Travel Time Studies With Global Positioning and Geographic

TRANSPORTATION
RESEARCH
PART C
Transportation Research Part C 6 (1998) 101127
Travel time studies with global positioning and geographic

information systems: an integrated methodology
Cesar A. Quiroga a,*, Darcy Bullock b,1
a
Texas Transportation Institute, Texas A&M University System, 3500 N.W. Loop 410, Suite 315, San Antonio,
TX 78229, USA
b
School of Civil Engineering, Purdue University, West Lafayette, IN 47907, USA
Received 24 March 1997; received in revised form 18 June 1998
Abstract
The paper describes a new methodology for performing travel time studies using global positioning
system (GPS) and geographic information system (GIS) technologies. It documents the data collection,
data reduction, and data reporting procedures, as well as analyses that illustrate the capabilities of the
GPS/GIS methodology. The data collection procedure uses GPS receivers to automatically collect time,
local coordinates, and speed at regular sampling periods, for example every one second. The data reduction
procedure lters and aggregates GPS data to compute travel time and speed along highway segments. The
data reporting procedure uses a GIS-based management information system to dene queries, tabular
reports, and color coded maps to document travel time data along these highway segments. These proce-
dures have been implemented in three metropolitan areas in Louisiana: Baton Rouge, Shreveport, and
New Orleans. In these cities, more than 180,000 segment travel time and speed records were derived
between 1995 and 1996 from nearly three million GPS data points collected on 30,000 miles of travel time
runs along 300 miles of urban highways. The three analyses included in the paper to assist in the process of
understanding the GPS/GIS methodology are the following: segment lengths, sampling rates, and central
tendency. The segment length analysis examines the eect of using dierent highway segment lengths and
shows that relatively short segments (0.20.5 miles long) are needed to detect localized trac eects. These
trac disturbances become visible only when segment lengths are at most half the length of the associated
disturbance. This means that traditional link-based segments, which are typically longer than 0.5 miles, are
not sucient to characterize localized eects properly. The sampling rate analysis addresses the eect of
collecting GPS data at dierent sampling periods and shows that for a segment to have GPS data asso-
ciated with it, the GPS sampling period should be smaller than half the shortest travel time associated with
the segment. The analysis also shows a tradeo between sampling rates and segment speed reliability, and
* Corresponding author. Tel.: +1-210-731-9938; Fax: +1-210-731-8904; e-mail: c.quiroga@ttimail.tamu.edu

1
Tel.: +1-765-494-2204; Fax: +1-765-496-1105; e-mail: darcy@purdue.edu
0968-090X/98/$19.00 # 1998 Elsevier Science Ltd. All rights reserved.

PII: S 0968-090X (9 8 ) 0 0 0 1 0 - 2
102 C.A. Quiroga, D. Bullock/Transportation Research Part C 6 (1998) 101127
emphasizes the need for even shorter GPS sampling periods (12 s) in order to minimize errors in the
computation of segment speeds. The central tendency analysis compares harmonic mean speeds and med-
ian speeds and shows that median speeds are more robust estimators of central tendency than harmonic
mean speeds. # 1998 Elsevier Science Ltd. All rights reserved.
Keywords: Travel time; Travel time studies; GPS; GIS; Global positioning systems; Geographic information systems;
Segmentation; Segment speed; Sampling rates; Accuracy; Errors; Median speed; Harmonic mean speed
1. Problem denition
Two techniques have traditionally been used to measure travel time: the license plate technique
and the oating car technique (May, 1990; Robertson, 1994). With the automation provided by
computers and other electronic devices, additional techniques have emerged over the years,
including automatic vehicle identication (AVI), automatic vehicle location (AVL), cellular
phone tracking, and video imaging (Liu and Haines, 1996; Turner, 1996). For discussion purposes,
these techniques can be grouped into two categories: roadside techniques and vehicle techniques.
Vehicle techniques, particularly those that involve the use of probe vehicles, are perhaps the
most commonly used techniques for collecting travel time data. These techniques are con-
ceptually very simple. However, most implementations tend to be quite labor intensive. For
example, in the traditional oating car technique two technicians are required in the car: one of
them to drive the vehicle, and the other one with a clipboard and a stopwatch to manually record
distance driven and the location and time stamp associated with individual checkpoints. Accuracy
with this technique varies from technician to technician. In addition to this, problems such
as missing checkpoints or inaccurately marked checkpoints are common. Nowadays, distance
measuring instruments (DMIs) can be attached to the vehicle's transmission to automatically
record distance, time, and speed. In this case, only one technician is needed in the probe
vehicle. However, DMIs are not free of diculties. For example, Benz and Ogden (1996) reported a
need for weekly calibrations of the probe vehicles and constant verication of the vehicle's
tire pressure (an inaccurate tire pressure provides inaccurate speed and distance readings). This
indicates the vehicle dependence of the DMI approach. In addition to this, DMIs tend to be rela-
tively expensive. Including a notebook computer for data storage, each unit costs about US $2000.
Recent research (Quiroga and Bullock, 1996; Quiroga, 1997) has demonstrated the feasibility of
using global positioning system (GPS) and geographic information system (GIS) technologies for
automating the travel time data collection, reduction, and reporting when using a probe vehicle.
The new automated procedures provide consistency, automation, ner levels of resolution, and
better accuracy in measuring travel time and speed than traditional techniques. As a result, large
amounts of reliable travel time and speed data can be collected and processed. GPS receivers can
record location and speed automatically at regular sampling periods. Consequently, only one
technician is required in the vehicle. However, because GI'S receivers record location as longitude
latitude pairs, additional tools are required to provide a linear reference to these point locations.
Fortunately, GIS can be used for this purpose, with the added advantage that the resulting travel
time data can be entered directly into existing geographic databases.
In most travel time studies involving probe vehicles, only one vehicle is used to characterize
trac ow along a specic direction of travel. Because lane dierentiation is not required, a GPS
C.A. Quiroga, D. Bullock/Transportation Research Part C 6 (1998) 101127 103
positional accuracy of around 23 m is sucient. A GPS receiver complying with this requirement
costs less than US $2000 (including dierential correction and a laptop computer for data sto-
rage), making GPS data collection price competitive with DMIs. GPS receivers are actually much
more versatile than DMIs because they can be used for additional purposes like sign inventories
and eet tracking with little or no modication in hardware conguration.
GPS les (and DMI les for that matter) tend to have huge numbers of records, particularly if
GPS data is collected at short sampling periods, say every 1 s. While collecting GPS data at such
short intervals has a very minor eect in terms of cost, it can pose a signicant data management
problem. A possible solution would be to keep for analysis only time and location data associated
with specic checkpoints along the corridor. However, this approach is appropriate only when
total travel time data is of interest and when localized variations in trac behavior are not of
concern. Moreover, because only a handful of travel time data points would be used, we can
expect the accuracy of the results to closely resemble those of the traditional, manual clipboard
and stopwatch approach.
Another solution would be to keep for analysis all the GPS data and build detailed speedtime
or speeddistance proles along corridors (Guo and Poling, 1995; Zito et al., 1995; Harding et al.,
1996). Because one gets a very detailed and rich picture of the trac situation (Harding et al.,
1996), this solution is suitable for detecting localized variations in trac behavior. However, it
overlooks the inevitability that trac may vary greatly from one day to the next both in space
and time. For example, in Fig. 1, which shows a sample of speedspace proles along the I-10,
I-12, and Airline Hwy corridors in Baton Rouge, LA, variations of more than 50 mph are evident
at some locations. Even where trac tends to be fairly uniform, variations of up to 20 mph are
noticeable. Dates and times associated with the proles shown in Fig. 1 are summarized in
Table 1
Dates and times associated with the runs shown in Fig. 1
Date Start time End time Total travel Date Start time End time Total travel
time (min) time (min)
(a) I-10 and I-12 EB (AM peak) (b) I-10 and I-12 EB (PM peak)
09/28/95 7:21:57 am 7:26:33 am 4:36 10/16/95 5:11:51 pm 5:21:33 pm 9:42
10/02/95 7:40:19 am 7:44:53 am 4:34 10/18/95 4:36:25 pm 4:46:11 pm 9:46
10/03/95 7:10:12 am 7:15:13 am 5:01 10/23/95 4:58:59 pm 5:10:10 pm 11:11
10/04/95 7:25:00 am 7:29:56 am 4:56 01/11/96 4:56:18 pm 5:09:39 pm 13:21
01/08/96 7:18:29 am 7:23:09 am 4:40 03/25/96 4:55:12 pm 5:03:16 pm 8:04
01/23/96 7:09:25 am 7:13:41 am 4:16 03/28/96 5:05:33 pm 5:11:43 pm 6:10
02/01/96 7:15:16 am 7:19:59 am 4:43
02/12/96 7:30:20 am 7:35:01 am 4:41
(c) I-12 WB (AM peak) (d) Airline Hwy SB (PM peak)
09/28/95 7:02:19 am 7:15:02 am 12:43 11/28/95 4:47:00 pm 5:00:56 pm 13:56
10/02/95 7:21:13 am 7:35:01 am 13:48 02/06/96 5:11:18 pm 5:26:20 pm 15:02
10/03/95 7:06:54 am 7:28:42 am 21:48 03/07/96 4:54:03 pm 5:03:44 pm 9:41
10/04/95 7:07:57 am 7:19:28 am 11:31 03/26/96 4:44:40 pm 4:58:45 pm 14:05
01/23/96 7:40:15 am 7:54:00 am 13:45 04/26/96 4:39:20 pm 4:54:15 pm 14:55
02/01/96 7:33:53 am 7:52:49 am 18:56
02/12/96 8:11:48 am 8:22:45 am 10:57
Fig. 1. Spacespeed proles using GPS on selected corridors in Baton Rouge, LA.
Table 1. Thus, keeping all GPS data for analysis may actually become counterproductive and
some kind of aggregation is both desirable and necessary to facilitate data interpretation.
2. Aggregating speed and travel time
2.1. Spatial model
In conducting travel time studies using GPS and GIS, the rst step is to obtain a good base
vector map with links to a database. One approach is to use existing planning maps like the
Topologically Integrated Geographic Encoding and Referencing (TIGER) les developed by the
US Census Bureau [Fig. 2(a)]. Unfortunately, such maps only provide a crude representation of
the corridors and their surroundings. To illustrate this point, Fig. 2(b) shows a map of the same
area obtained using GPS equipment having a positional accuracy of 25 m spherical error
probability (SEP). Notice that the oset of some of the physical discontinuities in the rst map
[Fig. 2(a)] with respect to those in the second map [Fig. 2(b)] is quite large: between 0.1 and
Fig. 2. Digital representation of I-10 between Acadian and College Drive in Baton Rouge, LA.
0.2 miles. This would also be the oset associated with GPS data. If we tried to calculate travel
time and speed using the physical discontinuities shown in Fig. 2(a) as checkpoints, the resulting
errors would be unacceptably large.
Alternatively, generating a network map using GPS is simple. It involves driving a vehicle equip-
ped with a GPS receiver to survey both directions of travel, as well as on-ramps, o-ramps, inter-
changes, and signalized intersecting streets [Fig. 3(a)]. To avoid problems associated with selective
availability, it is advisable to use dierential correction. The resulting GPS data is imported into a
GIS map to create a directional centerline network map [Fig. 3(b)]. Because this base map is
constructed directly from GPS data, GPS data collected during future travel time studies is guar-
anteed to match the vector base map within a tolerance dened by the GPS equipment positional
accuracy.
Traditional travel time studies record travel time and average speed between checkpoints along
the study route. To formalize these checkpoint denitions in the GIS map and link them to a
relational database, we can use three simple rules. First, we establish a checkpoint at all physical
discontinuities such as signalized intersections, signicant unsignalized intersections, lane drops,
on-ramps, and o-ramps [Fig. 3(c)]. Second, we segment the section of road between physical
discontinuity checkpoints so that there are nominally n checkpoints every mile [Fig. 3(d)]. Finally,
we link each discrete segment to a relational database by assigning unique identication numbers
to each segment [Fig. 3(e)]. For example, segment 12444 in Fig. 3(e) always represents the section
of I-10 east bound (EB) immediately before the I-10 and I-12 split. By creating these unique
identiers, each segment can have xed data such as number of lanes and posted speed limit
associated with it. It can also be used to index travel time data from travel time studies performed
on dierent dates and times.
Theoretically, the number of checkpoints per mile, n, can be set at any value. In practice, it is
necessary to dene a balance between GPS sampling rate, segment length, and vehicle speed.
Section 4.2 analyzes this situation with greater detail.
2.2. Mathematical model
Fig. 4 shows the timedistance diagram of a probe vehicle as it traverses a segment of length L.
The GPS equipment in the probe vehicle receives information from the constellation of GPS
satellites on a continuous basis and computes coordinates and speed values at discrete intervals.
In order to compute segment travel time and speed, we could interpolate the time stamps of the
two GPS points located immediately before and after the segment entrance, and the time stamps
of the two GPS points located immediately before and after the segment exit. However, in order
to perform a time interpolation, we would also need to know the location of the GPS points
involved. For example, in order to perform a time interpolation between points Pp1 and Pp in
Fig. 4, we would need to determine the location of these two points. Because each GPS point has
a positional error " associated with it, the interpolated time also has an associated error "t . This
error translates into an error "u in the computation of the segment speed u.
Let the interpolated times associated with the segment entrance and exit be tent and texit . The
average speed u associated with the segment of length L is (Fig. 4)
Fig. 3. Sample network map geocoding and segmentation of the I-10 and I-12 Split in Baton Rouge, LA.
L L
u 1
texit tent tL
where tL is segment travel time. Using error propagation theory (Gans, 1992) we can show that "u
can be expressed as (Quiroga, 1997)
p p
2" 2"
"u 2
texit tent tL
Eq. (2) shows that "u is a function of the positional accuracy of individual GPS points and the
segment travel time. It provides an upper bound for "u because the errors associated with tent and
texit are assumed to be independent. In practice, "u values may be lower if the errors associated with
tent and texit are both of the same sign and order of magnitude. However, because typical L values
are 0.2 miles or larger, we assume such errors to be independent for all practical purposes. It must
be noted that Eq. (2) is based on the assumption that a linear interpolation scheme is appropriate
for computing tent and texit . This assumption is reasonable in an environment in which GPS data is
collected every one second and distances between contiguous GPS points are much smaller than
segment lengths (Section 4 contains a more detailed discussion on this subject).
In our research we used GPS receivers which had the capability to record speeds in addition to
coordinates and time. For speed computations, the receivers poll pseudorange data (distance
from satellite to receiver) and pseudorange rate data for a fraction of a second and, as a result,
the computed speeds are almost instantaneous and essentially independent of position xes. This
Fig. 4. Timedistance diagram for GPS points on a segment.

allowed us to use GPS speeds to calculate distance driven which, in turn, allowed us to calculate
segment speeds. In Fig. 4, the total distance d covered by the probe vehicle between t0 and tp is
tp " p1 #
t t X tk1 tk1 t t
1 0 p p1
d vdt v0 k p 3
2 k1
2 2
t0
assuming a special case of the trapezoidal rule in which time intervals, o not remain constant (to
account for variations in the GPS sampling rate). As before, the trapezoidal approximation is
reasonable in an environment in which GPS data is collected every one second and distances
between contiguous GPS points are much smaller than segment lengths.
The average speed ud associated with d is
d d
ud 4
td tp t0
Normally, the segment length L is known. If the initial and nal GPS points associated with the
segment are close to the entrance and exit points, respectively, d and L should be very similar. In
this case, ud can be assumed to apply over the entire segment length and can be used to estimate
u. Eq. (4) then becomes
( " # )
1 t t X
p1 t tp tp1
1 0 k1 tk1
u 0 k p 5
tp t0 2 k1
2 2
The resulting travel time tL along the segment is then

L
tL 6
u
It is possible to derive an expression for "u when using Eq. (5) to estimate u. As before, using
error propagation theory (Gans, 1992), and assuming, for simplicity, that the sampling period or
time interval t between contiguous GPS points is constant, we can show that
p p
p 0:5 ttL 0:5t
"u " " 7
p tL
where " is the error in speed associated with individual GPS points and p is the number of GPS
points that can be associated with a segment (in addition to P0 ).
Eqs. (2) and (7) are comparable in the sense that they only measure errors in segment speed due
to errors associated with individual GPS data measurements (" and "v , respectively). They do not
measure errors due to other factors like GIS distance measuring accuracy or GPS sampling rates
(see Section 4.2 for a discussion on this subject). Because Eqs. (2) and (7) are comparable, we can
combine them to determine a threshold beyond which Eq. (5) would be preferable for computing
segment speeds. Such a threshold is given by
Table 2
Limiting error "v as a function of tL and ", assuming t 1 s [Eq. (8)]
tL (s) Limiting error "v in GPS speed measurements (mph)
" 0:1 m " 0:2 m " 0:5 m " 1m " 2m " 5m " 10 m " 20 m
5 0.15 0.30 0.75 1.49 2.98 7.46 14.92 29.83

10 0.10 0.21 0.51 1.03 2.05 5.13 10.27 20.53
20 0.07 0.14 0.36 0.72 1.43 3.58 7.17 14.33
30 0.06 0.12 0.29 0.58 1.17 2.91 5.83 11.65
60 0.04 0.08 0.21 0.41 0.82 2.05 4.10 8.20
120 0.03 0.06 0.14 0.29 0.58 1.45 2.89 5.79
240 0.02 0.04 0.10 0.20 0.41 1.02 2.04 4.09
p
2"
"v < p 8
ttL 0:5t
Table 2 shows a few sample values of "v for several combinations of " and tL values, assuming
t 1 s. Note that "v is directly proportional to ". This means that as " decreases, i.e. as the
positional accuracy of the GPS equipment increases, "v must also decrease in order for Eq. (5) to
maintain its relative advantage. Otherwise, Eq. (1) will become relatively more accurate than Eq.
(5). Note also that "v is inversely proportional to tL . This means that as the segment travel time
increases, the eect of positional errors for individual GPS points will be lower and, as a result,
Eq. (1) will tend to become relatively more accurate than Eq. (5). In our research, we used GPS
equipment (with dierential correction) which had a positional accuracy of 25 m spherical error
probability (SEP) and a speed accuracy of 0.1 mph (1 sigma) (Trimble, 1993). From Table 2,
assuming " 2 5 m, it is evident that the limiting error "v is much larger than 0.1 mph. This
means that for the GPS equipment we used, Eq. (5) turns out to be preferable for computing
segment speeds.
In order to apply Eq. (5), we need a procedure to dene which points can be associated with the
segment. A very simple approach involves using GIS neighborhood functions to search for the
points that are closest to the segment entrance and exit (P0 and pp in Fig. 4). Notice that P0 can
occur either before or after the segment entrance. Similarly, Pp can occur either before or after the
segment exit.
Eqs. (5) and (6) are valid for a single run on a single segment and provide the necessary tools to
transform a set of GPS point time stamp and speed values into a single pair of segment travel
time and average speed values for the segment. In general, however, several runs involving several
contiguous segments are typically made. In this case, it may be of interest to compute not only
speed values for individual segments due to individual runs, but also representative speed values
for each segment and for all segments combined. To keep the discussion general, the number of
runs per segment is assumed to be dierent. This is particularly important if there are inter-
changes or intersections along the route and some segments have more records than other seg-
ments. Because of this, the procedure to compute representative speed values involves
aggregating the data at the segment level rst.
Let the number of runs (or sample size) per segment be mi . The average travel time per segment is
1 X mi
1 X mi
Li 1 X mi
1
ti tLj Li 9
mi j1 mi j1 uji mi j1 uij
where Li is the length of each segment and uij is the jth speed record associated with segment i.
Let the number of segments be n. The total distance LT is
Xn
LT Li 10
i1
The total travel time tTL over LT is

" #
X n
1 X mi
1
tTL Li 11
i1
mi j1 uij
The average speed for all runs over LT is

LT 1
uL " # 12
tTL P
n P
mi
Li 1 1
LT mi uij
i1 j1
This equation represents a weighted harmonic mean, where the weight is the ratio of the length of
each segment to the total length considered. If all segments have the same length, the weight
associated with each segment will be the same.
One disadvantage of using Eq. (12) to estimate average segment speeds is that the harmonic
mean, by denition, is very sensitive to individual values which are much lower than the rest in
the series. As a result, outlying low speeds, which tend to occur on atypically adverse trac
conditions, could result in very small average speeds. Since the objective of using a formulation
like Eq. (12) is to obtain a solid, robust estimator of central tendency, the question then is how
appropriate the harmonic mean speed formulation is for estimating central tendency. After all,
the central tendency estimator does not necessarily have to be the harmonic mean speed.
One possible modication to the harmonic mean formulation would be to manually lter out
records that are ``atypical''. However, this does not appear to be feasible in an environment in
which large amounts of data are being collected. A second possible modication would be to
convert Eq. (9) so that median segment travel times are calculated instead of arithmetic mean
segment travel times. The median is known for not being seriously aected by others and in many
cases it is preferred by statisticians as a measurement of central tendency (Duncan, 1986). The
end result would be a formulation similar to Eq. (12), but that uses segment median speeds
instead of segment harmonic mean speeds. The median speed formulation would be
1
uL n h i 13
P Li 1
LT umi
i1
where umi is median speed associated with segment i during the time period being studied.
The equations described previously are general in the sense that they can be used either for one
or several runs, and either for one or several contiguous segments. To illustrate this point, con-
sider the GPS data from a run made on 13 August 1995 on I-10 in the vicinity of the I-10 and I-12
Split in Baton Rouge, LA [Fig. 3(e)]. Table 3 shows the GPS data and summarizes the compu-
tational procedure. For example, for segment 12444, using Eq. (3),

1 60:5 2:0 61:1 4:0 61:2 2:5 60:9 1:0
d ... 0:196 miles
3600 2 2 2 2
Using Eq. (5),

1 60:5 2:0 61:3 4:0 61:2 2:5 60:9 1:0
u ... 61:20 mph
11:5 2 2 2 2
Using Eq. (6),

0:200
tL 3600 11:76 s
61:20
Table 3 summarizes these values, as well as those corresponding to segments 12453 and 12454.
Now, using Eq. (10),
Table 3
Computation of travel time and speed for run of 13 August 1995
GPS travel time data Segment td (s) d (miles) u (mph) tL (s)
Time (AM) t (s) Speed (mph) Code L (miles)
6:26:51.3 60.5
6:26:53.3 2.0 61.1
6:26:55.3 2.0 61.3
6:26:56.8 1.5 61.4
6:26:57.8 1.0 61.4
6:26:58.8 1.0 61.4
6:27:59.8 1.0 61.3
6:27:00.3 0.5 61.4
6:27:01.8 1.5 61.2
6:27:02.8 1.0 60.9 12444 0.200 11.5 0.196 61.20 11.76
6:27:03.8 1.0 60.5
6:27:04.8 1.0 60.2
6:27:06.3 1.5 59.8
6:27:06.8 0.5 59.8
6:27:08.8 2.0 59.4 12453 0.106 6.0 0.100 60.03 6.36
6:27:10.3 1.5 59.3
6:27:11.3 1.0 59.5
6:27:12.3 1.0 59.2
6:27:14.8 2.5 57.2
6:27:15.8 1.0 56.7 12454 0.106 7.0 0.114 58.60 6.51
L 0:200 0:106 0:106 0:412 miles
Using Eq. (11),

0:200 0:106 0:106
tTL 3600 11:76 6:36 6:51 24:63 s
61:20 60:03 58:60
Finally, using Eq. (12),
0:412 3600
uL 60:22 mph
24:63
From Table 3, the td (or tp t0 ) values for the three segments considered are 11.5, 6.0, and 7.0 s.
These values are very similar to the three tL values obtained (11.76, 6.36, and 6.51 s). Further-
more, the summation of the three td values (24.5 s) is very close to the summation of the three tL
values (24.63 s). The close agreement of these values empirically indicates that the methodology to
compute segment travel times and speeds based on Eq. (5) is valid. To better substantiate this
assessment, Fig. 5 shows the relationship between tL and td for 155,300 segment travel time and
speed records collected in the Baton Rouge network between 1995 and 1996 for all trac condi-
tions and facility types. Fig. 5(a) shows a scatter plot of tL vs td and clearly demonstrate the lack
of bias of tL with respect to td . This is also shown in Fig. 5(b) with the symmetry of the dis-
tribution of dierences between tL and td .
Fig. 5. Relationship between tL and td for 155,300 segment travel time and speed records in Baton Rouge, LA.
3. Implementation
3.1. Data collection procedure
In general, runs are made during the AM peak and PM peak periods. For each time period, the
number of runs must comply with acceptable error tolerance specications. Many travel time
studies are carried out following guidelines included in the ITE Manual of Transportation Engi-
neering Studies (Robertson, 1994). Those guidelines dene minimum sample sizes as a function of
permitted errors in the estimate of the mean speed, average ranges in speed, and specied levels of
condence. Unfortunately, the ITE formulation seriously underestimates minimum sample sizes.
Details of this assessment and a new formulation we have proposed can be found in Quiroga and
Bullock (1998).
3.2. Data reduction procedure
To eciently transform GPS point travel time data into segment travel times and average
speeds, we developed a data reduction application with a convenient user interface within the
GIS. In a typical session, the application user loads a GPS data le and selects segments. Each
time the user selects a segment, the application nds the GPS points that are closest to the seg-
ment entrance and exit and evaluates a segment speed based on all GPS speeds associated with
the segment. Between June 1995 and July 1996, several undergraduate students used this data
reduction application to process travel time runs totaling nearly 30,000 miles on 329 miles of
urban highways in Baton Rouge, Shreveport, and New Orleans. They generated 183,000 segment
travel time records out of 2.9 million GPS data points collected.
We stored all segment travel time data in a relational database. The basic structure of this
database is composed of two tables: CORR_SEGMENTS and SEG_TRAVEL_TIME. Fig. 6
shows a description of the associated attributes and a sample of records that correspond to seg-
ments in Fig. 3(e). Table CORR_SEGMENTS contains basic segment data, including a unique
segment code, name, direction, and length. Table SEG_TRAVEL_TIME contains segment travel
time data, including segment code, date, time, vehicle ID, and segment speed.
3.3. Data reporting procedure
After the data is reduced to the format in Fig. 6, the next step is to report it. An obvious
approach, which is typical in traditional time travel studies, is to draw speeddistance or speed
time proles along the corridor of interest. However, this type of report can be greatly enhanced
by the use of GIS and database querying and reporting tools. This way, we can produce reports
on measures of eectiveness using travel time at various levels of resolution including system,
corridor/route, and local. In this paper, it would be clearly impractical to include a comprehen-
sive list of data querying and reporting possibilities. For the sake of brevity, we include only two
reporting examples: thematic maps and strip map tabular reports. Fig. 7 shows a sample gray-
scale shaded map showing average segment speeds from September 1995 to May 1996 between
4:30 and 5:30 pm (peak hour) in Baton Rouge. Obviously, we could modify the query used to
generate the map to include dierent time periods or show other speed values such as median
Fig. 6. Description of attributes and sample of records from the database.
Fig. 7. PM peak (4:305:30 pm) average speeds in Baton Rouge, LA, from September 1995 to May 1996.
speeds, minimum speeds, and free ow speeds. In practice the maps are color plots covering a
1010 miles2 area. They are extremely useful for explaining trac congestion in public meetings,
but are not particularly suitable for archiving and analysis.
For archiving and analysis, we implemented a second reporting format similar to that used to
document highway features like sign posts and culverts. We produced these reports on 1117
paper for 20 highway segments in each direction of travel (Fig. 8). The data tabulated above and
below the ``strip map'' are average speeds and average cumulative travel time for both directions
of travel during the morning peak, o peak, and afternoon peak periods. The shading next to the
average speed values is related to the ratio of observed speed to posted speed limit and makes the
identication of critical congestion areas easy. We automated the procedure to produce these
tabular reports using Microsoft Access report macros. In a typical session, the application asks
the database operator to select a range of dates, the report page(s) of interest, and the printing
date. The application automatically executes the necessary database queries, formats the report
pages, and sends the results to the printer.
4. Analysis
This section analyzes the GPS-GIS methodology described before from the following points of
view: segment lengths, sampling rates, and central tendency. The segment length analysis exam-
ines the eect of using dierent highway segment lengths. The sampling rate analysis addresses
Fig. 8. Example report page with travel time data from September 1995 to May 1996 in Baton Rouge, LA.
the eect of collecting GPS data at dierent sampling periods. The central tendency analysis
compares harmonic mean speeds and median speeds.
4.1. Segment lengths
Many segmentation schemes are possible in a highway network. We can dene a detailed net-
work composed of relatively short segments, say 0.1 miles in length, or a ``coarse'' network
composed of relatively long segments, say 2 miles in length. Shorter segments imply a larger
number of segments but, at the same time, segment speeds which are closer to the original GPS
point speeds. Conversely, longer segments imply a fewer number of segments but, at the same
time, segment speeds which may be farther from the original GPS point speeds.
One way to estimate required segment lengths would be to borrow concepts from sampling
theory. This theory, which originated in Nyquist's theorem (Oppenheim and Schafer, 1989), has
been extended to other areas like computer cartography where it is used to determine the rela-
tionship between detectibility of individual features and spatial resolution (Tobler, 1987). For
speeddistance proles, sampling theory would establish that for a disturbance (normally gener-
ated by a physical discontinuity) to be visible in the speed-distance prole, the required segment
length must be smaller than half the length of the disturbance.
In order to evaluate whether this holds in the case of speeddistance proles, we considered
seven segmentation schemes for each highway section referred to in Fig. 1 and applied the data
reduction procedure to process the GPS runs described in Table 1. We then compared the
resulting segment speeddistance proles with those of Fig. 1. In all cases, we conrmed the
validity of the sampling theory. As an illustration, consider the seven groups of segment speed
distance proles shown in Fig. 9. These groups of proles resulted from seven corresponding
segmentation schemes on the 4.15-mile section of I-10 and I-12 EB between Acadian Thruway
and Airline Hwy [Fig. 1(b)], as follows: 42 segments (or 0.099-mile segments), 21 segments (or
0.197-mile segments), 10 segments (0.415-mile segments), six segments (or 0.691-mile segments),
three segments (or 1.38-mile segments), two segments (or 2.07-mile segments), and one segment
(4.15 mile long). For visualization purposes, we constructed each prole by drawing straight lines
connecting points located half way along the segments. A comparison between the proles of
Fig. 9 and those of Fig. 1(b) clearly indicates that using segments 1.38 miles long resulted in an
almost complete loss of information compared to the original GPS data [Fig. 9(e), (f), and (g)].
When using 0.415-mile or 0.691-mile segments [Fig. 9(c) and (d)] the eects of the I-10 WB on-
ramp and the Jeerson Hwy on-ramp became visible. These observations were consistent with
sampling theory concepts because both 0.415 and 0.691 miles not only were smaller than 1.38
miles but also smaller than half the total length of the disturbances generated by the I-10 WB on-
ramp and the Jeerson Hwy on-ramp (around 2 and 1.5 miles, respectively). When using 0.099 to
0.197-mile segments [Fig. 9(a) and (b)] the eects of practically all physical discontinuities became
visible. These observations provide a clear indication that traditional travel time studies based on
long segments connecting contiguous physical discontinuities (or links), as recommended in the
literature (Robertson, 1994; TRB, 1994), are not sucient to characterize localized eects prop-
erly.
Speed distance proles like those included in Fig. 9 provide a rough, qualitative way of
determining segment lengths. In practice, unless we have a fairly good idea of the size of the
Fig. 9. PM peak speeds on the I-10 and I-12 corridor in Baton Rouge, LA, using various segment lengths for aggregation.
disturbances, we would need to consider fairly short segments for the data reduction process and
then aggregate the resulting segment speed data as needed, following the mathematical model
described in Section 2.2. In general, we would also need a quantitative measure of the improve-
ment in t for smaller segments. To do this, we analyzed the distribution of speed residuals, where
speed residuals represent the dierence between GPS speeds and the corresponding segment
speeds (Fig. 10). In most cases, regardless of segment length, the distributions of speed residuals
tended to be symmetric around zero. Initially, we thought these residuals would be normally
distributed. However, for relatively short segment lengths (<0.5 miles), the distributions of speed
residuals tended to be more peaked, i.e. they had a positive kurtosis (Duncan, 1986), than the
corresponding theoretical normal distribution (zero kurtosis). Using a Kolmogorov-Smirnov test
(Massey, 1951; Ostle and Mensing, 1975) we determined that these speed residual distributions
could be represented by double exponential distributions (Hahn and Meeker, 1991). The corre-
sponding cumulative probability distribution is of the form
Pr; expjrj= 100% 14
where r is speed residual, is standard deviation, and Pr; is probability that a speed residual
will be larger than r. For longer segment lengths, the distributions of residuals did atten con-
siderably (in several cases the kurtosis became negative) but, in general, they did not t a normal
distribution.
Fig. 10. Comparison between original GPS speeds and aggregated segment speeds (using 0.099-mile segments).
Fig. 11 shows plots of as a function of segment length. Each curve represents the relationship
between segment length and for each of the travel time runs described in Table 1. The eect of
segment length on is particularly signicant for segment lengths smaller than 0.5 miles. For
larger segment lengths, that eect tends to decrease until, eventually, it becomes negligible for
segment lengths larger than 1.5 miles. Obviously, there are numerical dierences between plots,
particularly between the AM and PM plots for the I-10 and I-12 section [Fig. 11(a) and (b)].
However, plots associated with congested conditions tend to exhibit a similar behavior, particu-
larly for segment lengths shorter than 1 miles. For example, the PM plots for the freeway sections
[Fig. 11(b) and (c)] suggest values of about 3 mph for 0.1-mile segments, and 5 mph for 0.2-mile
segments. The plots for the Airline Hwy runs [Fig. 11(d)] show larger values for the same seg-
ment lengths (6 and 9 mph, respectively). The Airline Hwy section is a signalized highway, as
opposed to I-10 and I-12 which are controlled access facilities. This is a clear indication that sig-
nalized highways need shorter segment lengths than controlled access facilities in order to achieve
similar speed residual levels.
We can use the knowledge that speed residuals tend to follow a double exponential distribution
to better characterize the relationship between GPS speeds and segment speeds. For example,
Fig. 11. Eect of segment length on standard deviations of speed residuals.

assume 5 mph (value obtained for 0.2-mile segments on the I-10 and I-12 section). Using Eq.
(14), the probabilities that a speed residual will be larger than 5, 10, and 15 mph are about 37, 14,
and 5%.
4.2. Sampling rates
Using a laptop computer to collect data directly from the GPS receiver makes one second
sampling periods easy to obtain. Frequently, however, GPS data is transmitted via radio to a
central location and much longer sampling periods are used. It is important then to discuss the
impact of these longer sampling periods on the number of GPS points that can be associated with
a segment. As before, borrowing concepts from sampling theory, the minimum sampling rate
would have to be such that the time interval between consecutive GPS points is at most half the
segment travel time. To determine the validity of this hypothesis, we generated six subsets from
the data associated with each of the GPS runs described in Table 1. The six subsets were the result
of assuming six dierent sampling rates as follows: 1, 0.5, 0.2, 0.1, 0.06667, and 0.05 Hz (or
sampling periods of 1, 2, 5, 10, 15, and 20 s, respectively). We then applied each subset to the
travel time runs described in Table 1. For simplicity, we considered only one segmentation
scheme (0.197-mile segments for I-10 and I-12 EB; 0.199-mile segments for I-12 WB and 0.194-
mile segments for Airline Highway SB). For each corridor and GPS data subset, we obtained the
corresponding number of GPS points per segment and segment speed. We then used all pairs of
number of GPS pointssegment speed values to generate the plots shown in Fig. 12. Each plot
represents a composite gure which summarizes the relationship between the number of GPS
points per segment, segment speed, and sampling period for all runs associated with a specic
corridor. For example, Fig. 12(b) contains all pairs of GPS pointssegment speed values obtained
for the 36 subsets obtained from the six runs (six subsets for each run) made on I-10 and I-12 EB
(PM peak). For completeness, Fig. 12 also shows the numerical percentage of segments that
ended up with GPS points associated with them, as well as a theoretical curve (dotted line) and a
best t curve (solid line) for each group of points. The theoretical curves shown in Fig. 12 resulted
from applying the following expression:
3600L
N 15
ut
where N is number of GPS points per segment; L is segment length in miles; u is segment speed in
mph; and t is nominal sampling period in s.
A comparison between the best t curves and the theoretical curves indicates that the actual
number of GPS points per segment was about 25% lower than what the theoretical equation
suggests. This means that the actual sampling period was about 30% longer than the nominal 1 s
sampling period. This larger value was due to occasional gaps in the GPS data (bridges, tree
canopy, and signal losses), perhaps to imperfections in the way the GPS receiver collected and
processed the data, and to the fact that the actual number of GPS points per segment N was an
integer.
For sampling periods up to 5 s, 100% of the segments ended up with GPS points. For sampling
periods larger than or equal to 10 s, the percentage of segments with GPS points started to
decrease as the sampling period increased. This trend was consistent with sampling theory con-
cepts. For example, for a sampling period of 10 s, a few segments were left without any GPS data
if the vehicle was traveling at around 55 mph. At this speed, the travel time was almost 13 s. From
sampling theory, the sampling period should be at most 6 s (assuming a sampling period of 1/2
the segment travel time). A sampling period larger than 6 s would almost inevitably result in some
Fig. 12. Number of GPS points per segment as a function of sampling period and vehicle speed.
segments not having any GPS data. The results obtained for a sampling period of 5 s vs those
obtained for a sampling period of 10 s conrm this assessment.
Because of their eect on the number of GPS points per segment, sampling rates also have an
eect on segment speeds. If the speeds of the GPS points associated with a segment are relatively
uniform, the eect of a varying sampling rate on the corresponding average segment speeds
should be minor. In general, though, variations in segment speed are expected to grow as the
sampling rate decreases (i.e. the sampling period increases) because of the uncertainty associated
with a lower number of GPS points per segment. To observe this eect, we computed and plotted
the relative variation in speed with respect to the 1-s segment speed for each value of speed as a
function of GPS sampling period (Fig. 13). For completeness, Fig. 13 also shows dotted lines
depicting average relative variations (almost horizontal lines), as well as 1-sigma and 2-sigma
bands. As the sampling period increases, these bands tend to spread. This is a clear indication of
the increasing uncertainty associated with the computed segment speed as the sampling period
increases.
Notice the dierence in behavior between plots according to trac condition and facility type.
Of particular interest is the larger spread observed for the Airline Highway plot [Fig. 13(d)]. This
Fig. 13. Relative variation of segment speed with respect to the 1 s segment speed as a function of sampling period.
is an indication that GPS sampling periods for signalized highways have to be shorter than GPS
sampling periods for freeways in order to achieve similar levels of certainty in the computation of
segment speeds. For example, assume a speed variation of 5% (2 sigma). From Fig. 13(b), the
sampling period for a freeway such as I-10 or I-12 would have to be around 4 s. However, from
Fig. 13(d), the sampling period for a signalized highway such as Airline Highway would have to
be only 2 s. In practice, this GPS sampling period would have to be even lower to account for the
fact that signalized highways require shorter segments than freeways (see Section 4.1).
4.3. Central tendency
As discussed in Section 2.2, the harmonic mean formulation in Eq. (12) is very sensitive to slow
speeds observed during atypical trac incidents. In order to test the eect of using Eq. (13)
instead of Eq. (12), we calculated both harmonic mean speed and median speed for nearly 1900
segments in the Baton Rouge network, using 26,000 segment records collected from September
1995 to May 1996 during the 7:008:00 am, 4:305:30 pm time periods. As an illustration, Fig. 14
shows the relationship between median speeds and harmonic mean speeds for the 4:305:30 pm
data. The corresponding gures for the 7:008:00 am data were very similar. Fig. 14(a) clearly
indicates that median speeds tend to be larger than harmonic mean speeds. This is also shown in
Fig. 14(b) as most dierences between median speed and harmonic mean speed turned out to be
positive. However, as shown in Fig. 14(b), the vast majority of dierences were smaller than
Fig. 14. Relationship between median speed and harmonic mean speed in Baton Rouge, LA (September 1995May
1996, 4:303:30 pm time period).
2 mph, clearly indicating that for most segments, median speeds can be used instead of harmonic
mean speeds and still produce essentially the same results. Fig. 14(a) shows this eect with the
extremely dense cluster of points around a hypothetical y x line.
In some cases, the dierences between median speeds and harmonic mean speeds were too
great to be ignored. Figure 14(a) shows this situation with the points that are farthest apart from
the linear trend. As an illustration, Table 4 shows the largest dierences observed for the 4:30
5:30 pm time period. For the most part, incidents like stalled vehicles, accidents, or extremely
heavy rain were responsible for the low harmonic mean speeds and, therefore, for the big dier-
ence between median and harmonic mean speeds. After removing speeds associated with inci-
dents, most harmonic mean speeds increased noticeably, while median speeds barely changed at
all. As a result, dierences between median speeds and harmonic mean speeds decreased. The
only segments that did not exhibit a signicant reduction in speed dierence were segments 13275
and 12271. Segment 13275 is located on Florida Blvd WB (inbound), just before the O'Neal Ln
signalized intersection. Florida Blvd WB is not usually congested during the PM peak. Only two
of the seven times the probe vehicle crossed O'Neal Ln while going west the stop light was on red
causing the vehicle to stop. Since WB trac is barely aected by the signalized intersection, we
can argue that a high median speed represents average trac conditions much better than a low
harmonic mean speed. Segment 12271 is located on I-10 EB halfway between Louisiana Highway
1 and Highland Rd (Fig. 7). Segment 12271 is located near the upstream limit of the congestion
caused by the I-10 and I-110 merge. As a result, trac on segment 12271 tends to be unstable and
unpredictable. However, dierences between median speeds and harmonic mean speeds for
neighboring segments tend to be much smaller. Consequently, we can still argue that the median
speed approach is still better for segment 12271 because it does not involve executing a separate
query to compute the harmonic mean speed for just one segment.
Table 4
Largest dierences between median speeds and harmonic mean speeds in Baton Rouge, LA (September 1995May
1996, 4:305:30 pm time period)
Segment Functional All records After removing records

class associated with incidents
No. of Median Harmonic Dierence No. of Median Harmonic Dierence

records speed mean speed (mph) records speed mean speed (mph)
(mph) (mph) (mph) (mph)
13275 Principal arterial 7 51.33 23.57 27.76 7 51.33 23.57 27.76

12300 Interstate 17 52.84 24.81 28.03 15 53.88 48.57 5.32
12270 Interstate 17 44.34 16.00 28.34 12 52.18 48.11 4.07
13564 Principal arterial 5 42.24 13.76 28.48 4 43.09 43.57 0.49
12269 Interstate 15 49.18 20.08 29.10 11 50.29 48.07 2.22
12198 Interstate 13 49.63 19.61 30.02 10 50.96 50.48 0.49
12199 Interstate 13 52.06 21.73 30.33 11 53.07 53.73 0.66
12271 Interstate 18 46.47 15.84 30.63 13 49.54 28.36 21.18
12268 Interstate 17 49.70 18.94 30.76 14 50.53 49.33 1.20
5. Conclusions and future work
In this paper, we described an integrated GPS-GIS methodology for conducting travel time
studies. It includes a spatial and mathematical model that can handle vast amounts of GPS data
in a GIS environment, a data collection procedure, a data reduction procedure, and a data
reporting procedure. It also includes procedures to analyze the signicance of the speed and travel
time data obtained. The spatial and mathematical model uses short highway segments, say 0.2
miles in length. However, it is suciently general so that other segment sizes can be easily
accommodated. The data collection procedure uses GPS receivers to automatically record time,
local coordinates, and speed of probe vehicles every 1 s. The data reduction procedure uses an
application with a convenient user interface within the GIS to eciently aggregate GPS time and
speed data into highway segment travel time and speed data. This data is then imported into a
geographic database where it can be used for the production of reports. The data reporting pro-
cedure uses graphical and tabular reports including color coded maps and archival tabular
reports and can be used to produce reports on measures of eectiveness using travel time at var-
ious levels of resolution including system, corridor, and local.
We included three analyses to assist readers in the process of understanding the GPS-GIS
methodology: segment lengths, sampling rates, and central tendency. The segment length analysis
examined the eect of using dierent segment lengths on the detectibility of changes in speed due
to physical discontinuities like signalized intersections, ramps, and interchanges. The analysis
shows that relatively short segments, say 0.2 miles long, are needed to characterize the eect of
most physical discontinuities, and provides a clear indication that link-based travel time studies
are not sucient to characterize localized eects properly (a link is a segment connecting physical
discontinuities). This is particularly noteworthy because typical procedures available in the lit-
erature are oriented toward the use of long segment sizes that are 12 miles in length. An analysis
of the distributions of dierences between GPS speeds and segment speeds (or speed residuals)
indicated a tendency for short length segmentation schemes to follow a double exponential dis-
tribution. This knowledge can be used to better characterize the relationship between original
original GPS speeds and segment speeds, and to provide a probabilistic framework to the deni-
tion of segment lengths.
The sampling rate analysis examined the eect of using dierent time intervals between con-
secutive GPS data points on segment GPS data coverage and segment speed variability. In
agreement with sampling theory concepts, the analysis shows that the sampling period or time
interval between consecutive GPS points has to be less than half the shortest segment travel time
in order to achieve a 100% segment coverage. The analysis also shows a tradeo between sam-
pling rates and segment speed reliability, and emphasizes the need for even shorter GPS sampling
periods (12 s) in order to minimize errors in the computation of segment speeds. This informa-
tion is important if an agency is considering collecting GPS data over radio links that cannot
support 1 s polling intervals.
The central tendency analysis compared harmonic mean speeds and median speeds as estima-
tors of central tendency of segment speed. The analysis shows that in most cases, median speeds
and harmonic mean speeds are very similar. However, for those cases where the two values dier
considerably, median speeds represent typical trac ow conditions better than harmonic mean
speeds. Based upon this observation, we recommend that Eq. (13) be used instead of Eq. (12).
There are areas in the work described in this paper which warrant further research. One of
them is the use of xed highway segments. By denition, trac ow is dynamic both in space and
time. This means that depending on the circumstances, it may be more appropriate to use either
shorter of longer segments than those dened by the hard segmentation process. This limitation
can be overcome by the use of dynamic segmentation techniques. It would be desirable to modify
the GPS/GIS methodology so that each GPS point is assigned a milepost or cumulative linear
distance tag along the route of interest before the GPS data is actually aggregated into highway
segments.
In this paper, we developed a model using error propagation theory to compare errors in GPS
position xes with errors in GPS speeds. Such a model led us to the conclusion that using GPS
speeds was preferable for computing segment speeds. The work we have done with nearly 3 mil-
lion GPS data records over the past 3 years has provided us with strong evidence for supporting
such an assessment. However, we believe a more conclusive answer would be obtained by setting
up an experimental procedure by driving a probe vehicle equipped with a GPS receiver on a
carefully measured test track under laboratory-controlled conditions.
Some of the results described in this paper, particularly those that involve probability dis-
tributions, are based on data we collected in three metropolitan areas in Louisiana. To generalize
these results, it would be necessary to conduct similar analyses using data from other metropoli-
tan areas.
Acknowledgements
The procedures described were developed as part of the Congestion Management System project
for Baton Rouge, LA. This project was supported by the Louisiana Transportation Research
Center (Awards No. 95-7SS and 97-7SS). The views expressed do not necessarily reect the poli-
cies of the Louisiana Department of Transportation and Development or the Louisiana Trans-
portation Research Center.
References
Benz, R.J., Ogden, M.A., 1996. Development and benets of computer-aided travel time data collection. Transporta-
tion Research Record 1551, Transportation Research Board, Washington DC, pp. 17.
Duncan, A.J., 1986. Quality Control and Industrial Statistics, 5th ed. Irwin, Homewood, IL.
Gans, P., 1992. Data Fitting in the Chemical Sciences. John Wiley & Sons, Chichester, UK.
Guo, P., Poling, A.D., 1995. Geographic information systems/global positioning systems design for network travel
time study. Transportation Research Record 1497, Transportation Research Board, Washington, DC, pp. 135139.
Hahn, G.J., Meeker, W.Q., 1991. Statistical Intervals. A Guide for Practitioners. John Wiley & Sons, New York.
Harding, J., Salwin, A., Shank, D., Ghaman, R., 1996. GPS applications for trac engineering studies. Preprint
960214, Transportation Research Board, 75th Annual Meeting, Washington, DC, 711 January 1996.
Liu, T.K., Haines, M., 1996. Travel time data collection eld testslessons learned. Report FHWA PL-96-010,
FHWA, US Department of Transportation, Cambridge, Mass.
Massey, F.J., 1951. The KolmogorovSmirnov test for goodness of t. Journal of the American Statistical Association
46(253), 6878.
May, A.D., 1990. Trac Flow Fundamentals. Prentice Hall, Englewood, NJ.
Oppenheim, A.V., Schafer, R.W., 1989. Discrete-Time Signal Processing. Prentice Hall, Englewood Clis, NJ.
Ostle, B., Mensing, R.W., 1975. Statistics in Research. Basic Concepts and Techniques for Research Workers, 3rd ed.
Iowa State University Press, Aimes, IA.
Quiroga, C.A., 1997. An integrated GPS-GIS methodology for performing travel time studies. Ph.D. dissertation,
Louisiana State University, Baton Rouge, LA.
Quiroga, C.A., Bullock, D., 1996. Congestion management system architecture for controlled access facilities. Trans-
portation Research Record 1551, Transportation Research Board, Washington DC, pp. 105113.
Quiroga, C.A., Bullock, D., 1998. Determination of sample sizes for travel time studies. ITE Journal, on the web,
August 1998, http://www.ite-org.
Robertson, H.D., 1994. Travel time and delay studies. In: Robertson, H. D. (Ed.), Manual of Transportation Engi-
neering Studies, 4th ed. Institute of Transportation Engineers, Washington DC, pp. 5268.
Tobler, W., 1987. Measuring spatial resolution. International Workshop on Geographic Information Systems, Beijing,
China, pp. 4248.
TRB, 1994. Highway Capacity Manual, 3rd ed. update. Special Report 209, National Research Council, Washington,
DC.
Trimble Navigation Limited, 1993. Placer GPS 400 Installation and Operator's Manual. Sunnyvale, CA.
Turner, S.M., 1996. Advanced techniques for travel time data collection. Transportation Research Record 1551,
Transportation Research Board, Washington DC, pp. 5158.
Zito, R., D'Este, G., Taylor, M.A.P, 1995. Global positioning systems in the time domain: how useful a tool for
intelligent vehiclehighway systems? Transportation Research C 3(4), 193209.

Travel Time Studies With Global Positioning and Geographic

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

Travel Time Studies With Global Positioning and Geographic

Uploaded by

Copyright:

Available Formats

TRANSPORTATION

Transportation Research Part C 6 (1998) 101127

Travel time studies with global positioning and geographic

Received 24 March 1997; received in revised form 18 June 1998

* Corresponding author. Tel.: +1-210-731-9938; Fax: +1-210-731-8904; e-mail: c.quiroga@ttimail.tamu.edu

0968-090X/98/$19.00 # 1998 Elsevier Science Ltd. All rights reserved.

2. Aggregating speed and travel time

2.1. Spatial model

2.2. Mathematical model

Fig. 4. Timedistance diagram for GPS points on a segment.

The resulting travel time tL along the segment is then

tL (s) Limiting error "v in GPS speed measurements (mph)

5 0.15 0.30 0.75 1.49 2.98 7.46 14.92 29.83

The total travel time tTL over LT is

The average speed for all runs over LT is

GPS travel time data Segment td (s) d (miles) u (mph) tL (s)

Time (AM) t (s) Speed (mph) Code L (miles)

L 0:200 0:106 0:106 0:412 miles

Using Eq. (11),

Finally, using Eq. (12),

3.1. Data collection procedure

3.2. Data reduction procedure

3.3. Data reporting procedure

Fig. 6. Description of attributes and sample of records from the database.

4.1. Segment lengths

Pr;  expjrj= 100% 14

Fig. 11. Eect of segment length on standard deviations of speed residuals.

4.2. Sampling rates

4.3. Central tendency

Segment Functional All records After removing records

No. of Median Harmonic Dierence No. of Median Harmonic Dierence

13275 Principal arterial 7 51.33 23.57 27.76 7 51.33 23.57 27.76

5. Conclusions and future work

You might also like

Pr; expjrj= 100% 14