
the source. To this end, information can disseminate quickly
and eciently and also reduce information spreading to
vehicles which do not require the information (as discussed
previously, this can cause more pertinent information to be
lost).
The probability for rebroadcasting, P, as described by
Nekovee [17] is obtained from
P =
1 if N
f
or N
b
= 0,
1 exp
N
f
N
b
N
f
+ N
b
otherwise,
(5)
where N
f
and N
b
are the number of times the car has received
that particular message from front and back, respectively,
and A is a protocol parameter which controls redundant
transmissions. In the case of directional message propaga-
tion, (5) is modied such that if a message is propagating in
either direction it is only kept alive by nodes near the head or
tail of the group. In this case the rebroadcasting probability,
P, is computed from
P =
1 if N
k
= 0,
1 exp
_
A
N
k
N
k
+ N
k
_
otherwise,
(6)
where N
k
is the count of messages received fromthe direction
of message propagation (e.g., if k is forward, N
k
is the count
of messages received from vehicles in front), and N
k
is the
count of messages received from the opposite direction.
4.2. Variable Speed Limit. The initial algorithm we introduce
reduces the value of the desired velocity (v
0
in (5)) by a xed
amount when the vehicle has received the warning message.
This achieves an overall slowdown in the network which can
reduce the time delay between free-ow and gridlock (where
v = 0) at the obstacle. This particular algorithm change has a
transient eect on the network, such that the system will still
become congested over time. The idea for this came from the
London Orbital (M25) variable speed limit which operates
on parts of the motorway.
The value we reduce v
0
by is of great importance. Initial
tests showed that reducing the desired velocity too much
(i.e., a reduction of over 10 m/s) when infected, the eect
on the network was to cause congestion further back in the
system, such that gridlock (i.e., all vehicles in the eld of
simulation are static) occurs much sooner. By reducing the
value of v
0
by 2.7 m/s the network slowed well, and the time
delay between free-ow and gridlock was increased without
causing congestion further back in the network.
An important algorithmic change is the return to the
normal value of v
0
once the obstacle has been passed
geographically, otherwise the recovery from the obstacle will
take a greater amount of time. We did test a proportional
change in the desired velocity as the obstacle was approached,
but this provided little observable eect at great distances and
a highly-negative eect closer to the obstacle, as cars were
congesting more smoothly but to a greater extent.
The results of this algorithm change were both transient
and often detrimental to the overall velocity of vehicles in the
system, so the change was dropped from the nal algorithm
described in Section 4.3. The reason for this negative eect is
thought to be related to the size of eld we are simulating.
In future simulations we will explore a much larger eld,
where we expect the eect of the algorithm to become more
pronounced as the distance from the obstacle increases.
4.3. Lane Changing Algorithm. The original lane change
model, shown in (7), operates by determining an acceleration
advantage (toward a goal velocity) to be gained by changing
lane and then testing if a gain threshold is reached by
the advantage. Early incarnations of our changes to the
algorithm worked to forcibly increase the advantage if a
message had been received and the vehicle was in the
lane with the obstacle. This has some positive eects, but
can cause problems when the message propagates a large
distance back through the system. In the case of the message
propagating beyond the reasonable extent of the need to
change lane, this approach causes unnecessary congestion in
the adjacent lane to the obstacle resulting in total congestion
in a short time.
At the start of the simulation several static variables
are applied to the model. A changing threshold is applied
that indicates the increased acceleration the lane change will
yield; this is set by default at 0.3 ms
2
for cars in the eld.
The other value is the politeness factor which reduces the
overall calculated advantage and which simulates the care
drivers take when changing lane (i.e., the model may say it
is advantageous to change lane, but the driver may be too
polite or hesitant to do so).
The basic algorithm operates by calculating a value of
advantage (A), the disadvantage this causes to other (B) and
then calculates whether a function of these values reaches a
changing threshold. If the threshold is reached the vehicle
changes lane:
A = a
n
a
o
+ B, (7)
where a
n
is the acceleration in the new lane and a
o
is the
acceleration in the old lane. B refers to a weighting to keep
the vehicle in the slow lane, as operates in reality:
B = a
b(o)
a
b(n)
, (8)
where a
b(o)
is the acceleration of the car behind in the old lane
if the vehicles changes lane and a
b(n)
is the acceleration of the
car behind in the new lane if I change lane. These values are
then entered into the following equation to return true or
false to changing lane:
_
A p
_
B > T, (9)
where p is the politeness factor and T is the changing
threshold. The form of (9) is multiplicative so the values
of A and B have a signicant impact on each other. In the
following equations, these values are only calculated if the
vehicle has been infected with a message. If a vehicle is
ignorant, it will continue to use the algorithm in (9). The
initial change we made was to add a value to A in (10) as
such
_
(A + V) p
_
B > T. (10)
This is very much a brute force approach and as such does
not truly represent real driving in a highway, where the value
of V would increase as the vehicle approaches the obstacle
and drop to zero after the obstacle has been passed. This
proportional addition to A is shown in
=
P
o
(P
m
P
o
)
,
_
(A + ) p
_
B > T.
(11)
In this adaption of the original algorithm, the value is
calculated as a function of the location of the obstacle and
the vehicles distance to it. This value is capped at a maximum
(currently 20) to prevent unrealistic behaviour (i.e., cutting
in with zero safety headway), which means the eect is
noticeable but quite subtle, when compared to the brute force
method in (10). This means that as the vehicle approaches
the obstacle the incentive to change lane becomes greater,
reducing the appearance of congestion at the obstacle in the
same lane.
5. Simulation Scenario
In order to deliver usable and realistic results, the simulated
system must be an accurate representation of a real highway.
The models for vehicular ow and network transmission
have been introduced in previous sections, and the metrics
Time 0:49
#S : 29
#R : 136
#I : 24
Main inow
Truck percentage
Imposed speed limit
Transmission range
Time warp factor
2350 vehicles/h
120
0%
120 km/h
100 m
5 times
0
25
50
(
%
)
75
100
Message spreading
Time
15 s
Igonrant
Infected
Active
Backing o
Sleeping
Stier
Figure 1: A graphic showing the simulator, with the obstacle in the lower quadrant of the curve to the left.
gathered with the results obtained are shown in Section 6.
The system we used for our study is shown in Figure 1. The
road length is approx. 1400 m and has two lanes with vehicles
moving in the same direction and an obstacle at 720 m, in
the left-hand lane. Vehicles originate in the top-right of the
road and travel around to the departure point in the bottom-
right. The ow of vehicles entering the road can be varied
from 0 to 4000 vehicles per hour, split across the two lanes.
The desired speed (v
0
in (1)) can be varied between 20 and
140 km/h, which will directly aect the vehicular density as
greater velocity will require greater separation, according to
the IDM.
The transmission of data between vehicles is controlled
by them operating in ve distinct states: ignorant, where
the vehicle has received no data and is not transmitting,
informed, when the vehicle receives data from the network,
transmitting, to resend received data through the network,
backing o, when the vehicle wishes to transmit but the
medium is busy, sleeping, when the vehicle has forwarded
data but is waiting to make sure all neighbours have received
that data and stiing, when the vehicle has performed all the
reception/transmission required and therefore is no longer
active. The movement through these states, from ignorant
to stier, can be seen as a loop. The propagation methods
described in Section 6.3 control how the vehicle progresses
through the states in dierent ways. The transmission range
of the vehicles is also set globally for the simulation, between
0 and 200 m(it is possible to extend the transmission range of
vehicular networks beyond 200 mbut only by using very high
transmission power and assuming little interference). The
vehicle that is causing the obstacle (surrounded by several
car-lengths of safety barriers) constantly transmits without
moving to the sleeping or stier states, so that any vehicle
coming within range will receive the data, and then begin the
process of receive-and-forward.
In the next section we present the ndings of our work.
6. Simulation Studies
In this section the collection of metrics is discussed with
diagrams and presentation of the results. All the simulations
were run with varying settings, so that an appreciation of
all the situations that may occur (that we can control in
the simulator) can be established. It is important to note, as
previously mentioned, that in the initial studies we assumed
that all vehicles in the simulation are equipped with the
technology for message propagation. In reality this market
penetration will take many years to achieve, but many car
manufacturers are working on supplying this technology
soon [18].
In later studies, as shown in Section 6.5, we have explored
dierent rates of equipped vehicles in the simulation. This
requires that all parameters remain static whilst testing
the equipped vehicle rates, such that we can rule out any
inuence from transmission range, vehicular density and
vehicle velocity. For each run at dierent parameters we
tested equipped vehicle rates from 0% to 100% in increments
of 10% and performed 20 runs to average out the data and
remove any transient or artifactual data.
6.1. Velocity Experiments. These experiments test the eec-
tiveness of the algorithm as vehicle velocity changes, to see
if the algorithms are suited to an urban (slow) or highway
(fast) environment. When ignorant (no messages have been
received), the cars will still attempt to change lane to avoid
the obstacle, but only as part of the original lane-changing
algorithm, and so congestion builds up in a short amount
of time, for most simulations. Below a certain network load
the road will never become congested, so the vehicular load
of the experiment was varied for each experiment. The
vehicular load on the road was also set low enough so that
the algorithm could aect the ow of cars as, at high loads,
this would not be possible. To this end there is, in any system,
a critical value of vehicular load after which no action can
prevent or reduce congestion.
Some early simulations with low vehicular loads showed
that it is sometimes more ecient to be ignorant of the
obstacle, and this must be taken into account, as in this
case the best course of action is to drive normally, using the
normal algorithm.
Figures 2 and 3 show results from a sample of the
experiments we ran to test this theory. By varying all
0
0 100 200
Infected
Ignorant
300 400
Time
500 600 700 800 900 1000
20
40
N
u
m
b
e
r
o
f
c
a
r
s
e
x
i
t
i
n
g
60
80
100
120
140
180
160
Low velocity limit (50 km/h)-load 4400 cars/hour
Figure 2: Comparison of exit aggregate for infected and ignorant
simulations at urban velocities (below 50 km/h).
0
0 100 200
Infected
Ignorant
300 400
Time
500 600 700 800 900 1000
50
N
u
m
b
e
r
o
f
c
a
r
s
e
x
i
t
i
n
g
100
150
200
250
300
High velocity limit (120 km/h)-load 4400 cars/hour
Figure 3: Comparison of exit aggregate for infected and ignorant
simulations for motorway velocities.
the parameters available we found that certain velocities,
transmission ranges, and vehicular load had dierent levels
of eectiveness to the overall congestion in the system. In the
main the results showed that our algorithm always produced
a positive eect. In order to produce valid data we ran several
simulations with the same parameters and then averaged this
data, to show results for a typical case.
Figures 2 and 3 show the number of cars exiting the eld
in a simulation run as an aggregate over time. Both show an
advantage for infected cars using the advanced lane-changing
algorithm, but the advantage is greater at higher velocities,
where the vehicles have more distance between them for the
same vehicular load, meaning they can more easily change
lane.
We found, and Figures 2 and 3 corroborate this, that
the advanced algorithm increased the time before congestion
0
0 100
Communication on
Communication o
Accident position
200 300
Time
400 500 600 700
100
P
o
s
i
t
i
o
n
o
f
l
a
n
e
c
h
a
n
g
e
200
300
400
500
600
700
800
Position of lane change over time
Figure 4: Chart showing location on the eld of lane changes with
and without communication.
began to build up and then once congested, the infected
cars still moved through the system more eciently. By
running the simulation for 15 minutes we can monitor the
development of congestion in the system and how the ow is
aected by the changed algorithm.
The results show interesting behaviour, beyond the
reduction of congestion in the system. By analysing the
rst 57 minutes of simulation time, we can see that the
development of congestion is also slower once vehicles do
start to slow down. This is because of the algorithm moving
vehicles into the opposite lane to the obstacle, reducing the
load on the lane with the obstacle and therefore reducing the
number of stopped vehicles behind the obstacle which, when
changing lane, cause a dramatic slowdown in the new lane.
This reduction in stop-and-go vehicular formation is also
seen elsewhere in the eld when the cars are infected with
the warning message and switch to our adapted algorithm.
6.2. Position of Lane Change. A factor that aects the build-
up of congestion in the system is related to the location
of the lane change. The following results show where the
lane change occurs with no communication and then using
our enhanced lane change algorithm with communication
active. The simulation settings were set at 2200 cars/hour
load, speed limit of 120 km/h and a transmission range of
100 m.
During the simulation the message propagates backwards
towards position 0 and, with the advanced algorithm, the
location of the lane-change also reduces. When the system
starts to slowand vehicular density increases, the lane-change
moves right back, causing a slower build-up of congestion
and a greater amount of free vehicular ow, as shown by
Figure 4. This does place greater load on the opposite lane
to the obstacle, but the reduction of stop-and-go behaviour
negates this. In Figure 4 the rapid reduction in position of
most lane changes between 395405 seconds and again at
0
0 100
Propagation o
Mixed
Edge detection
Distance detection
Simple ooding
200 300
Time (s)
400 500 600 700 800
10
F
l
o
w
p
r
o
l
e
(
c
a
r
s
o
u
t
/
c
a
r
s
i
n
)
(
%
)
20
30
40
50
60
70
80
90
Flow prole comparing propagation methods
Figure 5: Comparison of message propagation methodologies.
475500 seconds represents a period when the propagation
of the message is continuous, and the periods of little change
(of lane change position) are due to reduced propagation of
the warning message. The initial peak of lane change position
between 0 and 35 seconds represents the initialisation of the
system, that cars can change lane very close to the obstacle
due to the road being less loaded. The data to produce this
result came from a single typical simulation, with parameters
set as per the previous paragraph.
6.3. Transmission Method Experiments. In this experiment
the available transmission methods are tested with the new
algorithm, to see how they aect the overall congestion in
the system. Figure 5 shows a simulation run until congestion
is present at the origin of the eld (i.e., position = 0).
The simulation is stopped when the congestion reaches the
origin as after this point the algorithm is not aecting
the vehicular ow. The values represent the proportion
of cars leaving the eld in relation to the cars entering
(Cars Exiting/Cars Arriving). This indicates how vehicles are
owing through the system, where an increase of the gradient
represents free ow and a decrease represents congestion.
The models shown in Figure 5 are simple ooding, where
the message is rebroadcast just once, edge detection which
is explained in Section 3.1, and distance detection, which
is a dierent probabilistic method and a mixture of edge
and distance detection. These models are all running in
the simulation, but the mixed (edge detection and distance
detection) oers the best simulation of a real epidemic
protocol.
As can be seen in Figure 5 the edge detection method
alone oers little improvement over no propagation, and
the simple ooding and distance detection methods oer a
good initial advantage (0200 seconds) but then suer very
fast congestion build-up. The mixture of edge and distance
detection algorithm, with our changes to the lane change
model oers excellent results keeping near free ow until
0
200
400
600
T
im
e
P
o
sitio
n
800
1000
1200
1400
0
100
200
300
400
500
600
0
A
v
e
r
a
g
e
v
e
l
o
c
i
t
y
10
20
30
40
50
60
Average velocity map with no transmission
Figure 6: Average velocities across the simulated highway without
transmission.
0
200
400
600
T
im
e
P
o
sitio
n
800
1000
1200
1400
0
100
200
300
400
500
600
0
A
v
e
r
a
g
e
v
e
l
o
c
i
t
y
10
20
30
40
50
60
Average velocity map with transmission active
Figure 7: Average velocities across the simulated highway with
active transmission
ca. 420 seconds, when the network then slows and starts to
congest, but this takes longer (ca. 250 seconds from the rst
slowdown) than the other algorithms.
The addition of this propagation method and the
changed algorithm prevent several congestion-causing situ-
ations to occur. The main situation avoided is when vehicles
are unable to change lane to avoid the obstacle and begin to
slow down, but then do change lane causing the cars behind
to slow. This cause has been seen to initiate the build-up
of congestion by earlier warning of the obstacle so that the
cars can change lane at a high velocity. Another behaviour
of the system is that when the congestion is initiated, the
cars will ll up behind the obstacle, unable to change lane.
With the adapted algorithm the extra incentive to change
lane means that the opposite lane lls rst and so vehicles
can still move, increasing the time before the whole system
becomes congested.
6.4. System Velocity. The average velocity through a system
is of great importance. If a higher average velocity can be
achieved the number of vehicles passing through the area of
congestion will be higher than if there is much slowing of
vehicles. Figures 6 and 7 show the average velocity calculated
for intervals of 10 metres on the x axis and an interval of 30
seconds across the y axis. Each point represents the average
velocity at that time/position interval. In both gures there
0
0 50 100
Time
150 200 250 300 350
200
400
P
o
s
t
i
t
i
o
n
600
800
1000
1400
1200
40% equipped vehicles
(a)
0
0 50 100
Time
150 200 250 300 350
200
400
P
o
s
t
i
t
i
o
n
600
800
1000
1400
1200
(b)
0
0 50 100
Time
150 200 250 300 350
200
400
P
o
s
t
i
t
i
o
n
600
800
1000
1400
1200
(c)
0
0 50 100
Time
150 200 250 300 350
200
400
P
o
s
t
i
t
i
o
n
600
800
1000
1400
1200
(d)
Figure 8: Vehicle trajectories through the system for (a) 40% equipped vehicles, where we see little change from 0% equipped vehicles, (b)
60% where some improvement is shown but congestion still builds up, (c) 70% where the improvement is noticeable and (d) 80% where
the trajectory is only slighty aected by the obstacle. Each coloured trajectory represents a single vehicle as it enters and leaves the simulated
highway.
is a noticeable slowdown as the vehicles pass the obstacle.
This can be accounted for by the IDM attempting to retain
a minimum safe distance between vehicles.
Figure 6 shows that after an initial even velocity through
the eld, congestion begins to build up at the position of
the obstacle between 100 and 200 seconds, which causes a
slowdown further back to position 0. By time of 330 seconds
the congestion has reached position 0, and the average
velocity falls from 2025 ms
1
to 05 ms
1
.
As can be seen in Figure 7, there is a uniform average
velocity before and after the obstacle during the whole period
of the simulation (10 minutes). This reinforces the results
from the other simulations and proves that there is a better
ow of vehicles through the network, as well as a reduction
in the build-up of congestion, when eective transmission of
the road condition occurs.
We note that for both gures there is a spike in velocities
(between time 010 and position 12001400). This is the
period when the rst cars are leaving the eld. As they have
no cars in front they can accelerate up to the full speed
limit unhindered. To remove this artefact future simulations
will have a warm-up period, with low vehicular load, that
initialises the eld.
6.5. Varying Penetration Rate of Radio-Equipped Vehicles.
In all of the simulation studies and results presented in
Section 5, we have assumed that 100% of the vehicles in
the simulated eld are radio equipped and therefore able
to receive and retransmit (where necessary) the data about
the upcoming incident or congestion. This is a somewhat
idealised view of how deep uptake of radio-equipped vehicles
will be. More likely, as vehicle manufacturers and radio
equipment providers begin releasing the products to operate
and support a VANET, the ratio of equipped vehicles will
increase. Due to the critical safety-of-life applications of
VANET, it is important to study the eciency and resilience
of simulated applications, to study the point at which they
become usable and reliable. This simulation study involved
the same system as in previous experiments, but here the
proportion of equipped vehicles is increased from 10% to
90%. The plots in Figure 8 show the ow trajectories of
vehicles in the system.
We ran each simulation several times, to ensure no
artifactual or erroneous data was included in our results,
but the results shown here represent only one simulation. As
can be seen at low penetration rates (10%40%) the vehicles
reach the obstruction and begin to congest, but between 60%
and 70%there is a noticeable improvement, or smoothing, of
the vehicles velocity through the eld of study. The gures
show that above 70%, where there is still some conges-
tion build-up around the obstruction, the propagation of
information to vehicles further away, and the change to the
lane-changing algorithm this triggers, allows for more steady
driving through the eld of study, and less congestion.
7. Conclusions
This paper has presented a simulation of a specic road
condition, that of an accident blocking a lane on a dual
carriageway. The simulation uses the coupled approach to
mobility modelling and network simulation in a single
process. The models and algorithms that represent the
vehicular ow and network trac have been implemented
according to well-known and standard models. We have
shown that adapting the algorithms when information about
the accident is received via wireless transmission, we can
reduce the build-up of congestion and increase the ow of
vehicles passing the obstacle, both in terms of quantity of cars
and average velocity.
The various algorithms we tested achieved overall
improvement in the majority of cases. In some simulations
the improvement was not only in the prevention of con-
gestion overall, but by also keeping a consistent average
velocity through the network, which helps to reduce the
eects of stop-and-go vehicular ow and smooth possible
congestion waves that emanate from the source backwards.
The driver model and lane-changing algorithms come from
well validated sources, and so the adaptions we have made
are highly realistic and can show the eect of even simple
changes (as in the brute force addition of a value to A).
In order to fully test our algorithms we ran numerous tests
with a wide variety of parameters, to test for any transient or
artifactual eects. From these repetitive runs we established
that most eects were long-lasting and that where those were
transient, this was only due to a vehicular load on the road
where congestion could no be prevented (as corroborated by
control tests).
In the common situation where the obstacle is temporary
(i.e., a vehicle malfunction), any reduction in congestion
build-up allows for the obstacle to be removed, before the
velocity of all vehicles behind the obstacle drops to zero. In
more complex roadways the advanced warning could also
lead to a change of route, so drivers can avoid the section
of road where the obstacle exists.
The real-world application of this algorithm is limited
with current technology, as the recommendation for lane
change can be ignored by a human driver. The current possi-
bility for this algorithmis for an advanced driver information
system, where the details of the obstruction are suggested and
displayed in-car. The modelling of penetration rates that we
performed (Section 6.5) also studies the obedience of human
drivers. In the future, when automated driving systems are in
place and more relied on, this algorithmwould forma part of
the plethora of functions and processes used to drive a vehicle
under computer control.
B uscher et al. [19] discuss the challenges of systems
such as ours, from a social perspective. The authors work
motivates the need for accurate human-behaviour modelling
in vehicular network simulation and what impact the social
aspects of mobility provide. The simulation system we
designed to perform our studies uses accurate mobility
modelling, but for a real implementation the social aspects of
the driver and the context in which they are owing becomes
essential.
In the size of eld simulated here we are between the
microscopic and macroscopic scale of simulation, which
is achieved seamlessly by the use of a coupled model of
simulation. The tool we performed the simulations with
was lightweight and so we could easily implement new
algorithms and protocols. In order to run more complex
networks we would require a more complex simulation
engine for vehicle ow. With this increase in complexity the
eld size will increase and therefore a more powerful network
simulator is required. The proof of concept that this paper
provides will lead to further work in this eld, including the
use of parallel and distributed computational resource.
Acknowledgments
M. Nekovee acknowledges the Royal Society for supporting
his work through an Industry Fellowship. T. Hewer acknowl-
edges the support of British Telecom as part of a CASE
Industrial PhD and funding from the UK Engineering and
Physical Sciences Research Council.
References
[1] M. Nekovee, Modeling the spread of worm epidemics in
vehicular ad hoc networks, in Proceedings of the 63rd IEEE
Vehicular Technology Conference (VTC 06), vol. 2, pp. 841
845, May 2006.
[2] J. Yin, T. Elbatt, G. Yeung et al., Performance evaluation of
safety applications over DSRC vehicular ad hoc networks,
in Proceedings of the 1st ACM International Workshop on
Vehicular Ad Hoc Networks (VANET 04), pp. 19, October
2004.
[3] S. Eichler, B. Ostermaier, C. Schroth, and T. Kosch, Sim-
ulation of car-to-car messaging: analyzing the impact on
road trac, in Proceedings of the 13th IEEE International
Symposium on Modeling, Analysis and Simulation of Computer
and Telecommunications Systems (MASCOTS 05), pp. 507
510, September 2005.
[4] M. Torrent-Moreno, Inter vehicle communications, acheiving
safety in a distributed wireless environment: challenges, systems
and protocols, Ph.D. thesis, Universit at Karlsruhe, 2007.
[5] J. Anda, J. LeBrun, D. Ghosal, C.-N. Chuah, and M. Zhang,
VGrid: vehicular ad hoc networking and computing grid for
intelligent trac control, Tech. Rep., University of California
at Davis, 2005.
[6] K. Maeda, T. Umedu, H. Yamaguchi, K. Yasumoto, and
T. Higashino, MobiREAL: scenario generation and toolset
for MANET simulation with realistic node mobility, in
Proceedings of the 7th International Conference on Mobile Data
Management (MDM 06), p. 55, May 2006.
[7] O. Wellnitz, S. Lahde, E. Eden, and L. Wolf, A scenario editor
for mobile ad hoc networks, Tech. Rep., IBR, Technische
Universit at Braunschweig, 2006.
[8] A. Chen, B. Khorashadi, C.-N. Chuah, D. Ghosal, and M.
Zhang, Smoothing vehicular trac ow using vehicular-
based ad hoc networking & computing grid (VGrid), in
Proceedings of the IEEE Intelligent Transportation Systems
Conference (ITSC 06), pp. 349354, September 2006.
[9] M. Treiber, A. Hennecke, and D. Helbing, Congested trac
states in empirical observations and microscopic simulations,
Physical Review E, vol. 62, no. 2, pp. 18051824, 2000.
[10] A. Kesting, M. Treiber, and D. Helbing, General lane-
changing model MOBIL for car-following models, Trans-
portation Research Record, vol. 1999, pp. 8694, 2007.
[11] M. Nekovee and B. B. Bogason, Reliable and ecient infor-
mation dissemination in intermittently connected vehicular
adhoc networks, in Proceedings of the 65th IEEE Vehicular
Technology Conference (VTC 07), pp. 24862490, April 2007.
[12] IEEE, IEEE standard for information technology-
telecommunications and information exchange between
systems-local and metropolitan area networks-specic
requirementspart 11: wireless LAN medium access control
(MAC) and physical layer (PHY) specications, IEEE
Std 802.11-2007, (Revision of IEEE Std 802.11-1999), pp.
C11184, June 2007.
[13] J. Weinmiller, H. Woesner, and A. Wolisz, Analyzing and
improving the IEEE 802.11-MAC protocol for wireless LANs,
in Proceedings of the 4th IEEE International Workshop on
Modeling, Analysis, and Simulation of Computer and Telecom-
munication Systems (MASCOTS 96), pp. 200206, 1996.
[14] R. Struzak, Radio-wave propagation basics, Tech. Rep.,
ICTP-ITU-URSI School on Wireless Networking for Develop-
ment, 2006.
[15] K. Laasonen, Radio propagation modeling, Tech. Rep.,
University of Helsinki, 2003.
[16] M. Torrent-Moreno, Inter-vehicle communications: assess-
ing information dissemination under safety constraints, in
Proceedings of the 4th Annual Conference on Wireless on
Demand Network Systems and Services (WONS 07), pp. 59
64, January 2007.
[17] M. Nekovee, Epidemic algorithms for reliable and ecient
information dissemination in vehicular ad hoc networks, IET
Intelligent Transport Systems, vol. 3, no. 2, pp. 104110, 2009.
[18] D. Waters, Connected cars promise safer roads, BBC News,
2007.
[19] M. B uscher, P. Coulton, C. Efstratiou, et al., Intelligent mobil-
ity systems: some socio-technical challenges and opportuni-
ties, in Communications Infrastructure: Systems and Applica-
tions in Europe, R. Mehmood, E. Cerqueira, R. Piesiewicz, and
I. Chlamtac, Eds., pp. 140152, 2009.
doi:10.1155/2010/753256
Research Article
Reliable Delay Constrained Multihop Broadcasting in VANETs
Martin Koubek, Susan Rea, and Dirk Pesch
NIMBUS Centre for Embedded Systems Research, Cork Institute of Technology, Cork, Ireland
Correspondence should be addressed to Susan Rea, susan.rea@cit.ie
Received 26 November 2009; Accepted 5 September 2010
Copyright 2010 Martin Koubek et al. This is an open access article distributed under the Creative Commons Attribution
cited.
Vehicular communication is regarded as a major innovative feature for in-car technology. While improving road safety is
unanimously considered the major driving factor for the deployment of Intelligent Vehicle Safety Systems, the challenges relating to
reliable multi-hop broadcasting are exigent in vehicular networking. In fact, safety applications must rely on very accurate and up-
to-date information about the surrounding environment, which in turn requires the use of accurate positioning systems and smart
communication protocols for exchanging information. Communications protocols for VANETs must guarantee fast and reliable
delivery of information to all vehicles in the neighbourhood, where the wireless communication medium is shared and highly
unreliable with limited bandwidth. In this paper, we focus on mechanisms that improve the reliability of broadcasting protocols,
where the emphasis is on satisfying the delay requirements for safety applications. We present the Pseudoacknowledgments
(PACKs) scheme and compare this with existing methods over varying vehicle densities in an urban scenario using the network
simulator OPNET.
1. Introduction
The US Federal Communications Commission (FCC) and
later the European Telecommunications Standards Institute
(ETSI) approved a frequency band reservation in the 5.9 GHz
(in Europe 5 GHz) band for wireless communications
between vehicles (V2V) and roadside (V2R) infrastructures.
At present, the IEEE group is completing the nal drafts
of the IEEE 802.11p and IEEE P1609 Standard for Wire-
less Access in Vehicular Environments (WAVEs) [1]. The
European Commission through programmes like the i2010
Intelligent Car Initiative [2], which is a followup of eEurope
2005 [3] is driving the rollout of intelligent vehicle systems
in both European and international markets, by supporting
ICT research and development in the area of transport.
Under i2010, eSafety is a collaborative initiative involving the
European Commission, industry, and other stakeholders to
hasten the development, deployment, and use of Intelligent
Vehicle Safety Systems (IVSSs) as a means of increasing road
safety and reducing the number of road trac accidents
within Europe.
Integrating a network interface, GPS receiver, sensors,
and on-board computers presents an opportunity to build
a powerful car-safety system, capable of gathering, process-
ing, and distributing information. By collecting accurate
and up-to-date information concerning the status of the
surrounding environment, a driver assistance system can
quickly detect potentially dangerous situations and notify
the driver regarding this impending peril. Notifying other
drivers can be achieved via vehicle-2-vehicle (V2V) commu-
nications typically relying on broadcasting as the underlying
dissemination technique. However, broadcasting is a very
expensive dissemination technique that needlessly consumes
channel communication capacity with increased collisions
and packets losses [4]. A broadcasting protocol for VANETs
must guarantee fast and reliable delivery of information
to all vehicles in the neighbourhood, where the wireless
communications medium is shared, very unreliable, and
with limited bandwidth. It must guarantee high delivery
rates for priority messages with emergency payload data
in all situations from small vehicle densities (rural areas)
to crowded roads in cities during peak times with the
communication network may be well saturated.
Broadcasting protocols (e.g., [59]), that have been
proposed for VANETs have a common factor in that they
cannot guarantee high reliability for safety-related data
dissemination with [5] concluding that the probability
of successful reception of the data decreases with grow-
ing distance from the sender. These factors have serious
consequences for safety-related data dissemination where
dangerous situations can be aggravated through unsuccessful
broadcast communications.
In this paper, we propose a scheme called Pseudoac-
knowledgments (PACK) that interprets successful multi-
hop broadcast transmission through overhearing successive
rebroadcasts by its neighbours. As the broadcast packet
traverses the network, each hop creates dynamic time slots
in order to rebroadcast. Intermediate hops that receive the
broadcast wait until the dynamic slot time expires and then
rebroadcasts thereby acknowledging a link between itself and
a previous hop. If the previous hop does not overhear the
rebroadcast, it repeats the rebroadcasting. The dynamic slots
are created locally at individual nodes and do not require a
global clock.
The advantage of the PACK method is that it does not
need any extra hardware and can be implemented on top of
any broadcasting protocol, however, our simulation results
have demonstrated that most gains in eciency are achieved
with location-based p-persistent CSMA/CA broadcasting
protocols. The PACK schemes rapidly increase reception
probability of broadcasting protocols with minimal addi-
tional overhead in terms of latency and retransmissions. In
this paper, we compare the eciency of the PACK method
with existing schemes for reliable multihop broadcasting that
increase the reception probability. The network simulator
tool OPNET [10] is used to develop an accurate urban
scenario based on the VANET specic WAVE communica-
tions protocol with realistic vehicle mobility patterns, radio
propagation model using 802.11p.
2. Related Work
One of the primary concerns for broadcast protocols lies in
the unreliable packet delivery. Protocols such as ALOHA and
CSMAare some of the earliest works that focus on mitigating
packet collisions in uncoordinated networks. Following on
from this CSMA with Collision Avoidance was developed
which is the basis for the IEEE 802.11 suite of communi-
cations protocols of which IEEE 802.11p for V2V commu-
nications is part of it. An RTS/CTS handshake exchange
mechanism has been developed for unicast transmissions
to increase reliability, however, broadcast transmissions still
have to rely only on pure CSMA/CA protocol without
RTS/CTS. A common concern for broadcasting algorithms
in VANETs is their inability to achieve a packet reception rate
close to 100% [5].
2.1. Multihop Broadcasting Schemes. For multihop broad-
casting protocols, several works have proposed acknowl-
edging techniques to increase reliability in MARQ [11],
BACK [12], and BSMA [13] schemes. These methods are
based on reserving time slots where a sender allocates
virtual time slots for all its neighbours and transmits the
broadcast data. All its neighbours transmit ACKs in their
virtual slot. The reserving of virtual time slots for individual
ACK transmissions is problematic in denser networks as
it leads to a dramatic increase in latency, a fundamental
concern for the dissemination of safety related data. The
authors in [6] proposed a broadcasting protocol called UMB
that uses a handshake like RTS, CTS and ACK for one-
directional broadcasting, however, this protocol requires
the positioning of intersection repeaters that acknowledge
the broadcast along the physical roadways. Other multihop
broadcast protocols presented in [7] include V-TRADE and
HV-TRADE. A node wishing to transmit or retransmit
a broadcast transmits a position request at rst to all
neighbours and waits until all neighbours reply. After all
replies have been received, the node transmits the broadcast
with a list of selected nodes that act as forwarders similar
to OLSR [14]. This was one of the earlier works to address
broadcasting in VANET, and the overhead incurred with the
position request and reply packets at each hop can contribute
to network congestion in denser networks and also increase
delay.
From best of our knowledge, there is no method to
increase broadcast reliability in multihop broadcast protocols
for VANET networks that do not suer from dramatically
rising latency and/or increased load on the physical medium
through numerous redundant transmissions. As a precursor
to presenting the proposed PACK method, we discuss the
mechanisms previously developed to increase reliability for
1-hop broadcasting.
2.2. 1-Hop Broadcasting Schemes. In [15], the authors have
identied protocols that increase the reliability of 1-hop
broadcasting schemes and have grouped the schemes based
on their channel access methods.
(i) The rst group is based on CSMA/CA where pro-
tocols (e.g., [1113]) use a handshake mechanism
comprising of short packets similar to RTS/CTS/ACK
packets.
(ii) The second group of protocols relies on reserving
time slots in the physical medium. For the RR-
ALOHA [16], vehicles must continuously exchange
2-hop information to reserve free time slots without
any central coordination units. The RR-ALOHA
was proposed within the European research project
CarTalk2000 [17].
(iii) The third group relies on the repetition of broad-
casting transmissions. The SFR [18, 19] protocol
randomly repeats broadcasted transmissions. The
authors in [15] propose the OOC code that dynam-
ically aects the number of repetition. The OOC
method performed better against SFR [15, 20], but
for fast moving vehicles the OOC protocol has
diculties with codeword synchronisation.
(iv) Another group of protocols not discussed in [15]
investigated changing the transmission power of
broadcasting messages to control the wireless band-
width [21, 22]. The Adaptive Transmission Power
(ATP) protocol [21] changes the transmit power
depending on the number of 1-hop neighbours.
In recent years, several 1-hop broadcasting schemes
have been developed for VANETs whereas not many eorts
were invested in improving existing multihop broadcasting
schemes described in Section 2.1. For safety-related data
dissemination, there will be a prerequisite to dissemination
data beyond a single hop with high reliability for data
delivery over several hops with minimal delay and low data
collisions. In this paper, we propose a multihop scheme that
(i) improves the reliability of multihop broadcast proto-
cols,
(ii) with a marginal increase in latency and link load.
The proposed approach is based on creating exible time
slots at the transmitter and the pseudoacknowledging of
transmissions by rebroadcasting nodes through overhearing.
We choose three 1-hop reception schemes namely RR-
ALOHA, SFR, and ATP that we have extended for use with
multihop broadcasting and compared their performance
with our proposed scheme we refer to as Pseudoacknowl-
edgments (PACKs). We tested the schemes under dierent
vehicle densities where we emulated local (accidents) and
global (raining) events. An urban environment has been
selected for experimental evaluation as opposed to a rural
or motorway scenario as this environment will be densely
populated with slower moving vehicles that force the use of
multihop broadcast protocols as transmission distances are
severely attenuated with obstacles present in the environ-
ment such as buildings, trac lights, and restricted roadways
that cause a build up and congestion in trac ows.
3. Multihop Broadcast Protocol
The methods for increasing multihop broadcast protocol
reliability have been overlaid on the same underlying base
broadcast protocol namely the low latency Slotted Restricted
Mobility-Based (SRMB) protocol as opposed to using ood-
ing. The SRMB protocol is an extension of the Restricted
Mobility-Based (RMB) [23] broadcasting protocol with
SRMB minimising data collisions on forwarding broadcasts
by using a dynamic slot wait time generated in the upper
MAC layer in the order of milliseconds. PACK can be used
with any broadcasting protocol, but dynamic slot wait times
(SRMB) have been shown to reduce collisions by modifying
the channel access times. Protocols, which include already
some form of slot wait times, for example, in case of AODV
[24] and OLSR [14] random wait time in the range of 0 to
100 ms, do not need necessary integrate SRMB extensions to
use PACK.
The RMB, SRMB, and PACK algorithms are described
next, prior to the presentation of the experimental evalua-
tions.
3.1. Restricted Mobility-Based (RMB) Protocol. We have
previously presented the RMB algorithm in [23]. RMB is
a at (nonclustered), uncentralised, p-persistent CSMA/CA
S1 S2
S3
S4
Figure 1: Directional sectors are dened about the transmitting
node with a radius dened by the theoretical transmission distance
with each sector having a 90
spread.
broadcasting protocol that reduces redundant broadcast
transmissions using 1-hop location knowledge obtained
from beacons. RMB was compared with the DV-CAST
protocol [25], with RMB having fewer transmissions, lower
end-to-end delay, and a high delivery ratio.
The basic principle of this algorithm is that before
broadcasting (rebroadcasting) a transmitter M
i
determines a
small set of its neighbours MPR
i
1N
(Multipoint Relay set as
used in OLSR [14]) with each node lying in a geographically
dierent sector (maximum N 4 sectors)
as shown in Figure 1. The transmitter records the
shortened MAC addresses of the MPR
i
1N
nodes in the
packet header and broadcasts. A node M
j
that receives the
packet and has its MAC address recoded in the packet header
assigns a Backo time slot 0 for rebroadcasting in the MAC
buer. A node M
k
, which receives the packet and nds that
its MAC address does not match any address recorded in
the packet header, assigns its Backo time slot depending on
its position, speed, and motion vector compared against the
transmitter in range from 1 to the maximum value of the
particular Contention Window (CW). The maximum size of
CW depends on the type of trac (voice, video, and data)
and ranges from 3 to 15 [1, 26]. A Backo time slot of 1
refers to nodes that are suciently far from the transmitter
and have similar speed and motion vector as the transmitter.
Larger Backo time values indicate that nodes have dierent
motion vectors and speeds compared to the transmitter [23,
Section 3].
To avoid redundant transmissions during broadcasting,
each node M
i
(MPR, non-MPR) assesses whether all of its
neighbours have received the broadcast packet. This is based
on the knowledge of the position of the transmitter and all
neighbours and the knowledge of transmission distance. If
all neighbours are assessed by the M
i
to have received the
broadcast and the M
i
has the same broadcast to transmit,
then M
i
discards the packet and does not rebroadcast.
The RMB scheme ensures that during broadcasting if
a collision occurs at an MPR node, some other non-MPR
node with the second highest priority substitutes as the MPR
and rebroadcasts. A strong advantage of this scheme lies in
t
1
t
2
t
2
t
3
t
3
Collision
Figure 2: RMB.
the fast broadcasting process where nodes wait for retrans-
mitting generally less than a millisecond. A disadvantage is
that the contention window size of a trac class may not
be suciently large enough to transmit without collisions at
non-MPRs, for example, considering the Voice trac class,
there are only 3 Backo time slots. This implies that non-
MPRnodes can, with a high probability, be assigned the same
time slot which leads to collisions, thus, eectively stopping
the broadcast.
3.2. SRMB and PACK. The RMB algorithm suers from the
hidden terminal problem as illustrated in Figure 2.
A source node M
i
broadcasts at time t
1
with MPR set
MPR
i
j,k
. All neighbours of M
i
that received the broadcast
and because MPR nodes (M
j
and M
k
) set the Backo time
to 0 and are not within transmission range of each other
(Hidden Terminal problem), M
j
and M
k
rebroadcast at
time t
2
. Because M
j
and M
k
transmit in the same time
(or within a short proximity), a collision occurs around
M
i
in t
2
. The collision generally does not have an eect
on surrounding nodes of M
i
because these nodes already
received the broadcast in time t
1
. But M
i
does not overhear
(receive) the broadcasts sent by M
j
and M
k
correctly and
so M
i
does not know if its own broadcast transmission was
successfully received at its neighbours.
To minimise collisions at the source node (and likewise
at intermediate nodes that act as forwarders), we developed
the Slotted Restricted Mobility-Based (SRMB) algorithm and
the Pseudoacknowledgments (PACK) scheme.
3.2.1. Slotted Restricted Mobility-Based (SRMB) Algorithm.
The main contribution of the SRMB extension is that
rebroadcasting is carefully scheduled (spread in time) using
dynamic slot wait times (Figure 3). Each node that receives
a broadcast packet assigns a dynamic wait time slot for
rebroadcasting to ensure that nodes have sucient time for
rebroadcasting. The wait time slot is derived from the max-
imum transmission time T
L MAC
(1) including processing at
lower MAC layer and the time needed for transmission
T
L MAC
(ac) =
L
DATA
R
DATA
+
D
c
+ SIFS
+ T
BoSlot
(AIFSN + CW[ac]).
(1)
(The equation is valid only for lightly loaded networks, In
busier networks, if a transmission is heard while a node is in
Backo, then the new Backo time is set and transmission
delay (1) is increased.)
Table 1: Parameters in dierent trac categories.
Access Category AIFSN CW
max
CW[backgroundWSA] 7 15
CW[voiceWSM] 2 3
(i) L
DATA
is the size of data transmitted over the physical
medium in bits. It contains the data payload, WAVE,
and MAC headers.
(ii) R
DATA
is the data rate in bits per second.
(iii) D is the theoretical distance within which packets
can be successfully received. This depends on the
environment radio propagation characteristics. In
our simulation, we set the transmission distance to
200 m, which has been determined from empirical
data measurements and is described in Section 4.
(iv) c is the speed of light set to 3 10
8
m/s.
(v) SIFS is the short interframe space with a length of
16 s.
(vi) AIFSN species the number of slot periods within
the AIFS (Arbitration Interframe Space) value used
by an access category during contention (Table 1).
(vii) AIFS is the lag time between the medium becoming
idle and the time when the access category starts or
resumes a random Backo period.
(viii) CW is a number of slots in particular Contention
Window (Table 1).
(ix) ACs are the Access Categories used by 802.11e and
WAVE MAC to manage dierent trac classes (voice,
video, and data).
(x) T
BoSlot
is the duration of a slot, this is set to 9 s.
The SRMB algorithm extends the RMB principle and
works as follows.
A station M
j
receives the packet and encapsulates the
list of MPR
i
1N
addresses from the incoming packet. If any
addresses of MPR
i
1N
match the M
j
address, then before a
retransmission M
j
adds a delay in length of wait time slot
T
slot
as follows:
T
slot
(J) = (J 1) m max(T
L MAC
). (2)
(i) J is J (1 N) and is the order of the node M
j
in
the list of MPR
i
1N
.
(ii) m is a multiplier added to avoid collisions when
the networks become busy and (1) expires. The
value is set to 1.5, which has been determined from
simulation investigation.
Else if M
j
address does not match any of the addresses
in MPR
i
1N
,then before a transmission node M
j
adds a time
delay according to (2), where
(i) J = N + S;
t
1
t
2
t
3
t
3
t
4
Figure 3: SRMB.
(ii) N is the maximum number of nodes in MPR
i
1N
;
(iii) S is the order of the sector where M
j
is positioned
(Figure 1). A sector is dened about the transmitting
node with a radius dened by the theoretical trans-
mission distance with each sector having a 90
degree
spread [23].
Time slots are chosen based on an MPR node priority,
and MPR nodes transmit one by one leaving sucient
time to avoid collisions at a source node and also to avoid
collisions between other non-MPR nodes in dierent sectors.
3.2.2. Pseudoacknowledgments (PACK). The principle of
SRMB is that nodes M
jn
broadcast one by one without
collisions at the source node or previous forwarding hop
M
i
. As previously described, a broadcasting node denes
geographical sectors and selects its MPR set MPR
i
1N
and
broadcasts. Selected neighbours of M
i
that receive the broad-
cast say M
j
and M
k
then rebroadcast. The rebroadcasting
by M
j
and M
k
is also received (overheard) at M
i
(Figure 3)
assuming no collisions. Collisions are mitigated due to the
spreading of the retransmissions over dynamic wait time
slots, and so each rebroadcast node should transmit in turn
and be overheard by M
i
. This overhearing is interpreted by
the PACK method as a form of pseudoacknowledgement
for the individual sectors. If an unacknowledged sector(s)
remains after some predened time (as per (3)), then the
node M
i
repeats the broadcast with a new list of MPR
i
1M
that contains only the missing sector(s). The algorithm is
repeated until all sectors are acknowledged or a maximum
number of repetitions are reached for the broadcast. The
broadcast repetition interval T
rep
is calculated according to
the following equation:
T
rep
= 2 N max(T
L MAC
)
+ Rand (N max(T
L MAC
)).
(3)
(i) N is the maximum number of nodes in MPR
i
1N
.For
other broadcasting protocols other than SRMB, N
represents the number of nodes that can possibly
retransmit.
(ii) Rand is a random value uniformly distributed in the
range 0 to (N max(T
L MAC
)) to further randomise
repetitions over a short time interval to avoid colli-
sions.
t
1
t
1
t
2
t
3
Collision
t
4
t
5
t
6
Figure 4: SRMB+PACK.
The PACK scheme partly solves the Hidden Terminal
Problem by using repetitions Figure 4. The maximum
repetitions are set to 3 by default.
The fundamental dierence between SRMB+PACK and
slotted protocols such as RR-ALOHA is that SRMB+PACK
uses access CSMA to the physical medium. Only specic
nodes act as forwarders for the broadcast and in turn create
virtual time slots during the broadcasting process at the
upper MAC layer to further randomise the channel access
time to decrease packet collisions. Nodes set the start of the
repetition slots based on the time the packet is received so
global synchronisation is not required and the slot size is
determined using (3). After this wait time slot expires, the
broadcast packet is passed from the upper MAC layer to the
lower MAC layer for transmission according to the particular
MAC standard (e.g., [26]).
The converse is true for slotted protocols such as
RR-ALOHA, where TDMA is used to access the physical
medium. All nodes must rely on a global clock for syn-
chronization, and each node has its own reserved time slot
to transmit with a xed length, which makes this scheme
unsuitable for variable length packets or event/bursty trac.
3.3. Reliable Broadcast Schemes under Test. In this paper,
we compared the proposed multihop SRMB+PACK scheme
with 3 other reliable broadcast methods. These mechanisms
also used SRMB as the underlying broadcasting mechanism
with WAVE [1] as the communications protocol.
3.3.1. Synchronous Fixed Retransmission (SFR). SFR has
been presented in [18, 19] and is based on repetitively
broadcasting the same message by a sender. The number
of rebroadcasts is not constant and is randomly chosen
according to the following principle.
Messages are assumed to have a specic lifetime, and this
life time, is divided into time slots, and from this a random
number of these slots are chosen to repetitively transmit
the broadcast. The time slots are synchronized to a global
clock. The authors have proposed other mechanisms but
have shown that SFR achieves the better performance.
In [18], the message lifetime is set to 100 ms, which we
see as unsuitable because it signicantly lengthens safety
messaging. Considering this, we decreased the lifetime to
10 ms and derived a suitable slot length of 1 ms using (1) and
(2), giving sucient time to perform repetitive broadcasting.
Each sender can randomly choose from 0 to 9 repetitions for
broadcasting and then broadcast in the selected slot.
3.3.2. Adaptive Transmission Power for Beacons (ATPBs). The
Adaptive Transmission Power (ATP) protocol presented in
[21] is based on nodes listening to the medium and counting
the collisions that occur in this period. Depending on this
value and the number of neighbours, a node decreases or
increases its transmission power appropriately. In [21], the
threshold for the number of neighbours is set to 30, when
this value is exceeded the transmit power is controlled using
ATP.
Irrespective of the packet type, the same power is used
to transmit all messages. Authors [27] highlight that such
an approach leads to dangerously reduced transmission
ranges for emergency data and this is counter productive,
where emergency data is typically sent on the maximum
transmit power to cover as many nodes as possible. Improved
performance is achieved using the maximum transmit power
as opposed to broadcasting over multiple hops. As an
alternative to ATP, we developed the Adaptive Transmission
Power for Beacons (ATPBs), which relies on the same method
of assessing channel, but the transmit power is only modied
for the periodic beacons to spare communication capacity
for safety messages that are transmitted with the maximum
possible transmit power.
3.3.3. Reliable Reservation-ALOHA (RR-ALOHA). The RR-
ALOHA protocol presented in [16] has been developed
within the European research project CarTalk2000 [17]. This
is a slotted technique (TDMA access), where nodes rely on
synchronised time slots for communications, where nodes
are assigned a single dedicated slot for transmission. To
prevent nodes from using the same slot, the Reservation-
ALOHA (R-ALOHA) [28] protocol uses a central repeater
that announces used slots. This concept is impracticable for
use within VANETs because the inclusion of static infrastruc-
ture would restrict VANET communications to centralised
vehicle-2-Infrastructure communications. To avoid the use
of central repeaters, RR-ALOHA [16] was developed and
proposes that each node sends beacons containing informa-
tion identifying which slot is used for communications with
their 1-hop neighbours. Anode, which receives beacons from
its 1-hop neighbour nodes, indirectly receives information
identifying the used slots for its 2-hop neighbours. This
allows nodes to access free slots and to avoid the Hidden
Terminal Problem.
4. Simulation Environment
We have developed a VANET simulation environment using
the network simulation tool OPNET V.12 [10] to evaluate
the performance of the PACK algorithm over the Slotted
Restricted Mobility-Based (SRMB) broadcast algorithm and
integrated this with the VANET specic Wave Short Message
Protocol (WSMP) based on a simplied model of the
Wave communications standard (parameters are shown in
Table 2). The Wave model contains one Control Channel
Table 2: Scenario description.
Scenario Urban scenario
Transmit power 18 dBm
Frequency 5.9 GHz
Data rate 6 Mbit/s
Bandwidth of channel 10 MHz
Transmit range
Two-Ray Ground model
with shadowing
Minimum broadcast distance 500 m
Maximum num hops 10
Speed of vehicles 050 km/h
Scenario dimensions 2 km 2.5 km
Density of nodes/km
2
10140
Number of Hazardous locations 3 (accident), 1 (rain)
Repetition interval of safety messages 1 s
Size of beacons 480b
Size of safety messages 368b
Beaconing interval 100 ms
(CCH) and one Service Channel (SCH) interface with total
channel duration of 100 ms with 50 ms per channel that
switch periodically at 50 ms intervals.
To approximate real world radio propagation, we imple-
mented a realistic radio propagation models in OPNET.
The model is based on the Two-Ray Ground model with
shadowing, where the parameters are set based on empirical
testing of 802.11p radio modules [29]. The packet loss ratio
is in the region of 40% for distances up to 100 m between the
transmitter and receiver while the losses increase to 90%with
distances of between 100 m and 150 m, and 100% losses are
achieved with distances beyond 200 m.
For experimental investigation, we modelled an urban
scenario using the road trac simulator SUMO [30], where
the scenario represents a topology of collector roads in a
5 km
2
area in the Bishopstown district in Cork City, Ireland.
The trac model contained dynamically moving vehicles
with varying speeds that are restricted to a maximum speed
of 70 kmph along 2-lane roads with a mixture of signalled
intersections, trac circles, and stop signs. The density of
vehicles ranged from 10 to 140 vehicles per km
2
, which
represented trac ows at night time to peak time. Two types
of emergency situations were investigated representing safety
of life applications and low-priority hazard/environmental
warning applications.
The rst scenario emulates 3 accidents in 3 roads in
low, medium, and high density road sections. Accidents
can be detected by vehicles within 50 m of the accident
location. A vehicle entering this 50 m sensing range detects
and immediately invokes a broadcast relating to this emer-
gency. A vehicle that is within this 50 m range when the
accident occurs selects a random wait time over a uniformly
distributed interval of 100 ms (corresponds to the WAVE
SYNC INTERVAL) before broadcasting. This distributes the
generation of broadcasts over the complete WAVE frame and
randomises the intervals at which vehicles rebroadcast and
lessen collisions due to broadcast storms. The broadcasting
is repeated at 1 s intervals.
The second scenario was designed to focus on the
throughput of the whole network and emulates an environ-
mental network wide event, rain detection in this case. Each
scheme was tested with dierent loads in the network. All
vehicles detect the rain event uniformly distributed in time
over 1 s and repeatedly broadcast every 1 s.
5. Performance Analysis
In the simulated environment, only two types of messages
are transmitted. Beacon messages WSA [1] were transmitted
every 100 ms by each node, and safety messages were
encapsulated in WSM [1] packets and broadcasted with
the Minimum broadcast distance being set to 500 m and
the Maximum hops being set to 10 hops. We collected the
simulation results from 3 seeds with at least 200 runs for
each seed. The metrics recorded from the experiments are
outlined below and shown in Figures 510.
Network overview (Figure 5)shown in this diagram
is the mean number of 1st hop and 2nd hop neighbours that
nodes have in the network. The diagram shows a limitation
of ATPB and RR-ALOHA schemes.
Link Load (Figure 6)this is calculated as the mean
ratio of the number of nodes that transmit safety broadcast
packet against the number of nodes that receive the packet.
The lower this value the better as this indicates that fewer
transmissions are needed to disseminate the broadcast
packet.
End-to-End Delay (Figure 7)this is a measure of the
mean time delay between the source of a safety message and
the node that receives the broadcast last. This also covers
the time delay created by time slots CCH TS and SCH TS
in the Wave protocol. In the case of the SFR scheme, this
is measured as the delay between the source node and the
reception of last repetition broadcast.
Delivery Ratio (Figure 8)this measure is dependent
on the density of a network and it is the mean delivery
ratio taken as the number of nodes inside an area that
receive safety broadcast versus the number of nodes in that
area. The area was dened by a source node as an area
inside a circle with the source node at the centre, and the
radius is dened by the Minimum broadcast distance. For
the SFR scheme, this was measured based on the number
of nodes inside the area that received safety broadcast
(from any repetition) versus the number of nodes in the
area.
Delivery Ratio versus Distance (Figure 9)this shows the
eect on the mean delivery ratio against increasing distance
from the source up to the Minimum broadcast distance.
Throughput (Figure 10)this measure was collected
over the complete network and refers to global network
events. All vehicles in a scenario detect a global event (e.g.,
raining) using sensors and all vehicles broadcast this event.
The purpose of this measurement was to investigate the
impact of the broadcast repetition interval, which was varied
from 0.01 to 3 packets per second, on the delivery ratio (the
number of nodes that receive the broadcast) in a network that
was moderately busy, with 60 vehicles/km
2
.
6. Theoretical and Experimental Results
6.1. Theoretical Results. We compare the proposed PACK
scheme with 3 existing schemes, namely SFR, ATPB, and
RR-ALOHA. All the schemes were overlaid on the SRMB
broadcasting protocol. According to the WAVE standard [1],
time was divided to frames (Sync interval) with a length of
100 ms. Each frame contains two slots the Control Channel
(CCH TS) and the Service Channel (SCH TS) time slots,
each with a length of 50 ms. Each of these slots begins with
a Guardian time of 5 ms to allow a unit to switch from
one channel to another. In the Guardian time interval, no
messages can be sent. Beacon messages and safety messages
were sent only in CCH TS after the Guardian time. If a
safety message was sent in CCH TS, the beacon message was
omitted to prevent overloading the medium.
For repeated broadcasting of an event (local, global), the
invoking of safety messages was uniformly distributed across
the Sync interval with a length of 100 ms. If a safety message
was invoked during the SCH TS 50 ms interval or Guardian
time 5 ms duration, then it waited until the beginning of
the CCH TS where it was immediately transmitted. A mean
time delay T
H MAC
for waiting emergency data (WSM) at
the upper MAC layer before being passed to the lower MAC
layer to access the CCH TS is calculated as per the following
equation:
T
SRMB
H MAC
=
T
SCH+G
T
sync
T
SCH+G
2
15 ms. (4)
(i) T
SCH+G
is the time in length of SCH TS (50 ms) plus
Guardian time (5 ms) when emergency data cannot
be sent.
(ii) T
sync
is the length of Sync interval 100 ms species in
Wave [1].
The Mean theoretical overall time delay for multihop
broadcasting T
SRMB
is calculated as per equation (5), which
is derived from (1), (2), and (4) as follows:
T
SRMB
= T
SRMB
H MAC
+ H (T
slot
(J) + T
L MAC
)
T
SRMB
18 ms.
(5)
(It presumes that all transmissions were made in one CCH
TS. Otherwise the T
H MAC
was extended to 55 ms (length of
SCH TS and Guardian time).)
(i) L
DATA
in (1) is the size in bits of an emergency packet
(WSM) with a value of 368 bits.
(ii) H is the mean number of hops and is set to 6. The
number was taken from mean number of hops in the
simulations that increased with increasing density.
(iii) It is presumed that T
slot
with J (1 N) is the
delay applied mainly at the origin of the broadcast,
where broadcasts are sent in dierent sectors based
on the priority of the MPR nodes. Here, J represents
the average number of MPR nodes per hop, based on
simulation evaluation this was set to J = 1.5.
6.1.1. Pack. In case of using the PACK scheme, the overall
multihop delay T
SPACK
is slightly increased due to the
following repetitions:
T
PACK
= T
SRMB
+ k T
rep
,
T
PACK
22 ms,
(6)
where k is the number of repetitions. This value depends
on the data trac on the physical medium, where in
less busy network the repetition value was approximately
one repetition per the complete broadcast and this went
up to approximately 7 for busy networks. For theoretical
estimation, we set k = 2.5, which is compared to medium
busy network.
The SRMB+PACK scheme increased end-to-end delay of
SRMB protocol by 22% (18 ms compared to 22 ms).
6.1.2. Synchronous Fixed Retransmission (SFR). In case of
using the SFR scheme, the overall multihop delay T
SFR
was
calculated as follows:
T
SFR
= T
SRMB
+ k T
SFR slot
,
T
SFR
23 ms.
(7)
(i) k is the mean number of broadcast repetitions equally
distributed from 0 to 9 as specied by the SFR
scheme.
(ii) T
SFR slot
is a slot in length of 1 ms species by SFR
scheme.
The SFR scheme increased end-to-end delay of SRMB
protocol by 28% and by 5% when compared against
SRMB+PACK.
6.1.3. Adaptive Transmission Power of Beacons (ATPBs). The
theoretical overall time delay of multihop broadcasting was
kept the same as in SRMB protocol. From the perspective
of broadcasting delay, the ATPB and SRMB schemes work
on the same principle. ATPB only aects the transmission
power of the beacons and does not straight impact on the
dissemination of emergency (WSM) data.
6.1.4. Reliable Reservation-ALOHA (RR-ALOHA). In RR-
ALOHA, the beacon (WSA) contained a list of all time slots,
where each entry relates to particular time slot. Each entry
in the list had a size of 11 bits and contained information
relating to the state of the channel (busy or idle) and the
short MAC address of the node transmitting on that time
slot. Because we implemented RR-ALOHA over the WAVE
standard, we had to derive the maximum number of slots
rst. The size of beacons L
DATA
used by RR-ALOHA was
calculated in as follows:
L
RR-ALOHA
DATA
= L
MAC
+ L
WSA
+ L
RR-ALOHA
,
L
RR-ALOHA
= 11 bits S,
(8)
where L
MAC
is the size of the MAC header with 272 bits, and
L
WSA
is the size of the WSA beacons with length of 480 bits.
From a knowledge of the maximum available time of 45 ms
in the CCH TS and from maximum transmission delay (1),
we determine that the maximum number of time slots S used
by RR-ALOHA is 90 with length of a slot being 0.5 ms. The
overhead L
RR-ALOHA
was calculated as 11 bits 90 time slots,
which is 990 bits.
The mean theoretical overall multihop delay T
RR-ALOHA
was calculated as follows:
T
RR-ALOHA
H MAC
50 ms,
T
RR-ALOHA
= T
RR-ALOHA
H MAC
H + T
L MAC
,
T
RR-ALOHA
300 ms.
(9)
The delay T
RR-ALOHA
depends on the number of hops H
and how many retransmit nodes are chosen. Theoretically,
with 10 hops (10 hops in the maximum number of allowable
hops for a broadcast) the delay can vary from 18 ms (see (5))
to 1000 ms (see (9)) depending on the selecton of forwarding
hops and their time slot.
RR-ALOHA gives the longest delay, 17 times higher than
SRMB and 14 times higher than SRMB+PACK and SFR.
6.2. Experimental Results. All the results presented are
represented by mean values for individual data points which
are averaged over approximately 600 values with 3 seeds. The
data sets in most cases have a skewed distribution, so it is
preferable to use the rst and third quartiles (q
25
, q
75
) as
descriptive statistics.
6.2.1. Network Overview. Network overview(Figure 5)this
shows the mean number of 1-hop and 2-hop neighbours that
nodes have in the network. In [21], for the ATP scheme the
neighbour threshold is set to 30 nodes, meaning that if a
node has more than 30 neighbours, then the node should
change its transmission power which would then aect the
broadcasting performance. As can be seen in Figure 5, the
number of neighbours exceeds 30 between x = 40 and
x = 60 (vehicles/km
2
). The RR-ALOHA protocol uses time
slots, where the number of time slots, was set to 90 using
(1) and (8). The number of 1st hop and 2nd hop neighbours
exceeds the maximum number of slots, that is, 90 at x = 60
(vehicles/km
2
), beyond this density some nodes will have to
share the same time slot.
The results show a limitation of the ATPB and RR-
ALOHA as the number of neighbouring nodes can aect
the broadcast performance. SRMB+PACK and SFR are not
restricted by number of neighbours and can work across all
neighbour densities.
0
50
100
150
200
0 50 100 150
N
u
m
b
e
r
o
f
n
e
i
g
h
b
o
u
r
s
(vehicles/km
2
)
1st neighbours
2nd neighbours
Figure 5: Network overview shows the mean number of 1st hop
and 2nd hop neighbours that nodes have in the network.
6.2.2. Link Load. Link Load (Figure 6) showed that all
schemes (except SFR because of repetitions) have a rapidly
decreasing link load trend. As the vehicle density increases,
the network connectivity goes from sparsely connected to
well connected. After SFR, the SRMB protocol performs
the next worst in terms of link load (with a Link Load
Ratio mean value of #LL = 0.25, with 1st and 3rd
quartiles being q
25
= 0.14, q
75
= 0.25, taken at the
highest density of vehicles with x = 140 vehicles/km
2
)
in denser networks. The PACK scheme in lower density
networks performs marginally poorer (5%, this drop in
performance is attributed to the repetition of broadcasts
for unacknowledged sectors) than SRMB with #LL = 0.77,
q
25
= 0.5, and q
75
= 1, at a vehicle density x = 10/km
2
in less busy networks. For higher density networks with
x = 140 vehicles/km
2
,values of #LL = 0.22, q
25
=
0.18, and q
75
= 0.27 are achieved, and this represents an
improvement of 12% when compared with SRMB. For more
saturated networks, the pseudoacknowledgements used by
PACK to acknowledge sectors reduce the probability of non-
MPR nodes rebroadcasting and thus reduce the probability
of collisions which results in fewer transmissions in the
congested medium. The ATPB and SRMB schemes have a
similar performance as the power control aspect of ATPB
only applies to the beacons. The best performance across
all densities was achieved by RR-ALOHA as expected with
a 40% improvement over SRMB at a vehicle density of x =
140/km
2
and #LL = 0.15, q
25
= 0.12, and q
75
= 0.15. This
performance is attributed to the fact that RR-ALOHA uses
one slot per node transmissions and will always outperform
CSMA/CA methods, on which the other schemes are based.
The better performance in terms of link load is oset by
the poor end-2-end delay and throughput achieved with RR-
ALOHA. The worst performance, that is, the greatest number
of transmissions was attributed to SFR, which signicantly
diers from the other schemes. In the lightest density, SFR
reached a value (#LL = 4, q
25
= 2, q
75
= 6, and density
x = 10) 5 times greater than SRMB. In the heaviest density
0
1
2
3
4
5
6
7
8
0
0.2
0.4
0.6
0.8
1
0 50 100 150
L
i
n
k
l
o
a
d
r
a
t
i
o
(vehicles/km
2
)
SRMB
+PACK
+ATPB
+RR-ALOHA
+SFR
Figure 6: Link Load Ratio is calculated as the mean ratio of the
number of nodes that transmit a broadcast packet against the
number of nodes that receive the packet. Second right y axis is for
SFR scheme, which signicanty dier from the others.
network, SFR ooded the network, which led to rapidly
increasing unsuccessful transmissions (#LL = 7, q
25
= 0.5,
q
75
= 2, and density x = 140) with values 30 times greater
than in SRMB.
The link load results show that all schemes (except
SFR) perform broadcasting with a very low number of
transmissions and decrease the number with increasing
density of vehicles as network increasing in connectivity due
to a larger number of nodes. At higher densities, SFR ooded
network because of repetitions and is actually worse than
using a simple ooding protocol making SFR unsuitable for
VANETs.
6.2.3. End-to-End Delay. End-to-End Delay (Figure 7)As
expected due to their similar operation, the results for end-
to-end delay showed that the SRMB protocol and the ATPB
scheme maintain the same relatively constant short time
delay (End-to-End Delay, #EE = 20 ms, q
25
= 4, q
75
= 37,
and density x = 140), which matched the theoretical result
achieved with (5). In comparison, the PACK method had
a slightly increased delay across all densities from lighter
densities (#EE = 18 ms, q
25
= 0.4, q
75
= 33, and density
x = 10) with a deterioration in performance when compared
with SRMB of 12% and in larger densities (#EE = 33 ms,
q
25
= 8, q
75
= 50, and density x = 140) a deterioration of
50% again comparing to SRMB. Using (6) and a repetition
factor of 2.5 and looking at a medium density network with
x = 40/km
2
,theoretical results matched experimental result
(#EE = 22 ms, q
25
= 4, q
75
= 39, and density x = 40).
The SFR scheme gave the 2nd longest delay across a low-
density network (#EE = 22 ms, q
25
= 5, q
75
= 40, and
density x = 10) to a high-density network (#EE = 47 ms,
q
25
= 11, q
75
= 60, and density x = 140) with a deterioration
in performance ranging from 40% to 240% when compared
against SRMB. Using (7), the theoretical end-to-end delay
0
200
400
600
800
0
20
40
60
80
100
E
n
d
-
t
o
-
e
n
d
d
e
l
a
y
(
m
s
)
0 50 100 150
(vehicles/km
2
)
Figure 7: End-to-End Delay is a measure of the mean time delay
between the source of a safety message and the node that receives
the broadcast last. Second right y axis is for RR-ALOHA scheme,
which signicanty dier from the others. Label is the same as at
Figure 6.
does not match the empirical result. Equation (7) is derived
using the maximum transmission time T
L MAC
from (1)
which does not consider a saturated case (i.e., collisions are
not considered). Equation (1) is valid only for lightly loaded
networks. In more dense networks if a transmission on the
medium is detected while a node is in Backo, a new Backo
time is set and the transmission delay (1) is increased. With
the SFR protocol, we have an increasing load on the physical
medium as a consequence of repetitions that saturate the
network and lead to collisions. The longest delay is given
by RR-ALOHA across all densities from the lowest (#EE =
120 ms, q
25
= 40, q
75
= 176, and density x = 10) to
highest (#EE = 740 ms, q
25
= 580, q
75
= 920, and density
x = 140) with a deterioration from 7.5 times to 35 times
that of SRMB. Using (9), we see that the theoretical result
depended strongly on the number of hops and the selection
of the next hops based on their time slots. This variation was
described in Section 6.1.4 and matched experimental results.
The results showed that the SRMB, ATPB, PACK, and
SFR schemes reach a fraction of driver reaction time (around
0.05 s of 0.7 s [31]). On the basis of the results, we show that
these schemes in terms of end-to-end delay are appropriate
for VANETs. As RR-ALOHA has prohibitively long end-
to-end delays across all densities, we conclude that this
method based on comparison with driver reaction speeds is
unsuitable for emergency data dissemination in VANETs.
6.2.4. Delivery Ratio. Delivery Ratio (Figure 7)Results
showed that the SRMB protocol reached relatively constant
values for Delivery Ratio, #DR = 0.62, q
25
= 0.48, q
75
= 0.90,
and density x = 40 to #DR = 0.61, q
25
= 0.43, q
75
=
0.92, and density x = 140. Similar results were achieved
with ATPB and acknowledged that sparing communication
capacity by decreasing transmit power of beacons did not
have signicant eect on delivery ratio. The PACK method
in low-density network gave values of (#DR = 0.35, q
25
=
0.18, q
75
= 0.45, density x = 10) and in high density
0.2
0.4
0.6
0.8
1
0 50 100 150
(vehicles/km
2
)
SRMB
+PACK
+ATPB
+RR-ALOHA
+SFR
D
e
l
i
v
e
r
y
r
a
t
i
o
Figure 8: Delivery Ratio is a measure of the mean delivery ratio
taken as the number of nodes inside an area that receive a broadcast
versus the number of nodes in that area.
gave (#DR = 0.83, q
25
= 0.77, q
75
= 0.96, and density
x = 140). These results reect improvements of 2% to
36% when comparing against SRMB from medium- to high-
density networks (#DR = 0.70, q
25
= 0.57, q
75
= 0.97,
and medium density x = 40) with PACK improving overall
other methods in medium- to high-density networks. SFR
gives the best performance in lower density networks because
of the repetitions in a sparsely connected network (#DR =
0.42, q
25
= 0.30, q
75
= 0.55, and density x = 10) with
improvements of 32% over SRMB. In medium busy densities
(#DR = 0.70, q
25
= 0.60, q
75
= 0.98, and density x = 40),
SFR has a slight deterioration of 4% when compared to
PACK and in the highest density (#DR = 0.78, q
25
= 0.77,
q
75
= 0.97, and density x = 140), the decline in performance
falls to 6% when compared to PACK. The RR-ALOHA
scheme gave a slightly poorer results when compared to
PACK (low density (#DR = 0.33, q
25
= 0.18, q
75
= 0.45, and
density x = 10) to high density (#DR = 0.80, q
25
= 0.76,
q
75
= 0.96, and density x = 140) with a 5% decline in
performance.
The results show that PACK, SFR, and RR-ALOHA
signicantly improved delivery ratio across all densities with
ATPB giving a performance again similar to SRMB. Again
this shows the unsuitability of the ATPB protocols for reliable
broadcasting in VANETs as it only refers to beacon frames.
6.2.5. Delivery Ratio versus Distance. Delivery Ratio versus
Distance (Figure 8)these results were captured at a density
of 60 vehicles/km
2
(medium busy network) and showed that
for all schemes the Delivery Ratio fell of with increasing
distance. Again SRMB and ATPB give similar results. PACK,
SFR, and RR-ALOHA improve on SRMB and give very
similar values up to a distance, x = 250 m from a sender,
with an improvement of 18% over SRMB.
0.2
0.4
0.6
0.8
1
0 100 200 300 400 500
D
e
l
i
v
e
r
y
r
a
t
i
o
Distance (m)
SRMB
+PACK
+ATPB
+RR-ALOHA
+SFR
Figure 9: Delivery Ratio shows the eect on the mean Delivery
Ratio against increasing distance.
Further, from the sender at a distance x = 350 m, SFR
gives an improvement of 14%, PACK gives 22% and RR-
ALOHA 28% over SRMB. On the boundary of the Minimum
broadcast distance (500 m), improvements with SFR were
23%, RR-ALOHA 25%, and PACK 28% over SRMB.
The results showed that with increasing distance the
delivery ratio falls o, but again SFR, PACK, and RR-ALOHA
improve broadcasting performance.
6.2.6. Throughput. Throughput (Figure 10)results showed
that all schemes have a decreased delivery ratio for broadcasts
with increasing load in the network. As before, SRMB and
ATPB perform similarly. SRMB (likewise ATPB) give the
lowest throughput ratio #TP = 0.45, q
25
= 0.27, q
75
= 0.72,
and load x = 0.01 packets/1 s under low load and this
decreases #TP = 0.32, q
25
= 0.04, and q
75
= 0.53, with
increasing load x = 3 packets/1 s (high load). In low load,
x = 0.01 packets/1 s RR-ALOHA has #TP = 0.49, which
improves on SRMB by 9% and maintained a trend similar to
PACK until reaching a moderately loaded network x = 0.3
packets/1 s, where #TR = 0.25, q
25
= 0.02, q
75
= 0.43,
and load x = 3 packets/1 s fell to 22% below SRMB. This
performance deterioration is attributed to the fact that RR-
ALOHA uses TDMA access and will always perform worse
than CSMA/CA access, on which the other schemes are
based. SFRreached the highest values with #TR = 0.58, q
25
=
0.30, and q
75
= 0.87 under low load x = 0.01 packets/1 s and
deteriorated #TR = 0.42, q
25
= 0.30, and q
75
= 0.73 under
greater load x = 3 packets/1 s due to saturating the network
with repeat broadcasts but still improved on SRMB by 31%.
The PACK scheme reached values similar to RR-ALOHA
under low load. With increasing load, the PACK improves
on these with #TR = 0.45, q
25
= 0.22, and q
75
= 0.68, high
load x = 3 packets/1 s giving an improvement of 40% over
SRMB, 80% over RR-ALOHA, and 7% over SFR.
The results showed that with increasing load the
throughput ratio dropped across all schemes. TDMA-based
0
0.2
0.4
0.6
0.8
1
1 0.01 0.1 10
T
h
r
o
u
g
h
p
u
t
Invoked broadcasts per second per vehicle in 60 vehicles/km
2
SRMB
+PACK
+ATPB
+RR-ALOHA
+SFR
Figure 10: Throughput measure an impact of the broadcast
repetition interval on the delivery ratio.
channel access in RR-ALOHA shows reduced throughput
against CSMA/CA-based access in highly loaded networks.
SFR scheme showed decreased throughput due to overload-
ing of network with many repetitions but still maintains
a higher performance then SRMB. The PACK scheme in
moderate to highly loaded networks gives a better perfor-
mance than the SFR, RR-ALOHA, and SRMB (ATPB). These
advantages of the PACK algorithm can be further highlighted
when safety application is required to report with a higher
rate (<1s) as it can maintain high delivery ratio for moderate-
to high-density networks (e.g., vehicle platooning).
7. Discussion, Conclusions, and Future Work
In this paper, we have concentrated on techniques that
increase reliability of a multihop broadcast protocol. We have
proposed the Pseudoacknowledgments (PACK) scheme that
improves reliability in multihop broadcasting protocols by
repeating broadcast transmissions on unsuccessful links. The
scheme was compared with existing mechanisms over an
urban scenario using the network simulator tool OPNET
[10] with an empirical-based propagation model [29], realis-
tic mobility patterns using the road trac simulator SUMO
[30], and the Wave [1] standard. All schemes were overlaid
on the low-latency p-persistent CSMA/CA broadcasting
protocol called the Slotted Restricted Mobility-Based (SRMB)
protocol.
The scenarios were designed to test safety-related data
dissemination and measured relevant broadcasting statistics:
link load, end-to-end delay, delivery ratio, and throughput.
From our results, we draw the following conclusions and
identify tasks for future work.
(i) Changing transmission power of beacons does not
have a signicant eect on performance of broad-
casting, as has been demonstrated by examining the
performance of ATPB.
(ii) Repeating broadcasts leads to increased delivery ratio,
but it also increases the number of transmissions in
the network. This can lead to ooding the network
with repetitions and can decrease the delivery ratio
in denser networks. The throughput results showthat
SFR scheme can easily saturate the network under
higher loads which leads to a rapidly decreasing deliv-
ery ratio. The redundancy incurred as a consequence
of repetitions which can lead to ooding makes this
scheme unsuitable for VANETs.
(iii) Using small time slots for broadcasts leads to a
high delivery ratio due to minimum collisions in
particular slots but this increases end-to-end delay
as broadcasts are delayed at successive rebroadcast
nodes. In the case of RR-ALOHA, the delay reaches
large values that cannot be tolerated for safety-related
data dissemination. Another disadvantage of slotting
is that it decreases throughput in densely loaded
network which corresponds with the throughput
performance when comparing CSMA/CA access with
ALOHA access. When the number of nodes exceeds
the maximum number of available slots, nodes must
share slots. However, this limitation does not have a
signicant eect on the broadcasting performance in
our simulations.
(iv) In Summary, PACK in moderate-to high-density
networks achieves an increase in delivery ratio of
approximately 5%over SFRand 3%over RR-ALOHA
while this may appear to be a minor increment
RR-ALOHA is unsuitable for VANETs due to its
excessive end-to-end delay, for example, 741 ms for
RRALOHA against 32 ms for PACK in high-density
networks. While SFR has a tolerable end-to-end
delay for VANETs, it does so at the expense of
saturating the network with repeat broadcasts and
this is highlighted by the link load metric which
at a vehicle density of 140/km
2
is 35 times higher
than of PACK. While PACK and SFR give similar
delivery ratio metrics, SFR does so at the expense of
excessive bandwidth usage in comparison to PACK.
Furthermore, due to less saturating of the network
PACK achieved the highest throughput in moder-
ate to high density networks. This makes PACK
suitable for applications that need to frequently
report (e.g., vehicle platooning, crashed vehicle
detection).
(v) From the experimental results presented in this
paper, we can conclude that the PACK mechanism
increases the reliability of multihop broadcasting and
is suitable for safety-related data dissemination. In
terms of end-to-end delay and bandwidth savings,
PACK outperforms RR-ALOHA and SFR, respec-
tively, making PACK a more reliable protocol for
safety data dissemination in VANETs.
Although SRMB+PACK protocol has been primarily
tested on a specic case of safety application (accident
reporting), the SRMB+PACK mechanism can be used as
a multihop data dissemination mechanism for a range
of applications that require high packet delivery and low
latency in very dynamic ad hoc networks. The SRMB+PACK
protocol could be used in route discovery for reactive routing
protocols in VANETs. From the route discovery, perspective
routes would be built based on latency, bandwidth con-
sumption, and mobility of nodes in the source-destination
path. Nodes with similar mobility behaviour (speed, motion
vector) would be selected as intermediate hops as this
supports the generation of stable routes and reduces route
maintenance overhead.
References
[1] Intelligent Transportation Systems Standards Fact Sheet,
IEEE 1609Family of Standards for Wireless Access in
Vehicular Environments (WAVE), ITS Standards Program.
[2] http://europa.eu/legislation summaries/transport.
[3] http://ec.europa.eu/information society/eeurope/2005/index
en.htm.
[4] W. Chen, R. K. Guha, T. J. Kwon, J. Lee, and I. Y. Hsu, Asurvey
and challenges in routing and data dissemination in vehicular
ad-hoc networks, in Proceedings of the IEEE International
Conference on Vehicular Electronics and Safety, pp. 328333,
Columbus, Ohio, USA, September 2008.
[5] M. Torrent-Moreno, D. Jiang, and H. Hartenstein, Broadcast
reception rates and eects of priority access in 802.11-
based vehicular ad-hoc networks, in Proceedings of the 1st
ACM International Workshop on Vehicular Ad Hoc Networks
(VANET 04), pp. 1018, Philadelphia, Pa, USA, 2004.
[6] G. Korkmaz, E. Ekici, F. Ozguner, and U. Ozguner, Urban
multiHop broadcast protocol, in Proceedings of the 1st ACM
International Workshop on Vehicular Ad Hoc Networks, pp. 76
85, Philadelphia, Pa, USA, 2004.
[7] M. Sun, W. Feng, T. Lai, K. Yamada, H. Okada, and K.
Fujimura, GPS-based message broadcasting for inter-vehicle
communication , in Proceedings of the International Confer-
ence on Parallel Processing, pp. 279286, Toronto, Canada,
2000.
[8] M. Heissenbuttel, T. Braun, M. Walchli, and T. Bernoulli,
Optimized stateless broadcasting in wireless multi-hop net-
works, in Proceedings of the 25th IEEE International Confer-
ence on Computer Communications (INFOCOM06), pp. 112,
Barcelona, Spain, April 2006.
[9] W. Si and C. Li, RMAC: a reliable multicast MAC protocol for
wireless ad hoc networks, in Proceedings of the International
Conference on Parallel Processing (ICPP 04), pp. 494501,
August 2004.
[10] http://www.opnet.com.
[11] J. Peng, Slim ARQ for reliable broadcasting in wireless
LANs, in Proceedings of the IEEE Wireless Communications
and Networking Conference (WCNC08), pp. 21642169, 2008.
[12] S.-T. Sheu, Y. Tsai, and J. Chen, A highly reliable broadcast
scheme for IEEE 802.11 multi-hop ad hoc networks, in
Proceedings of the IEEE International Conference on Commu-
nications (ICC 02), vol. 1, pp. 610615, New York, NY, USA,
2002.
[13] K. Tang and M. Gerla, Random access MAC for ecient
broadcast support in ad hoc networks, in Proceedings of
the IEEE Wireless Communications and Networking Conference
(WCNC 00), pp. 454459, Los Angeles, Calif, USA, 2000.
[14] http://www.ietf.org/rfc/rfc3626.txt.
[15] F. Hassanzadeh, Reliable broadcast of safety messages in vehic-
ular ad hoc networks, M.S. thesis, Graduate Department of
Electrical and Computer Engineering University of Toronto,
Toronto, Canada, 2008.
[16] F. Borgonovo, A. Capone, M. Cesana, and L. Fratta, RR-
ALOHA, a Reliable R-ALOHA broadcast channel for ad-hoc
inter-vehicle communication networks, 2002, ftp://ftp.elet
.polimi.it/users/Flaminio.Borgonovo/PUBBC/2002medhoc
.pdf.
[17] http://www.cartalk2000.net.
[18] Q. Xu, T. Mak, J. Ko, and R. Sengupta, Medium access control
protocol design for vehiclevehicle safety messages, IEEE
Transactions on Vehicular Technology, vol. 56, no. 2, pp. 499
518, 2007.
[19] Q. Xu, R. Senguptay, T. Mak, and J. Ko, Vehicle-to-vehicle
safety messaging in DSRC, in Proceedings of the 1st ACM
International Workshop on Vehicular Ad Hoc Networks, pp. 19
28, 2004.
[20] F. Farnoud and S. Valaee, Repetition-based broadcast in
vehicular ad hoc networks in Rician channel with capture, in
Proceedings of the 27th Conference on Computer Communica-
tions (INFOCOM 08), pp. 16, 2008.
[21] L. Yang, J. Guo, and Y. Wu, Channel adaptive one hop
broadcasting for VANETs, in Proceedings of the 11th IEEE
Conference on Intelligent Transportation Systems (ITSC 08),
pp. 369374, Beijing, China, 2008.
[22] M. Torrent-Moreno, P. Santi, and H. Hartenstein, Distributed
fair transmit power adjustment for vehicular ad hoc net-
works, in Proceedings of the 3rd Annual IEEE Communications
Society on Sensor and Ad Hoc Communications and Networks
(Secon 06), vol. 2, pp. 479488, Reston, Va, USA, 2006.
[23] M. Koubek, S. Rea, and D. Pesch, Eective emergency
messaging in WAVE based VANETs, in Proceedings of the
1st International Conference on Wireless Access in Vehicular
Environments (WAVE 08), December 2008.
[24] http://www.ietf.org/rfc/rfc3561.txt.
[25] O. Tonguz, N. Wisitpongphan, F. Bai, P. Mudalige, and V.
Sadekar, Broadcasting in VANET, in Proceedings of the
Mobile Networking for Vehicular Environments (MOVE 07),
pp. 712, May 2007.
[26] 802.11e-2005, IEEE Standard for Information technology
telecommunications and information exchange between
systemslocal and metropolitan area networksspecic
requirementspart 11: Wireless LANMediumAccess Control
(MAC) and Physical Layer (PHY) specications Amendment
8: MediumAccess Control (MAC) Quality of Service Enhance-
ments, 2005.
[27] J. Mittag, F. Thomas, J. H arri, and H. Hartenstein, A
comparison of single- and multi-hop beaconing in VANETs,
in Proceedings of the 6th ACM International Workshop on
Vehicular Inter-Networking (VANET 09), pp. 6978, 2009.
[28] W. Crowther, R. Rettberg, D. Waldem, S. Ornstein, and F.
Heart, A system for broadcast communications: reservation
ALOHA, in Proceedings of the 6th Hawaii International
Conference on System Sciences, January 1973.
[29] O. Brickley, M. Koubek, S. Rea, and D. Pesch, A network
centric simulation environment for CALM-based cooperative
vehicular systems, in Proceedings of the 3rd International
Conference on Simulation Tools and Techniques (SIMUTools
10), 2010.
[30] http://sumo.sourceforge.net.
[31] M. Green, How long does it take to stop? Methodological
analysis of driver perception-brake times, Transportation
Human Factors, vol. 2, no. 3, pp. 195216, 2000.
doi:10.1155/2010/148303
Research Article
Trafc FlowCondition Classication for Short Sections Using
Single Microwave Sensor
Muhammed G. Cinsdikici
1
and Kemal Memis
2
1
International Computer Institute, Ege University, 35100 Bornova, Turkey
2
Trac Systems Department, Aselsan Mil.Electrnoics Company, ASELSAN Inc., 10016 sok.No. 16, A.O.S.B, C i gli, 35620

Izmir, Turkey
Correspondence should be addressed to Muhammed G. Cinsdikici, cinsdikici@ieee.org
Received 19 October 2009; Accepted 2 September 2010
Copyright 2010 M. G. Cinsdikici and K. Memis. This is an open access article distributed under the Creative Commons
Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is
properly cited.
Daily observed trac ow can show dierent characteristics varying with the times of the day. They are caused by trac incidents
such as accidents, disabled cars, construction activities and other unusual events. Three dierent major trac conditions can be
occurred: Flow, Dense and Congested. Objective of this research is to identify the current trac condition by examining
the trac measurement parameters. The earlier researches have dealt only with speed and volume by ignoring occupancy. In
our study, the occupancy is another important parameter of classication. The previous works have used multiple sensors to
classify trac condition whereas our work uses only single microwave sensor. We have extended Multiple Linear Regression
classication with our new approach of Estimating with Error Prediction. We present novel algorithms of Multiclassication with
One-Against-All Method and Multiclassication with Binary Comparison for multiple SVM architecture. Finaly, a non-linear
model of backpropagation neural network is introduced for classication. This combination has not been reported on previous
studies. Training data are obtained from the Corsim based microscopic trac simulator TSIS 5.1. All performances are compared
using this data set. Our methods are currently installed and running at trac management center of 2.Ring Road in Istanbul.
1. Introduction
Trac ow characteristic shows dynamical change at dif-
ferent time periods of the day. Many trac incidents such
as accidents, disabled cars, construction activities, high
demands on trac, and other unusual events cause this
change. For dealing with these unstable trac problems,
trac conditions should be claried. Mainly there exist three
dierent major trac conditions: Flow Trac, Dense
Trac, and Congested Trac. Clarication of them req-
uires careful/detailed examination of the ow parameters of
speed, volume, and occupancy (SVO).
Accurate interpretation of SVO supports trac manage-
ment centers to make proper decision on directing the trac
to the less intensive roads. Hence, the response time for
intervention in an incident will be reduced.
Measurement methods for obtaining SVO have changed
for the last 60 years of the span (i.e., especially the last 40
years with the rapid rise in the number of freeways). Indeed,
they are still changing [1]. Some of them are
(i) measurement at a point,
(ii) measurement over a short section (about 10 meters),
(iii) measurement over a length of road (at least 0.5 km),
(iv) measurement with mobile observer in the trac
stream,
(v) measurement with multiple simultaneous mobile
vehicles, as part of ITS (Intelligent Transportation
Systems).
Although all measurement methods above produce speed
and volume of SVO, the only method producing occupancy
(i.e., the percentage of unit time that the detection zone of
the instrument is occupied by vehicles) is the measurement
over a short section [1].
Detectors used for short section measurement can be
based on inductive loop (IL), microwave, radar, photocell,
ultrasonic, and analog/digital camera technologies [1]. Since
the quick and eortless installation is possible only on IL
technology, we have used Radio Transmissions Microwave
Sensor-based (RTMS) IL detector. We have collected SVO
parameter values at dierent times of day by using RTMS-IL.
The collected SVO data is analyzed through our three
distinct novel approaches to classify trac ow as Flow,
Dense, or Congested. These approaches are estimating
error prediction for multiple linear regression analysis, two
improved variants of support vector machines (SVM), and
backpropagation neural network. First two methods are
linear classiers, whereas NN is non-linear.
In Section 2, the contributions of our paper will be given
in the related works. Then, background is presented. In the
Section 4, all details about the proposed study are going to
be given. Section 5 summarizes experimental results. Finally,
conclusion is given.
2. Related Works
Flow theory has been tried to analyze trac through the
speed, volume, and the vehicular concentration parameters.
Temporal vehicular concentration named as occupancy
can be measured only over a short section (i.e., shorter than
the minimum vehicle length). So, this parameter becomes
unmeaning for long section measurements.
Density as an alternative vehicular concentration has
been a part of trac measurement since 1930s. It depicts
the number of vehicles over a long section (i.e., one mile or
kilometer) in contrast to occupancy.
Although vehicular concentration encompasses both
density and occupancy parameters, indeed, it would be fair
to say that the majority opinion is in favor of using density
during the evolution of trac ow theory.
However, a minority view has intended to use occupancy
in theoretical works. Although there are well-dened facts
put forward by the majority for the continued use of density,
the minority also propounds major reasons for making more
use of occupancy. The most crucial reason among them
can be given as density (i.e., vehicles per length of road)
ignores the eects of vehicle length and trac composition.
Occupancy, on the other hand, is directly aected by both of
these variables and therefore gives a more reliable indicator
of the amount of a road being used by vehicles.
First mathematical model-based on speed, volume, and
density variables had been developed by Greenshild in
mid 1930s [2] using the aerial photographs. In his work,
relationship between speed and density is introduced relying
on simple linear regression approach.
After World War II, with the tremendous increase in
use of automobiles and the expansion of the highway
system, there was also a surge in the study of trac
characteristics and the development of trac ow theories.
In 1950s, theoretical developments based on a variety
of approaches, such as car-following, trac wave theory
(hydrodynamic analogy), and queuing theory has emerged.
Some of the seminal works of that period include the works
by Reuschel (1950) [35], Wardrop (1952) [6], Pipes (1953)
[7], Lighthill and Whitham (1955) [8], Newell (1955) [9],
Webster (1957) [10], Edie and Foote (1958) [11], Chan-
dler et al. (1958) [12], and other papers by Herman et al.
Reuschel and Pipes oered a microscopic trac model
that identies the linear dependency between the speed of
a vehicle and the distance between the vehicles in a single
lane. The models described by Reuschel and Pipes were
reasonable in concept, but no experimental verication of
their conclusions was pursued for many years [13].
Wardrops theory was based on two major principles.
The rst one was stating that travel times between the same
origin and destination pairs for any used routes are less
than or equal to the travel times for all unused routes. This
is referenced as Dynamic User Equilibrium (DUE) in the
literature and used for large-scale networks. Diverting trac
with DUE is inecient and dicult-to-implement system
optimum (i.e., min. average travel time) [14]. The second
one aims at minimizing both the total travel times and
average travel times for all assigned routes for all drivers on
the whole network. However, individual choice of drivers was
by no means guaranteed to satisfy this principle and in most
cases did not [13].
A few years later, Lighthill and Whitham (L-W) together
set out the rst comprehensive theory of kinematic waves.
In L-W trac model, there exists correlation between
trac ow (cars/hour) and trac density (cars/mile) [8].
Propagation of shock waves, generated by trac transitions
from one steady state to another, was also determined by
L-W model. Nevertheless, this model was viable only for
describing the density changes. Unfortunately, the model
produces larger densities exceeding the possible maximum
vehicle density.
Greenberg improved the existing mathematical models
by adding the nonlinearity into the model structure [15]. He
treats the trac stream as a continuous uid and derives the
relations between speed, density, and ow by uid dynamics.
The distinction of free (i.e., non-congested) and con-
gested ow on the speed-density model was carried out by
Eddie [16] and Underwood [17] and was investigated in an
important empirical test by Drake et al. [18].
Athol [19] suggested using the occupancy rather than
the density for the ow-concentration works, however his
suggestion became popular one decade later.
Speed, ow, and occupancy (SFO) relationships have
been studied at the free ow level by Hurdle and Datta
[20], Persaud and Hurdle [21], F. L. Hall and L. M. Hall
[22], Banks [23], Smith et al. [24], Hall et al. [25], and
Wemple et al. [26]. Also, SFO have been studied at congested
ow level and by Hall and Montgomery [27], Zhou and Hall
[28], and Banks [29, 30].
SVMs are usually employed for incident detection algo-
rithms identifying the anomalies of trac ow model using
single or two sensors [31]. The process for determining
the presence of an incident is twofold. The rst is a
determination of congestion (exits/not-exists), one of the
states we also want to identify. Next is the binary analysis
of type of congestion (incident occurred/not-occurred). Past
researches and applications usually tend to use the trac
descriptors of two trac sensors as their input parameters
for incident detection with SVM methods.
In ITS, neural networks can be found on the areas like
vehicle detection, road detection, and single loop vehicle type
classication [32, 33]. Rarely some work related with trac
ow control can be found and they are also not well dened
[34].
All retrospective researches so far show the binary
relationships between the trac parameters such as ow rate
and occupancy (FR-O), ow rate and speed (FR-S), speed
and occupancy (S-O) at congested or ow levels. Trac
status is assessed according to these binary relationships or
by inspecting only one parameter (like speed) rather than
examining all of them.
The rst contribution of this paper to the trac studies
is focusing on another major trac ow level ignored in
earlier researches called dense ow level. All the previous
works studies congested and ow trac levels.
The second contribution of the paper is that no one in
the earlier works is based on single sensor (i.e., all of them
used double sensor placed within some distance) with using
SVO triplets (i.e., triple relations of trac parameters).
The third one is our new approach of estimation with
error prediction for multiple-linear regression.
The fourth and most important one is our novel
algorithms of Multiclassication with One-Against-All and
Multiclassication with Binary Comparison as variants for
support vector machine (SVM) classier for the trac ow
model.
The nal contribution is that none of the earlier works
uses neural network model to classify the trac control by
short section with single ILD.
3. Background
One of the short section detectors, RTMS, is capable of
producing some kind of trac parameters periodically for
each lane on the freeway. These parameters are volume,
speed, and occupancy.
(i) Volume shows the count of total vehicles passed
through this short section for one period.
(ii) Speed shows the average speed of total cars passed
through this short section for one period.
(iii) Occupancy shows the sum of the time; vehicles
occupy the short section divided by one period time.
3.1. Linear Multiple Regression Analysis Method as a Classier.
The objective of linear multiple regression analysis (LMRA)
is to dene which of the independent variables are important
on predicting the model [35]. Multiple regression analysis
provides a predictive equation
Y = a +
1
x
1
+
2
x
2
+ +
n
x
n
, (1)
where, a is interception constant,
i (i=1, 2, ...,n)
are standard-
ized partial regression coecients (reecting the relative
impact on the criterion variable). x
i (i=1, 2, ...,n)
are the metric
scores (i.e., interval or ratio data) of dierent independent
variables. Y is the single dependent variable structuring the
X
2
X
1
Optimal hyperplane
Maximum margin
Figure 1: Maximum margin of optimal hyperplane.
model. Equation (1) actually symbolizes a linear hyper plane.
The purpose of the LMRAclassier is to minimize the energy
function E for each data point, dened by LMS as in (2). The
parameters a and
i (i=1, 2,...,n)
are obtained from solving the
partial derivatives of
E =
_
_
y
p
_
a +
1
x
p
1
+
2
x
p
2
+ +
n
x
p
n
__
2
. (2)
for given data.
3.2. Support Vector Machine as a Classier. Another inno-
vative supervised pattern classier technique SVM was rst
proposed by Vapnik in 1995 [36]. The formulation applied
by SVM embodies the Structural Risk Minimization (SRM)
principle, which has been shown to be superior to traditional
Empirical Risk Minimization (ERM) principle [37]. While
SRM minimizes an upper bound on the expected risk,
ERM minimizes the error on the training data [38]. It is
the dierence which equips SVM with a greater ability to
generalize, which is the goal in statistical learning. SVMs were
developed to solve the classication problem (i.e., data must
belong to either Class 1(+1) or Class 2(1)) but recently they
have been extended to the domain of regression problems
[39].
The decision function in (3) determines the classes of all
input vectors (x = x
i
m
) [where i is the input and m is the
tuple indexes]. is xed feature-space transformation, W is
m dimensional weights, and b is bias term [40]:
D(x) = W
T
(x) + b. (3)
If there are multiple solutions, we should nd the
smallest generalization error. So, margin (i.e., smallest
distance between decision boundary and any of the samples)
should be found through [controlling the separability] [41]:
min
W
T
x + b
= 1. (4)
Although there are innite number of solutions to
separate hyperplanes, by maximizing the margin, Figure 1
shows only two decision functions satisfying (4).
Structure
Description of
decision regions
Exclusive or
problem
Classes with
meshed regions
General region
shapes
Most general
region shapes
Single layer
Two-layer
Three-layer
Half plane
bounded by
hyperplane
Arbitrary
(complexity
limited by #
of hidden
units)
Arbitrary
(complexity
limited by #
of hidden
units)
Structure
Description of
decision regions
Exclusive or
problem
Classes with
meshed regions
General region
shapes
Most general
region shapes
Single lay aa er
Tw TT o-lay aa er
Three-lay aa er
Half plane
bounded by
hy hh perplane
Arbitrary
(complexity
limited by #
of hidden
units)
Arbitrary
(complexity
limited by #
of hidden
units)
Figure 2: Classication interpretation of ANN model selecyion [51].
The margin is given by the perpendicular distance to the
closest point x
i
from the data set and we wish to optimize
the parameters W and b in order to maximize this distance
(using target t
i
). Thus the maximum margin solution is
found by solving
argmax
W,b
_
1
W
_
min
i
_
t
i
_
W
T
(x) + b
__
. (5)
The dominating approach for solving multiclass prob-
lems using SVM has been based on reducing single multi
class problem into multiple binary problems. For instance,
a common method is to build a set of binary classiers
where each classier distinguishes between one of the labels
to the rest. This approach is a special case of using output
codes for solving multi class problems [42]. However, while
multi class learning using output codes provides a simple and
powerful framework it cannot capture correlations between
the dierent classes since it breaks a multi class problem into
multiple independent binary problems [43].
The idea of casting multi class problems as a single
constrained optimization with a quadratic objective function
was introduced by Vapnik [44] and Watkins [45]. These
attempts to extend the binary case are achieved by adding
constraints for every class and thus the size of the quadratic
optimization is proportional to the number of categories in
the classication problems. The result is often a homoge-
neous quadratic problem which is hard to solve and dicult
to store. The idea of breaking constrained optimization prob-
lem into smaller problems was extended by Joachims [46].
Multiple class problems were then discussed by Sch olkopf
[47]. Todays panorama of SVM is well summarized by
Cristianini [48] and more completely by Sch olkopf and
Smola [49].
3.3. Backpropagation Neural Network as a Classier. Articial
neural networks (ANN) are used to serve two important
functions as pattern classiers and as nonlinear adaptive
V
3,4
W
4,2
W
1,1
V
1,1
Input
Hidden
X
1
X
2
X
3
Z
1
Z
2
Z
3
Z
4
Y
1
Y
1
Output
Figure 3: Simple Three-layered Neural Network Structure (3-4-2).
lters. Figure 2 gives a brief overview about the ANN
architecture for proper pattern classication [50, 51]. Since
the data is not guaranteed to be linearly separable (i.e.,
overlapped), three-layered Backpropagation NN (BPNN)
architecture is chosen in this paper as classier. This type
of architecture does not need any prior knowledge about
data (i.e., exemplar pattern initialization). Since there is no
need for distribution of the data (for fault-tolerance), asyn-
chronous update in weights and self-classication, BPNN is
suitable as supervised classier [32, 50]. In Figure 3, simple
three-layered BPNN is indicated.
In our paper work, we used alternative learning methods
for comparing the performances. They are Gradient Descent
with Momentum (gdm), Conjugate Gradient Descent (cgd),
Scaled Conjugate Gradient (scg), and Levenberg-Marquardt
Gradient Descent with momentum and Scaled Conjugate
Gradient. In [52], you can nd algorithms and related details
0
10
20
0
2
4
6
8
10
12
0
20
40
60
80
100
120
V
o
l
u
m
e
Speed
O
c
c
u
p
a
n
c
y
Congested trac
Figure 4: An example simulator data for Congested trac.
0
20
40
60
80
100
0
50
100
0
5
10
15
20
25
V
o
l
u
m
e
Flow trac
S
p
e
e
d
O
c
c
u
p
a
n
c
y
Figure 5: An example simulator data for Flow trac.
about the formulation and theory about these learning
methods.
4. Proposed Method
4.1. Data. For each trac condition, dierent values and
types of data can be acquired through the FHWAs (Federal
Highway Agency) TSIS 5.1 software (i.e., a CORSIM-based
microscopic trac simulation tool). Dierentiation of data
is supplied by conguring simulation tool parameters such
as distribution function of generated trac, lane speeds, car
distances, and incident creation intervals.
20
30
40
50
20
25
30
35 40
45
50
55
10
15
20
25
30
35
40
45
50
V
o
l
u
m
e
Dense trac
Speed
O
c
c
u
p
a
n
c
y
Figure 6: An example simulator data for Dense trac.
Each ILD sensors generates speed, volume, and occupancy
parameters values for congurable one period of time (i.e.,
60 seconds) for each lane on the freeway.
4.1.1. Congested Condition. Congested trac condition can
be observed by creating long-term incidents along the
freeway. According to the number of lanes they occupy,
incidents can be assorted. In our study, kinds of incident
scenarios are created for four-lane freeway. The sensors are
placed 100 feet upstream from the incident point. As seen
in Figure 4 occupancy reaches the maximum values, whereas
both speed and volume reach the minimum values.
4.1.2. Flow Condition. In order to generate the ow trac
condition, CORSIMs average speed input is determined
from high values (i.e., 90 km/h) and no incidents are created.
Figure 5 illustrates the ow condition. It is clear from the
gure that, when the speed is high and volume is low, the
occupancy approximately reaches to zero. Under the low
speed and high volume conditions, occupancy rises but never
reaches to the maximum value.
4.1.3. Dense Condition. Dense trac condition can be iden-
tied by generating low speed values, short-term incidents,
and decreased distances between the cars in trac. This
condition is observed between the ow and congested
conditions. As seen in Figure 6 the occupancy never reaches
to min or max values.
4.2. Multiple Regression Analysis. There exist two signicant
facts through the basis of this method. The prior one says
it is obvious that occupancy approximates to zero whether
the average speed of vehicles approximates to innity in
0
50
100
150
0
20
40
60
80
0
20
40
60
80
100
S
p
e
e
d
98.28.5x-2.5y
V
o
lu
m
e
O
c
c
u
p
a
n
c
y
Figure 7: Regression plane (red) for Congested trac.
ow trac condition or not. On the contrary, occupancy
goes to maximum value when speed approximates to zero in
congested trac ow condition.
It is also obvious that, until the volume reaches its
maximum, occupancy is also increasing. After the maximum,
the volume is going to be monotonically decreasing whereas
occupancy is still increasing.
Under these circumstances, in owtrac condition more
volume leads to more occupancy. However, in congested
trac ow, the less volume observed leads to more occupancy.
According to the facts mentioned above, occupancy is
dependent both on speed and volume.
Occupancy = a +
1
Volume +
2
Speed + Error. (6)
Equation (6) is not viable in real world aairs. Once
the trac seems to be in congested, both speed, and
volume can show the zero, although occupancy approaches
to maximum. Unless the vehicles exist along the freeway
(i.e., ow condition), speed, volume, and also the occupancy
indicate zero. A linear dependency of SVO for each trac
condition with respect to the discrepancy are modied from
(6) as in (7). For each ow condition a,
1
, and
2
parameters
are dierent,
Occupancy
condition
= a +
1
Volume +
2
Speed + Error.
(7)
4.2.1. Regression Planes. Using the CORSIM data, ,
1
, and
2
values for each trac condition are calculated using LMS
(Least Mean Squares) method. The obtained results can be
tracked through following subtitles.
0
10
20
30
40
50
60
0
50
100
150
0
10
20
30
40
50
60
70
70
80
90
100
Volume
0.05 + 0.3x 0.004y
S
p
e
e
d
O
c
c
u
p
a
n
c
y
Figure 8: Regression plane for Flow trac.
(a) Congested condition. The plane evaluated for congested
trac is depicted in Figure 7 and its equation is in
Occupancy
congested
= 98, 02 8, 5 Volume 2, 5 Speed.
(8)
Through the crosschecking (8), it can easily be seen that
occupancy reaches max (98.2) when the speed and volume
are marked as zero. This veries the expectations of the real
trac condition.
(b) Flow Condition. Figure 8 shows the plane obtained from
the ow trac. Its equation is dened as
Occupancy
ow
= 0, 05 + 0, 3 Volume 0, 004 Speed. (9)
In real trac, if no car is detected on the freeway during
the time period, detectors will produce the occupancy as 0
whereas speed and volume also share the same value. Through
the crosschecking (9), it can easily be seen that occupancy
reaches min (0.05) when the speed and volume are marked
as zero. This veries the expectations of the real trac
condition.
(c) Dense Condition. Figure 9 depicts the dense trac ow
conditions regression plane. Linear equation of the plane can
be dened as
Occupancy
dense
= 34.5 + 0, 58 Volume 0, 7 Speed.
(10)
(d) Combined Data and Regression Planes. Data and regres-
sion planes of all trac conditions are shown in Figure 10.
Three dierent methods are applied, respectively, for
classication of trac conditions using linear regression.
(1) Using Analytic geometry (AG). From the CORSIM
simulation, regression planes are dened for each trac
conditions. AG calculates the distances between all data
0
20
40
60
0
50
100
150
V
o
lu
m
e
34.5 + 0.58x 0.7y
Sp
eed
O
c
c
u
p
a
n
c
y
0
10
20
30
40
50
60
70
80
90
100
Figure 9: Regression plane for Dense trac.
points and each plane. The result (i.e., nearest plane) supplies
us to infer the class (status) of current trac.
From (8), we can obtain (11) for congested trac as
0 = 98, 02+8, 5Volume2, 5SpeedOccupancy
congested
,
d
congested
=
98.028, 5Volume2, 5SpeedOccupancy
_
8, 5
2
+2, 5
2
+1
2
.
(11)
From (9), we can obtain (12) for ow trac as
0 = 0, 05+0, 3Volume0, 004SpeedOccupancy
ow
;
d
ow
=
0, 050, 3Volume0, 004SpeedOccupancy
_
0, 3
2
+0, 004
2
+1
2
.
(12)
From (10), we can obtain (13) for dense trac as
0 = 34.5+0, 58Volume0, 7SpeedOccupancy
dense
,
d
dense
=
34.50, 58Volume0, 7SpeedOccupancy
_
0, 58
2
+ 0, 7
2
+ 1
2
.
(13)
(2) Estimating Occupancy without Error Prediction (EO). In
EO, linear regression (6) is used and error is assumed as zero.
Using this assumption occupancy estimator can be calculated
for each simulated trac condition (8), (9), (10) using the
real speed and real volume. After acquiring these estimators,
they are compared with the real occupancy. The one nearest
0
10
20
30
40
50
60
70
0
50
100
150
0
20
40
60
80
100
S
p
e
e
d
V
o
l
u
m
e
O
c
c
u
p
a
n
c
y
Figure 10: Composite regression planes of all trac.
to real occupancy becomes the result of our trac ow
condition.
(3) Estimating Occupancy with Error Prediction (EOwEP):
Another method for nding the trac ow class is extended
version of EO. This method calculates occupancy estimators
by adding the predicted error values.
EOwEP depends on the training data values and their
residuals. Nearest point is found by measuring L2 distance
between real trac data point and CORSIM simulated data
points from each trac condition. The residual of the found
point is assumed to approximate our error. As a result,
occupancy estimator for that condition can be determined.
So the obtained equation as
Occupancy
congested
= 98, 02 + 8, 5 Volume 2, 5
Speed + res
i
,
(14)
where i is the nearest point, found by (15), to the real data
point among the congested training data points
Min
_
(Vol
i
Vol)
2
+
_
Speed
i
Speed
_
2
+ (Occ
i
Occ)
2
_
.
(15)
4.3. Support Vector Machines. As mentioned in background,
SVM is generally applied to binary classication problems.
Since the addressed problem is mapping the input data into
one of three ow conditions, SVM can be seen inapplicable
at rst. However, without changing its calculation style, SVM
can also be used in multiclassication problems. The idea lays
behind multi classication is just using more than one SVM
and classing the data according to the outputs of multiple
SVMs. We use our two novel approaches for Multiple SVM.
Multiclassication with One-Against-All Method (OAAM).
The designated architecture (our novel approach) for this
method is composed of three SVMs. Each SVM represents
one of the trac ow conditions. Detector values can
be fed into each SVM, respectively. The bipolar output
BPNN Algorithm;
(i) Initialize weights (v
i, j
, w
j,k
) with small random values
(ii) Broadcast the input data to the input layer x
i
Feed-forward phase is starting here;
(i) Calculate the hidden layer unit signals
(f is sigmoid function in our work)
z
j in
=
n
i=1
x
i
v
i j
z
j
= f (z
j in
) =
2
1 + e
z
j in
1
(ii) Calculate the output unit signal
y
k in
=
p
j=1
z
j
w
jk
y
k
= f (y
k in
) =
2
1 + e
y
k in
1
Backpropagation phase is starting here;
(i) Calculate the residual. By using expected value for output signal (t
k
).
k
= (t
k
y
k
) f

(y
k in
) = (t
k
y
k
)
_
1
2
[1 + f (y
ki n
)][1 f (y
k in
)]
_
w
jk
=
k
z
j
where is learning rate. In the learning phase, w
jk
is updated according to learning rule.
(ii) Then calculate reectance of the residuals and propagate it to the input weights.
j
=
m
k=1
k
w
jk
f

(z
j in
)
v
i j
=
j
x
i
(iii) All weights are updated with learning rule (i.e., gdm/scg).
w
jk
(new) = w
jk
(old) + w
jk
v
i j
(new) = v
i j
(old) + v
i j
(iv) Test the stopping condition (i.e., reaching to goal; predened Mse mean square error- of the total residuals)
Algorithm 1
[+1 (Class 1), 1 (Class 2)] indicates if the input data
belongs to this SVM or not.
For modeling OAAM, data is acquired from TSIS 5.1.
Decision is made through. KKT (Karush-Khun-Tucker
method for minimizing Quadratic Problem solution for
SVM) [45] .
min
_
1
2
W
2
_
. (16)
(a) Congested SVM (C-SVM). In the training, C-SVM target
is trained with +1 whereas the other SVM targets are
trained with 1. Plane for C-SVM can be shown in
Figure 11. Its equation is
D
_
vol, spd, occ
_
= 0.1235 vol 0.0443 spd + 0.1052
occ 0.9996.
(17)
(b) Dense SVM (D-SVM). In the training, D-SVM target is
trained with +1 whereas the other SVM targets are trained
with 1. Plane for D-SVM can be shown in Figure 12. Its
equation is
D
_
vol, spd, occ
_
= 2.6577 vol 1.7462 spd 0.2456
occ 0.9998.
(18)
(c) Flow SVM (F-SVM). In the training, F-SVM target is
trained with +1 whereas the other SVM targets are trained
with 1. Plane for F-SVM can be shown in Figure 13. Its
equation is in (19);
D
_
vol, spd, occ
_
= 0.0105 vol + 0.0788 spd 0.3218
occ + 1.0003.
(19)
After training is done, each member of Multiple-SVMs
has its own linear planes for OAAM. In order to get the
correct class of the queried data is just easy as to look at
the responses of each SVMs. The +1 response of the SVM
claries the class of the data. For instance if F-SVMs response
is +1 (i.e., the others are expected as 1) then the class
of the data belongs to ow trac.
However, this is not sucient alone. Following cases
must be taken into account and supplemented approaches
must be applied in these conditions.
(i) If at least two SVMs produce output +1, then the
one with the longest L2 distance from data point
to its own SVM plane can be picked out. Since the
furthest distance fromthe data point to plane exposes
the stricter and more accurate data for current trac
ow condition, we choose the furthest one.
(ii) If all SVMs produce output 1, the one with the
shortest L2 distance to SVM plane must be chosen.
0
10
20
30
40
50
60
70
0
50
100
150
0
20
40
60
80
100
S
p
eed
9.5019 + 1.1739x + 0.4211y
V
o
l
u
m
e
O
c
c
u
p
a
n
c
y
Figure 11: Congested Trac Plane for C-SVM of OAAM.
0
10
20
30
40
50
60
70
0
50
100
150
0
10
20
30
40
50
60
70
80
90
100
Volum
e
4.0708 + 10.8212x 7.1099y
S
p
e
e
d
O
c
c
u
p
a
n
c
y
Figure 12: Dense Trac Plane for D-SVM of OAAM.
Producing output 1 implies that the data does not
belong to the current SVMs trac condition. Since,
the shortest distance from data point to SVM plane
is the nearest one to +1 (Class 1), we can pick this
SVM out as our desired class.
Multiclassication with Binary Comparison (BC). Our other
novel approach applied to classify our data is again composed
of three SVMs. However, it can be distinguished from the
OAAM method by its architecture. In BC method, each
SVM is responsible for the following binary combinations,
respectively; (Congested, Dense), (Congested, Flow), and
(Flow, Dense). This multi classication method depends
on basic voting principle. There exist three candidates (i.e.,
Congested, Flow, and Dense trac conditions) and three
voters (each SVM). The one among the candidates wins the
election if it gets the most votes from voters.
0
20
40
60
0
50
100
150
0
10
20
30
40
50
60
70
80
90
100
V
o
l
u
m
e
Speed
3.1084 + 0.0326xx + 0.2448y
O
c
c
u
p
a
n
c
y
Figure 13: Flow Trac Plane for F-SVM of OAAM.
(a) Congested-Dense (CD-SVM). In the training, CD-SVM
target +1 indicates that data belongs to Class 1 (i.e.,
congested trac). The other alternative of 1 indicates
that the data belongs to Class 2 (i.e., dense trac).
So the voter gives its vote to Congested candidate if its
output equals +1. In adverse condition, it gives to Dense
candidate. Figure 14 shows the decision plane for CD-SVM.
The plane equation is
D
_
vol, spd, occ
_
= 0.0642 vol 0.0440 spd 0.0041
occ + 1.8692.
(20)
(b) Congested-Flow SVM (CF-SVM). In the training, CF-
SVM target +1 indicates that data belongs to congested
trac. The other alternative of 1 indicates that the data
belongs to ow trac. Figure 15 shows the decision plane for
CF-SVM. The plane equation is
D
_
vol, spd, occ
_
= 0.0044 vol 0.0160 spd + 0.0711
occ 0.9998.
(21)
(c) Flow-Dense SVM (FD-SVM). In the training, FD-SVM
target +1 indicates that data belongs to ow trac. The
other alternative of 1 indicates that the data belongs to
dense trac. Figure 16 shows the decision plane for FD-
SVM. The plane equation is
D
_
vol, spd, occ
_
=0.0105 vol + 0.0788 spd 0.3218
occ +1.0003.
(22)
A supplemented way must be found for handling the
situation when all candidates have the equal vote. Using the
longest distance from data point to each SVMs plane will
0
20
40
60
0
50
100
150
0
10
20
30
40
50
60
70
80
90
100
Speed
455.9024 15.6585x 10.7317y
V
o
lu
m
e
O
c
c
u
p
a
n
c
y
Figure 14: Plane of CD-SVM for BC.
0
50
0
50 100
150
0
10
20
30
40
50
60
70
80
90
100
V
o
l
u
m
e
Speed
14.0618 + 0.0618x + 0.2250y
O
c
c
u
p
a
n
c
y
Figure 15: Plane of CF-SVM for BC.
give us the most accurate class for that data point. Then the
SVM with the longest distance will choose the leader class.
Backpropagation Neural Network Architectures . The network
architecture in our work is selected as three-layered back-
propagation neural network (BPNN). three groups of data
are used as input. The rst group is fed into the network as
training data set. The next one is used as verication data
set. The last one is used as query set (i.e., test set) to measure
the classication performance of the neural net.Training data
set is obtained from TSIS 5.1 simulator. This data is veried
through the verication set also gathered from the simulator.
The test set is gathered from single sensor planted on the
2.Ring road in Istanbul. Test set indicates real data. Our
model is trained and veried through the simulated data and
tested with real trac information.In the learning phase we
have used gradient descent with momentum (Gdm) update
50
0 50 100 150
0
10
20
30
40
50
60
70
80
90
100
V
o
lu
m
e
Speed
3.1084 + 0.0326xx + 0.2448y
O
c
c
u
p
a
n
c
y
Figure 16: Plane of FD-SVM for BC.
function in (23) (i.e., learning rule). This base rule is used for
all architectures
w
jk
(t + 1) = w
jk
(t) +
k
z
j
+
_
w
jk
(t) w
jk
(t 1)
_
,
v
i j
(t + 1) = v
i j
(t) +
j
x
i
+
_
v
i j
(t) v
i j
(t 1)
_
.
(23)
In the learning phase, we have used the algorithms of only
Gdm, Levenberg-Maquardt with Gdm and Scaled Conjugate
Gradient with Gdm [52]. Matlab 7.6.0 (Rev.2008a) Neural
Network Toolbox where our related work algorithms and
learning rules are ready is used.
We have used default max iteration as 1000, Mse goal
value as 10
4
, and verication iteration count as 6. The
network architecture is performed as three-layered models
of 3-10-3, 3-20-3, and 3-50-3. Since one hidden layer usage
is almost identical with two hidden layered architecture, we
prefer to use single hidden layer model. For the connec-
tions lying between Input/Hidden Layers are trained with
LogSig or PureLin activation functions. LogSig activation
function normalized the input and applies Sigmoid. Pure
Linear activation function transfers input to hidden.We give
the performance analysis of our neural net architectures
in Table 1. Gray labeled values are average performances
obtained from 100 runs of neural net models. According
to the average performances, our 3-20-3 backpropagation
model gives the best result for classication.
5. Experimental Work
Turkey General Directorate of Highways has carried out
works in managing trac in Istanbul hence trac manage-
ment center is planned to be opened at the end of December
2007. The system is installed by Turkeys leading electronics
company ASELSAN Inc. In Istanbul trac management
system, microwave radar type detectors are used for occu-
pancy, speed, and trac volume measurements. There exist
almost 30 radio transmission microwave sensors along the
Table 1: Backpropagation Neural Net Performances.
GDM LM SCG
LogSig-PurLin PurLin-PurLin LogSig-PurLin PurLin-PurLin LogSig-PurLin PurLin-PurLin
BackPrp 3-10-3
Cong 100 100 0.9604 0.9472 100 0.4073
Dense 0.5733 0.5612 0.5134 0.5671 0 0
Flow 0.9502 0.9189 0.9210 0.9470 0 0.062
Total (Rate/Iteration) 0.9678/1000 0.9258/1000 0.9352/360 0.9170/5 0.0068/156 0.0615/15
BackPrp 3-20-3
Cong 0.6227 0.9472 100 0.9393 100 0.9393
Dense 0.6914 0.6878 0.4677 0.6778 0.8531 0.6725
Flow 0.9804 0.8950 0.7492 0.8967 0.9914 0.9468
BackPrp 3-50-3
Cong 0.9721 0.9393 0.9732 0.9893 0.9789 0.9472
Dense 0.4552 0.6686 0.1256 0.6800 0.6318 0.6711
Flow 0.9571 0.9869 0.9899 0.9907 0.9879 0.9867
Table 2: Estimation Performance of Our Model for Each Trac
condition.
Method Flow Dense Congested
Analytic Geometry 99% 84% 70%
Without Error Prediction 100% 78% 35%
Error Prediction 99% 24% 46%
SVM with One-Against-All Method 100% 46% 97%
SVM with Binary Comparison 100% 74% 100%
BackPrp NN-SCG (3-20-3) 99.14% 85.31% 100%
BackPrp NN-LM (3-50-3) 99.07% 68% 98.93%
BackPrp NN-GDM (3-10-3) 95.02% 57.33% 100%
2.Ring Road. The measurements recorded at critical points
on the road are transmitted periodically to the management
center [53].
Occupancy, speed, and trac volume measurements
are both recorded in the database, and the trac status
according to these measurements is mapped to color codes
and displayed on a large screen. Whenever any congested
state or dense state occurs along the road, alarms with
severity levels are generated and the operators are informed.
Information regarding the trac status is also displayed on
LED-based Variable Message Signs (VMS) on the road.
During one day, data from 7 dierent microwave sensors
is acquired every 60 seconds periodically. Occupancy, volume
and speed measurements are shown in Figure 17. Videos
from the trac surveillance cameras, which are capable of
watching the places that microwave sensors are installed,
are also captured in order to synchronize them with the
sensors measurements. Then, measurements are grouped as
congested, dense and ow according to these videos. [We
have used training data set of total size 18.000 x 3 (i.e., each
trac condition has balanced data subset size of 6000 x 3)
and we have used testing data set of total size 55.000 x 3, and
3 indicates SVO values]. These data sets are used to for the
Table 2.
0
100
200
0 5 10 15 20 25 30 35 40 45 50
0
10
20
30
40
50
60
70
80
90
100
Speed
Real trac (2. Ring Road Istanbul)
V
o
l
u
m
e
O
c
c
u
p
a
n
c
y
Figure 17: Real Trac data from 2.Ring Road in Istanbul (by 7
sensors).
6. Conclusion
Eight methods discussed above have applied the one day data
captured from 7 sensors on 2.Ring Road in Istanbul. Their
performance for each trac condition can be seen in Table 2.
Although our novel approaches OAAM and BC for Mul-
tiple SVM have good performances, our BPNN architecture
of 3-20-3 with SCG (3-20-3) is better than all other seven
methods. As a result of this comparison, this method has
chosen and installed for the evaluation of trac condition in
Istanbul trac management center (at Istanbul 2.Ring road).
According to the result it calculates, our algorithm colors
the road map for each status changes on GIS projection
screens. Once congested level is detected, it produces an
alarm with high severity to inform the operators. Never-
theless, dense level also creates alarm with lower severity.
After verication of this assessment by operator, VMS can be
supplied by messages to inform drivers about the road ow
condition.
References
[1] F. L. Hall, Trac stream characteristics, in USA Transporta-
tion Research Board. Trac Flow Theory, A State of the Art
Report, 1992.
[2] B. D. Greenshields, A study of trac capacity, Highway
Research Board Proceedings, vol. 14, pp. 448477, 1935.
[3] A. Reuschel, Fahrzeugbewegungen in der Kolonne, Oesterre-
ichisches Ingenieur, vol. 4, no. 3/4, pp. 193215, 1950a.
[4] A. Reuschel, Fahrzeugbewegungen in der Kolonne bei gle-
ichformig beschleunigtem oder verzogertem, Leitfahrzeug,
Zeitschrift des Oesterreichischen Ingenieurund Architekten-
Vereines, vol. 95, no. 9, pp. 5962, 1950.
[5] A. Reuschel, Fahrzeugbewegungen in der Kolonne bei gle-
ichformig beschleunigtem oder verzogertem, Leitfahrzeug,
Zeitschrift des Oesterreichischen Ingenieurund Architekten-
Vereines, vol. 97-98, no. 10, pp. 7377, 1950.
[6] J. G. Wardrop, Some theoretical aspects of road trac
research, Proceedings Institute of Civil Engineers, Part II, vol.
1, pp. 325378, 1952.
[7] L. A. Pipes, An operational analysis of trac dynamics,
Journal of Applied Physics, vol. 24, no. 3, pp. 274281, 1953.
[8] M. J. Lighthill and G. B. Whitham, On kinematic waves. II.
A theory of trac ow on long crowded roads, Proceedings of
the Royal Society of London, pp. 317345, 1955.
[9] G. F. Newell, Mathematical models for freely-owing high-
way trac, Journal of the Operations Research Society of
America, vol. 3, pp. 176186, 1955.
[10] F. V. Webster, Trac signal settings, Road Research Techni-
cal, Road Research Laboratory, London, UK, 1958.
[11] L. C. Edie and R. S. Foote, Trac owin tunnels, Proceedings
of Highway Research Board, vol. 37, pp. 334344, 1958.
[12] R. E. Chandler, R. Herman, and E. W. Montroll, Trac
dynamics: studies in car following, Operations Research, vol.
6, pp. 165184, 1958.
[13] D. C. Gazis, The origins of trac theory, Operations
Research, vol. 50, no. 1, pp. 6977, 2002.
[14] J. R. Correa, A. S. Schulz, and N. E. Stier-Moses, Selsh
routing in capacitated networks, Mathematics of Operations
Research, vol. 29, no. 4, pp. 961976, 2004.
[15] H. Greenberg, An analysis of trac ow, Operations
Research, vol. 7, pp. 7985, 1959.
[16] L. C. Edie, Car following and steady-state for non-congested
trac, Operations Research, vol. 9, pp. 6676, 1961.
[17] R. T. Underwood, Speed, volume, and density relationships:
quality and theory of trac ow, Yale Bureau of Highway
Trac, pp. 141188, 1961.
[18] J. S. Drake, L. Schofer, and A. D. May, A statistical analysis of
speed density hypotheses, Highway Research Record, vol. 154,
pp. 5387, 1967.
[19] P. Athol, Interdependence of certain operational characteris-
tics within a moving trac stream, Highway Research Record,
vol. 72, pp. 5887, 1965.
[20] V. F. Hurdle and P. K. Datta, Speeds and ows on an
urban freeway: some measurements and a hypothesis, in
Transportation Research Record 905, pp. 127137, TRB, NRC,
Washington, DC, USA, 1983.
[21] B. N. Persaud and V. F. Hurdle, Some new data that
challenge some old ideas about speed-ow relationships, in
Transportation Research Record, vol. 1194, pp. 191198, TRB,
NRC, Washington, DC, 1988.
[22] F. L. Hall and L. M. Hall, Capacity and speed ow analysis of
the Queue in Ontario, in Transportation Research Record, vol.
1287, pp. 108118, TRB, NRC, Washington, DC, USA, 1990.
[23] J. H. Banks, Freeway speed-ow concentration relation-
ships: more evidence and interpretations, in Transportation
Research Record, vol. 1225, pp. 5360, TRB, NRC, Washington,
DC, USA, 1989.
[24] W. S. Smith, F. L. Hall, and F. O. Montgomery, Comparing the
speed-owrelationship for motorways with newdata fromthe
M6, Transportation Research Part A, vol. 30, no. 2, pp. 89101,
1996.
[25] F. L. Hall, V. F. Hurdle, and J. H. Banks, Synthesis of recent
work on the nature of speed-ow and ow-occupancy (or
density) relationships on freeways, in Transportation Research
Record, vol. 1365, pp. 1218, TRB, NRC, Washington, DC,
USA, 1992.
[26] E. A. Wemple, A. M. Morris, and A. D. May, Freeway capacity
and level of service concepts, in Proceedings of International
Symposium on Highway Capacity, pp. 439455, Karlsruhe,
Germany, 1991.
[27] F. L. Hall and F. O. Montgomery, Investigation of an
alternative interpretation of the speed-ow relationship for
U.K. motorways, Trac Engineering and Control, vol. 34, no.
9, pp. 420425, 1993.
[28] M. Zhou and F. L. Hall, Investigation of speed-ow relation-
ship under congested conditions on a freeway, Transportation
Research Record, no. 1678, pp. 6472, 1999.
[29] J. H. Banks, Investigation of some characteristics of congested
ow, Transportation Research Record, no. 1678, pp. 128134,
1999.
[30] J. H. Banks, Review of empirical research on congested
freeway ow, Transportation Research Record, no. 1802, pp.
225232, 2002.
[31] F. Yuan and R. L. Cheu, Incident detection using support
vector machines, Transportation Research Part C, vol. 11, no.
3-4, pp. 309328, 2003.
[32] C. Oh and S. G. Ritchie, Recognizing vehicle classication
information from blade sensor signature, Pattern Recognition
Letters, vol. 28, no. 9, pp. 10411049, 2007.
[33] R. Mussa, V. Kwigizile, and M. Selekwa, Probabilistic neural
networks application for vehicle classication, Journal of
Transportation Engineering, vol. 132, no. 4, pp. 293302, 2006.
[34] E. I. Vlahogianni, J. C. Golias, and M. G. Karlaftis, Short-
term trac forecasting: overview of objectives and methods,
Transport Reviews, vol. 24, no. 5, pp. 533557, 2004.
[35] http://www.isical.ac.in/ddroy/multistat.doc.
[36] V. N. Vapnik, The Nature of Statistical Learning Theory,
Springer, New York, NY, USA, 1995.
[37] S. R. Gunn, M. Brown, and K. M. Bossley, Network
performance assessment for neurofuzzy data modeling, in
Intelligent Data Analysis, vol. 1208 of Lecture Notes in Com-
puter Science, pp. 313323, 1997.
[38] S. R. Gunn, Support vector machines for classication and
regression, Tech. Rep., Image Speech and Intelligent System
Research Group, University of Southampton, Southampton,
UK, 1998.
[39] V. Vapnik, S. Golowich, and A. Smola, Support vector
method for function approximation, regression estimation,
and signal processing, in Advances in Neural Information
Processing Systems, vol. 9, pp. 281287, MIT Press, Cambridge,
Mass, USA, 1997.
[40] S. Abe, Support Vector Machines For Pattern Recognition, 2005.
[41] C. Bishop, Pattern Recognition and Machine Learning,
Springer, New York, NY, USA, 2007.
[42] T. G. Dietterich and G. Bakiri, Solving multiclass learning
problems via error correcting output codes, Journal of
Articial Intelligence Research, vol. 2, pp. 263286, 1995.
[43] K. Crammer and Y. Singer, On the algorithmic implementa-
tion of multiClass kernel based vector machines, Journal of
Machine Learning Research, vol. 2, pp. 265292, 2001.
[44] V. N. Vapnik, Statistical Learning Theory, John Wiley & Sons,
New York, NY, USA, 1998.
[45] J. Weston and C. Watkins, Support vector machines for
multi-class pattern recognition, in Proceedings of the 7th
European Symposium on Articial Neural Networks, 1999.
[46] T. Joachims, Making large-scale support vector machine
learning practical, in Advances in Kernel MethodsSupport
Vector Learning, B. Scholkopf, C. Burges, and A. Smola, Eds.,
MIT Press, 1998.
[47] B. Sch olkopf, Statistical Learning and Kernel Methods,
Microsoft Research, 2000.
[48] N. Cristianini and J. S. Taylor, An Introduction to Support
Vector Machines, and Other Kernel-Based Learning Methods,
Cambridge University Press, Cambridge, UK, 2000.
[49] B. Sch olkopf and A. Smola, Learning with Kernels, MIT Press,
Cambridge, Mass, USA, 2002.
[50] L. Fausett, Fundementals of Neural Networks: Architectures,
Algorithms and Applications, Prentice-Hall, Upper Saddle
River, NJ, USA, 1994.
[51] A. K. Jain, J. Mao, and K. M. Mohiuddin, Articial neural
networks: a tutorial, Computer, vol. 29, no. 3, pp. 3144, 1996.
[52] S. Haykin, Neural Networks: A Comprehensive Foundation,
Prentice-Hall, Upper Saddle River, NJ, USA, 1998.
[53] E. Dorken, D. M. Barut, and B. T. Oranc, Integrated elec-
tronic tolling and trac management on Turkish highways,
in Proceedings of 12th World Congress on Intelligent Transport
Systems, p. 2901, San Francisco, Calif, USA, 2005.

Eurasip Journal

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Eurasip Journal

Uploaded by

Copyright:

Available Formats

Vehicular Ad Hoc Networks

Guest Editors: Hossein Pishro-Nik, Shahrokh Valaee,

Figure 23: Face orientation.

, des] means the state of a vehicle at

is the action of the trac

| [node, dir, pos, des], L,

, des]) gives the probability that the trac

at the next time step given the current state

, des]) is a reward function as follows: if L =

= Red, which means the vehicle moving to a green

) is the number of times the light

after such a transiting procedure.

([node, dir, pos, des], [node

, des]) is the reward

([node, dir, pos, des],

, des]) indicates the reward from the queue

([node, dir, pos, des], [node

Figure 19: Relation between distance and bandwidth using the

Figure 9: Position of preceding vehicle during cornering, with

, velocity dierence v, a safe

98.028, 5Volume2, 5SpeedOccupancy

0, 050, 3Volume0, 004SpeedOccupancy

34.50, 58Volume0, 7SpeedOccupancy

You might also like