You are on page 1of 15

JOURNAL OF INFORMATION SCIENCE AND ENGINEERING 26, 769-783 (2010)

A Distributed Threshold Algorithm for Vehicle Classification Based on Binary Proximity Sensors and Intelligent Neuron Classifier*
WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN
Department of Computer Science and Engineering Dalian University of Technology Dalian, Liaoning, 116023 P.R. China E-mail: gztan@dlut.edu.cn
To improve the accuracy of real time vehicle surveillance, utilize the advances in wireless sensor networks to develop a magnetic signature and length estimation based vehicle classification methodology with binary proximity magnetic sensor networks and intelligent neuron classifier. In this algorithm, we use the low cost and high sensitive magnetic sensors to measure the magnetic field distortion when vehicle crosses the sensors and detect vehicle via an adaptive threshold. The vehicle length is estimated with the geometrical characteristics of the proximity sensor networks, and finally identifies vehicle type from an intelligent neural network classifier. Simulation and on-road experiment obtains high recognition rate over 90%. It verified that this algorithm enhances the vehicle surveillance with high accuracy and solid robustness. Keywords: real-time traffic surveillance, vehicle detection, vehicle classification, wireless sensor networks, binary proximity sensor networks, intelligent neurons, distributed threshold, adaptive, clustering

1. INTRODUCTION
Nowadays, the urban traffic became a big problem with the rapid increase of vehicle quantity, and it disturbs the normal life of urban residents and travelers. Especially the traffic jams is a difficult problem confront the global with great financial loss every year. Intelligent traffic control system is proved the most effective approach to resolve this problem. Vehicle surveillance, including detection and classification, that provides realtime traffic data for traffic light control system with the needs to optimize the spatial and temporal allocation of traffic resource. And consequently, the performance of vehicle surveillance is significant to traffic light control, optimal traffic resource allocation and maintenance of the pavement system [1]. Currently there many vehicle surveillance technologies including loop sensor, video camera, image sensor, infrared sensor, microwave radar and GPS, etc. [2, 3]. The performance is acceptable but not sufficient because of their limited coverage and expensive costs of implementation and maintenance. They have defects include line-of-sight, low exactness, depending on environment and weather, can not perform no-stop work whether daytime or night, high costs for install and maintenance, etc. Consequently, in actual application the traffic data is insufficient or bad in real-timeness owing to detector
Received March 31, 2009, revised August 28, 2009, accepted September 30, 2009. Communicated by Chih-Yung Chang, Chien-Chung Shen, Xuemin (Sherman) Shen, and Yu-Chee Tseng. * This paper was supported in part by the National Natural Science Foundation of China (No.60873256) and National Basic Research Program of China (No.2005CB321904).

769

770

WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN

quantity and cost. And thus the actual performance of traffic control system such as SCOOT/SCATS is influenced. With the increase of vehicle in urban road networks, the vehicle detection technologies are confronted with new requirements. Wireless sensor network is the-state-of-art technology and a revolution in remote information sensing and collection applications [4]. Sensor node has advantages such as low costs, small size, wireless communication, high sensing accuracy, and can be deployed with great quantity. It has broad prospect of application in intelligent transportation system [5]. In the PATH (California Partners for Advanced Transit and Highways) project of University of California, Berkeley, the possibility of replacing traditional methods, such as loop detector, with wireless sensor networks is creatively researched. Their ATDA (Adaptive Threshold Detection Algorithm) is an efficient vehicle detection algorithm with high precision of 97%, but the classification scheme is not so efficient with low performance that overall recognition rate is below 60% [6, 7]. The length is an important distinguishing factor for vehicle classification [7, 8]. The main challenge is that occupancy will be influenced by velocity, and consequently the length of vehicle cannot be exactly estimated via single detector. Actually, locating and tracking for the certain part of vehicle are indispensable when estimate the length of vehicle. Tracking and locating are hot topics in research of wireless sensor networks, and in recent the approach of tracking with the geometric topology is introduced [9, 10]. Among them the BPSN (Binary Proximity Sensor Networks) is a simple and efficient method, and extremely suitable for uncomplicated topology scenario such as traffic information detection. It draws attention owing to its good performance [11-14]. Under the background, a new algorithm, Magnetic Sensors based Vehicle Classification Algorithm (MSVCA) is developed in this paper. In this algorithm, magnetic sensors are deployed as BPSN to detect the magnetic field distortion with a distributed threshold, and estimate the length of vehicle via the geometric characteristics of the topology. Finally the important features are exacted to identify vehicle type with neural network classifier. The on-road experiment and simulation show that this algorithm enhances vehicle classification with good performance and solid robustness.

2. RELATED WORK
2.1 Binary Proximity Sensor Networks The binary proximity sensor network is a special sensor network. Every sensor node has definite coordinate and finite detection range R, and sends single bit information about the target, detected or not, to the access point in a fixed timeslot. The master node for computation locates and tracks target with the bits information, its own coordinate and the geometric characteristics of the wireless sensor network [12]. It shows as Fig. 1. Assume a BPSN with m sensor nodes, which detect target periodically in a certain interval , and a m-dimensional binary vector set S as Eq. (1) is obtained. And the + 1 means target moving towards the detector and 1 means leaving away from the detection range. 0 means null state that no target is detected. The moving trace can be calculated according to this vector set and timestamp.

DISTRIBUTED THRESHOLD ALGORITHM FOR VECHICLE CLASSIFICATION

771

X t1

t2

t3 t4 t5

Z Y X

Z t1 t2 t3 t4 t5 t6 t6 Fig. 1. Locating and tracking with proximity sensor networks.

s(ti) (+ 1, 1)m 2.2 Vehicle Magnetic Signature and ATDA

(1)

In PATH project, they creatively use networked high precise magnetic sensors to detect magnetic field distortion caused by moving vehicle, and introduce a traffic surveillance approach based on the electromagnetism that ferrous materials, such as vehicles, distort the Earths field which is uniform over a wide area on the scale of kilometers [6, 7]. The Magnetic sensor such as Honeywell HMC1002 two-axis detector can measure the magnetic field change of Earth with high accuracy. The magnetic field distortion caused by moving vehicle is the fundamental of ATDA [7]. The arrival of the magnetic frontier and trail of the vehicle will influence the background magnetic field. As a typical distortion signal, namely the magnetic signature, raw data r(k) (X-axis, MICAz node) and the detection sequence s(k) generated by ATDA are showed as Fig. 2.

scope

r(k) s(k)

t/sec

tup

tdown

Fig. 2. Magnetic field distortion and ATDA detection result.

The raw data r(k) is smoothed to a(k) and input to ATDA for automatic detection and finally generate a detection result according to the judgment of an adaptive threshold h(k). Here k is time interval. The output includes an impulse sequence s(k), the corresponding timestamp tup(k) and tdown(k), and finally the detection flag d(k) is generated

772

WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN

according to the output state of the state machine. At a finer scale, the change in detection flag occurs within 0.1s immediately after the magnetic frontier of the vehicle crosses the sensor. Different vehicle has different ferrous structure that results differentiated magnetic signatures with distinctive timing and amplitude characters. The magnetic signature can be used in vehicle classification. About the classification algorithm of ATDA, there three main drawbacks as follows: (i) the feature of vehicle signals took in classification is insufficient. Only the macroscopic feature (hill pattern) is considered but the timing character and influence from velocity, viz. the duty cycle, are neglected; (ii) the length of vehicle is beyond consideration; (iii) identify vehicle type directly via the amount of signal crest and hollow, thus the capability of fault tolerance is not good when consider noise.

3. NETWORK ENVIRONMENT AND PROBLEM STATEMENT


According to the application for on-road traffic surveillance to detect and classify vehicles traveling at the upstream of the intersections, we employ magnetic sensors and design a binary proximity networks to detect vehicle and estimate length based on sensor readings of magnetic field distortion signal and an adaptive threshold. The network topology as Fig. 3 is designed according to the traffic application scenario and the characteristics of binary proximity sensor networks.
d ij
R S0 S1 Sn-1 The lane A vehicle L Fig. 3. The deployment of sensor nodes. V AP

Assume to deploy n sensor nodes in the same straight line parallel with the lane that constitutes a binary proximity network S = {S0, S1, , Sn-1}. The offset distance from the lane is Doffset and the sensing range of every sensor node is R. And the distance between node Si and proximity node Sj is dij. The AP node, which is the head of data acquisition cluster, has more resources and capability for computing, and thus bears the synchronization, computation, communication and topology maintenance. The sensor node reads magnetic field distortion to detect vehicle based on ATDA and then reports the result to AP within single hop.
+ 1 when s (t j ) = 1 & s (t j ) = 0 sq (t j ) = 1 when s (t j ) = 0 & s (t j ) = 1

(2)

Assume a vehicle runs along the parallel trace with the lane, and AP collects data in

DISTRIBUTED THRESHOLD ALGORITHM FOR VECHICLE CLASSIFICATION

773

a certain interval that generates the binary detection sequence sq(tj) as Eq. (2). The sequence s(k) is vehicle detection status that is generated in ATDA. The locating precision is inversely proportional to the distances among sensor nodes [11]. To reduce node amount, the distances between nodes are belonged to normal distribution N(, ) according to vehicle length intervals. And in ATDA, the sensing range can be adjusted via the base value of the adaptive threshold h(k).

4. METHODOLOGY
4.1 Distributed Threshold for Vehicle Detection To the magnetic field of the Earth, there is an uncontrollable drift because of affect from environmental factors such as temperature, and the rate of the drift is on the order 1 measuring unit per minute. In order to account for the drift in the long term, in ATDA, an adaptive threshold h(k) is setup to track the background magnetic reading, which is used to determine the adaptive threshold level for the detection state machine, as Eq. (3). Here, is the forgetting factor, a(k) is the smoothed magnetic signature from raw signal r(k), and s(k) is the detection sequence.
h ( k 1) (1 ) + a ( k ) h(k ) = h ( k 1) if s ( ) = 0 otherwise

(3)

In MSVCA, the adaptive threshold is shared by every node for reducing energy consumption. The master node computes the adaptive threshold in idle state and distributes to slave nodes when vehicle is detected. All slave nodes are in sleeping state when there no vehicle, and waked up by master node, then synchronize their timers and receive the latest threshold as baseline and threshold to detect vehicles.
Init Init Init_done Count0

a (k ) p T (k )

T W

Ns

a (k ) f T (k )
Ms
Count01

t down

a (k ) p T (k )
Car

Count10

a (k ) f T (k )

a (k ) p T (k )
Count1

p
Ns

a (k ) f T (k ) tup

p Ms

a (k ) f T (k )

a (k ) f T (k )

Fig. 4. State transition in state machine for vehicle detection.

The state transition in vehicle detection is showed in Fig. 4. The input parameters include vehicle signal a(k), threshold h(k) and timestamp, and the output includes current state, state sequence s(k), detection flag d(k) and time tup and tdown. Here Ms and Ns are experiential threshold to reduce effect from signal fluctuation.

774

WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN

Time window W is used to avoid the cycle counting and vehicle missing. Meanwhile, aiming at the abrupt fluctuation in rise and decline edge of the signal, add two temporary states Count01 and Count10 and the corresponding counting threshold to enhance the robustness. In Fig. 4, there two conditions, current counting and the comparison of a(k) and h(k), for every state transition. In the detected state (Car), s(k) = 1, and d(k) = 1 when there continuous s(k) = 1 detected.
4.2 Vehicle Length Estimation

For describing the current location M of target at the overlapped area, define the event equation I(t, t) of node i, that denotes event I happened at time t on node i.

I (i, t ) =

i, t(i +1), enter t t(i ), exit 0, otherwise

(4)

When vehicles move in detection area, the time that target enters detectable range tenter = ti,up and the time target exits texit = ti,down, as showing in Fig. 5. The event is defined as Eq. (4).
d(k) s(k) sq(t)
t>W

Init Init_done Count0

Output

s(k)=0

s(k)=1

S i Ri
Di,i+1

Si +1
Doffset

lane
Computing

< Ns

d(k)=1

i n
Confirmed

s(k)=1 > Nm

Count1

head ( p, t p ), tail (q, t q )

Fig. 5. Event in length estimation.

Fig. 6. State machine for length estimation.

Simply, define the event as the leaving from the detectable range of node i. If node p and node q have detected the events of vehicle entering and exiting the detection range respectively, thus the length of vehicle can be estimated from Eq. (5) as follows. The parameters Rp and Rq denote the detect radius of node p and q respectively, and di,i+1 denotes the distance between the sensor node i and its proximity node i + 1. The factor Doffset denotes the offset distance from the pivot of sensor detect range to the lane.
2 2 2 L = (1 + ) ( di , i +1 R 2 Doffset Rq Doffset ) p i= p q

(5)

Actually there difference between the magnetic length and the physic length be-

DISTRIBUTED THRESHOLD ALGORITHM FOR VECHICLE CLASSIFICATION

775

cause of the spread of magnetic line. But on other hand the exact length of vehicle is not so meaningful, and it just needs an attribute of length that can differentiate vehicles. Assume the estimated length is LE, which belongs to a certain length interval separated by the n sensor nodes, and decided by their deployment and target location.
2 LE {Li | i [0, N ]}, N = Cn

Li = D p , q = d j , j +1 , i [0, N ], p, q [0, n]
j= p

(6)

Algorithm 1: LENGTHESTIMATION 1: Init S; Init(interval t); 2: {h(k), d(k), s(k), aq(t)} buffers; 0 T; Count0 S; 0 C(1); 3: while T < W do 4: if a (k ) > h(k ) then 5: 1 s (k ) ; Count1 S; 6: C(1) + 1 C(1); 7: COUNT(1); 8: if C(1) < Ns then 9: Count0 S; 10: 0 C(1); 11: else if C(1) > Nm then 12: Confirmed S; 13: 14: 15: I(i, t) i n ; if d(k) = 1 then Output S;

16: TRACK (in , aq(t)) LE; LE L; 17: Init S; 18: end if 19: end if 20: end if 21: end while

In Eq. (5) there is a fuzzy factor utilized to reduce the difference between the magnetic length and physical length. Based on above, design a state machine as in Fig. 6 to detect events and then calculate the length. The processing of state machine is as Algorithm 1. The vehicles length can be estimated via Eq. (5). And consequently, the average velocity in timeslots toccupy can be calculated according to the occupy time as well, based on vehicles length and occupy time on single sensor node i [7]. The estimated velocity is given by the following equations.
D v= i =
2 L + 2 Ri2 Doffset

t occupy

(t exit t enter )

(7)

776

WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN

5. VEHICLE CLASSIFICATION
The intelligent neural networks is a computational model based on biological neural networks with weighted training to non-linear differentiable function, which is widely used in pattern recognition and classification with good performance. There are broad applications in intelligent transportation as well [15-17]. In MSVCA, time-domain features are extracted from vehicle magnetic signature with consideration of duty cycle and fault tolerance. After analyzing the characteristics of vehicle magnetic signature, the enhanced feature vector set (FVS) is extracted from signals for clustering and vehicle classification. The features vector set makes up vehicle length and time-domain features. In the meantime the two important factors, length and velocity are taken into account, and the time feature of signals, namely duty cycle, is sufficiently considered. In addition, intelligent neuron based classifier is used for vehicle type recognition to quicken error reduction and enhance the fault tolerance capability.
5.1 Preprocessing and Feature Extraction

According to the principle of magnetic field distortion caused by moving vehicle [7] and abovementioned analysis to vehicle signature detected via sensor nodes, there are three conclusions as follows, (i) the amplitude variety is directly related with the ferrous materials distribution of vehicle and offset to sensor node; (ii) the time-domain width of magnetic field distortion signal is decided by length and velocity of moving vehicle; (iii) the sensor is high sensitive so the signal is variable in amplitude and easy to be influenced. To enhance the precision and overall performance of vehicle classification, all aspects need to be taken into account. To extract the time-domain features as Fig. 7 (a), sample data from the smoothed signal a(k) inside window W with frequency f (in times of reading slot), and extract the feature sequence Hf(n) with a running average to enhance fault tolerance, as showed in Fig. 7 (b). Assume the reading frequency of sensor is and the running average width is 2, thus there relation between a(k) and Hf(n) as Eq. (8).

H f (i) =

1 2

m+ j = m

a( j ) here m = Wi/f and

i [0, W/f]

(8)

Normalized Amplitude

t/sec

0 2 (b) Running average.

t/sec

H f (n) H f (1) H f (i ) (a) Time-domain features extraction.

Fig. 7. Features extraction from vehicle magnetic signature.

DISTRIBUTED THRESHOLD ALGORITHM FOR VECHICLE CLASSIFICATION

777

Thus with different frequency f, the H sequence can denote the variety in amplitude and time-domain width in different extent. Above all, the Features Vector Set (FVS) of vehicle is defined as follows. V = (Hf(n), L)
5.2 The Clustering Algorithm

(9)

Assume the class set of vehicle is C, and the FVS data set is X.
C M = {ci , i [1, M ]}
X N = {xi , i [1, N ]}

(10)

Use the improved fuzzy c-mean (FCM) clustering algorithm [18, 19], the clustering lost function defined via membership function is denoted as Eq. (11).

J f = [ j ( xi )]b || xi m j ||2
j =1 i =1

(11)

The factor mj is clustering center of class cj, and b > 1 is a constant, which can adjust the fuzzy extent of clustering. j(xi) is the membership function of the ith sample to the jth category, and it under the loose normalized restriction of Eq. (12).

j ( xi ) = n
j =1 i =1

(12)

Evaluate the minimum of Eq. (11) under the condition restriction of Eq. (12) with iteration method, and obtain the membership function.

j ( xi ) =

n(1 / || xi m j ||)1 /(b 1)

(1 / || xl xk || )
k =1 l =1

1 /( b 1)

(13)

And finally the sample is classified to a certain category according to the membership function of FVS. The membership value after clustering will be used as expected output for samples of this class in neural network training, which is described in next section.
5.3 Intelligent Neuron Classifier

The main idea of neural network is modify the weights according to the deviation between the real output of neural network and the target vectors, and minimize the summary of square deviation in output layer. According to foregoing research and experiential applications [20], a three-layered neural network with infinite hidden layer nodes can make arbitrary non-linear mapping from input to output, and thus MSVCA uses threelayered neural networks for vehicle type recognition. The general architecture of intelli-

778

WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN

gent neuron classifier used in MSVCA is as in Fig. 8 (a). To enhance the capability of information storage on neurons and classification performance, and accelerate error reduction speed, the intelligent neurons are used in the classifier. The model of intelligent neuron is showed in Fig. 8 (b). In the model, x is input, ai are connection weights, f(x) is sigmoid function with adjustable parameter m which will be modified during study, and y is output.
y = f ( ai x i ) and here f ( x ) =
i =1 n

1 1+ ex m

(14)

The input includes FVS of magnetic signature and vehicle length, and the output is the membership values belonging to every classes. During training, to the samples of a certain class, the expected output is the corresponding clustering result of membership, and thus neural network can calculate the membership function after training. A vehicle can be classified into one predefined vehicle class in FHWA (Federal Highway Administration, U.S.A.) scheme based on the max membership value from neural network classifier.
Common Neuron Intelligent Neuron

J f (ci , xi ) H f (0) H f (1) H f (2)


x n 1

x
x

a2 a1 a3

an 1

f (x)
an

xn

Length (a) Architecture of intelligent neuron classifier. (b) Model of intelligent neuron. Fig. 8. Architecture of intelligent neuron classifier.
s (k )
Length
r (k )

a (k )

ATDA

L
v
Classifier

t (i )
Velocity
an (k )

J f ( ci )

Band Filter

ci
Features
H f (n)

Normalized

Fig. 9. Block diagram of MSVCA algorithm.

5.4 The Algorithm

The overall block diagram of MSVCA is as in Fig. 9. It detects vehicle magnetic signature signals based on ATDA and then extracts the time-domain features from magnetic signature of travelling vehicle. Synchronously, it estimates vehicle length via prox-

DISTRIBUTED THRESHOLD ALGORITHM FOR VECHICLE CLASSIFICATION

779

imity sensor networks and finally identifies vehicle and classifies to a certain class. For sake of the limited computation power on sensor node and reduce the complexity of pattern recognition in high-dimensional spaces, the PCA (Principal Component Analysis) can be introduced to lower the dimension of FVS [7, 21].

6. EXPERIMENTAL RESULTS AND ANALYSIS


In the on-road experiment, six MICAz nodes with Honeywell HMC1002 two-axis linear magnetic sensors are used. The physic characteristics of sensor are listed in Table 1 and the snapshot of experiment scene is as in Fig. 10.
Table 1. Characteristics of sensor node. Characteristics Data Nominal Sensitivity 3.2 mV/V/Gauss Resolution 40 Gauss Supply Current < 20 mA Operating Temperature 40 ~ 85 Celsius Noise Density 29 nV/Hz Bandwidth 5 MHz
20 20

-20 0 40 20

10

-20 0 50

10

0 0 -20 0 5 10 -50 0 5 10

Fig. 10. Experiment scene and snapshot.

Fig. 11. Reading samples (X-axis).

Five vehicle types and magnetic signatures on X-axis are obtained as databank with sensor nodes, and the reading samples are showed in Fig. 11. And meanwhile, vehicle types are recorded manually by eyeballing according to 13-classes FHWA vehicle classification scheme. The parameters used in experiment are listed as follows. The reading frequency = 128Hz, R = 3.2m, W = 64 timeslots, f = 8, = 50ms, = 0.15, Doffset = 1.8m, and the distances between sensor nodes dij~N(, ). In the simulation, about 500 samples from the databank are used. The performance of MSVCA is analyzed based on this dataset. 140 samples are utilized to clustering and train neural networks and use the rest to verify the performance of MSVCA. Fig. 12. shows the clustering process in multiple iterations that both time-domain features of magnetic signature and vehicle length are took into account. In Fig. 13, the error reduction curve (for instance of class bus) verifies the satisfactory result achieved via intelligent neural network classifier.

780

WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN

0.45 0.4 0.35 0.3


error

0.25 0.2 0.15 0.1 0.05 0 0 200 400 600 800 1000

iteration times

Fig. 12. The processing of clustering.

Fig. 13. The error reduction curve (Bus).

The statistics of recognition rate (abbr. RR) is as Table 2. Its efficient to classify vehicles with high precision. The experimental results show that the improvement both in recognition rate and fault tolerance capability. The recognition performance of several latest classification algorithms with different technologies is compared as in Table 3. Its obvious that MSVCA improves the performance of vehicle classification with wireless sensor networks and enhances the utility of passive magnetic sensor in traffic surveillance.
Table 2. Vehicle recognition rate. Actual type (record manually) C1 C2 C3 C4 107 2 3 3 122 1 5 1 46 4 3 1 34 116 92.24 128 95.31 50 92.00 38 89.47

MSVCA C1. (Bus) C2. (Car) C3. (Truck) C4. (Van) C5. (Motorcycle) Total RR (%)

C5

Total 115 123 52 42 28 360 93.61

28 28 100

Table 3. Comparison of vehicle classification algorithms. Classification algorithm RR (%) Remarks PATH [6][7] 60 WSN based, magnetic sensor MSVCA 93.61 WSN based, magnetic sensor, intelligent neuron ILD [16] 91.5 Loop sensor, BP neural network classifier Anshuls method [17] 86 Camera, BP neural network classifier NN&SVM based [22] 94.8 Image sensor, BP neural network classifier MW Sensor [23] 87 Microwave radar Partial Gabor filter bank [24] 95.17 Camera Repetitive pattern [25] 73.5 Satellite image T2 FLRBC [26] > 80 Acoustic sensor, type-2 fuzzy logic classifier

DISTRIBUTED THRESHOLD ALGORITHM FOR VECHICLE CLASSIFICATION

781

7. CONCLUSION
Wireless sensor network is a revolution in applications of information sensing and collection, and consequently it has broad prospect in intelligent transportation system. This paper developed a novel vehicle classification algorithm via magnetic field distortion signals, based on binary proximity sensor networks and intelligent neuron networks, which efficiently improves the correctness and robustness and makes it possible to replace traditional costly technologies such as loop detector, microwave radar and camera in traffic surveillance. On another hand, MSVCA (including ATDA) is only suitable to normal traffic condition, and in operating conditions the sensors have difficulty differentiating between closely spaced vehicles. Under heavy traffic volume conditions, the superposition of vehicle signals will influence the final performance. Vehicle detection and classification in heavy traffic volume conditions or traffic jam is still an important and unfathomed problem that needs further research.

ACKNOWLEDGMENT
The authors would like to thank the anonymous referees for their comments and kindly help.

REFERENCES
1. R. Alastair, et al., Intelligent transportation system, IEEE Pervasive Computing, Vol. 5, 2006, pp. 63-67. 2. A. Arora, P. Dutta, S. Bapat, et al., A line in the sand: a wireless sensor network for target detection, classification and tracking, Computer Networks, Vol. 46, 2004, pp. 605-634. 3. A. K. Lawrence, K. M. Milton, and R. P. G. David, Traffic detector handbook: third edition volume I/II, Federal Highway Administration, U.S. Department of Transportation, 2006. 4. J. Z. Li and H. Gao, Survey on sensor network research, Journal of Computer Research and Development, Vol. 45, 2008, pp. 1-15. 5. M. Tubaishat, P. Zhuang, Q. Qi, et al., Wireless sensor networks in intelligent transportation systems, Wireless Communications and Mobile Computing, Vol. 9, 2009, pp. 287-302. 6. S. Y. Cheung, S. Coleri, et al., Traffic measurement and vehicle classification with a single magnetic sensor, Transportation Research Board, 2005. 7. S. Y. Cheung and V. Pravin, Traffic surveillance by wireless sensor networks: final report, Department of Electrical Engineering and Computer Science, University of California, Berkeley, 2007. 8. B. A. Coifman, Vehicle classification from single loop detectors, Department of Electrical and Computer Engineering, The Ohio State University, 2007. 9. B. Marc, F. M. Jean, and S. Idrissa, An operational model for mobile target tracking in wireless sensor networks, in Proceedings of the 2nd International Symposium on

782

WEI ZHANG, GUO-ZHEN TAN, HUI-MIN SHI AND MING-WEN LIN

Wireless Pervasive Computing, 2007, pp. 99-104. 10. K. Seema, M. Eric, and I. Volkan, Triangulation based multi target tracking with mobile sensor networks, in Proceedings of IEEE International Conference on Robotics and Automation, 2007, pp. 3283-3288. 11. J. Aslam, Z. Butler, F. Costantin, et al., Tracking a moving object with a binary sensor network, in Proceedings of ACM International Conference on Embedded Networked Sensor Systems, 2003, pp. 150-161. 12. W. Y. Kim, K. Mechitov, J. Y. Choi, et al., On targets tracking with binary proximity sensors, in Proceedings of the 4th International Symposium on Information Processing in Sensor Networks, 2005, pp. 301-308. 13. R. Beresford, N. Shriyastava, R. Mudumbai, et al., Target tracking with binary proximity sensors: fundamental limits, minimal descriptions, and algorithms, in Proceedings of the 4th International Conference on Embedded Networked Sensor Systems, 2006, pp. 251-264. 14. J. Singh, R. Kumar, S. Suri, et al., Tracking multiple targets using binary proximity sensors, in Proceedings of the 6th International Symposium on Information Processing in Sensor Networks, 2007, pp. 529-538. 15. W. Wu and Q. S. Zhang, A method of vehicle classification using models and neural networks, in Proceedings of IEEE VTS 53rd Vehicular Technology Conference, Vol. 4, 2001, pp. 3022-3026. 16. Y. K. Ki and D. K. Baik, Vehicle classification algorithm for single loop detector using neural network, IEEE Transactions on Vehicular Technology, Vol. 55, 2006, pp. 1701-1711. 17. G. Anshul and V. Brijesh, A neural network based approach for the vehicle classification, in Proceedings of IEEE Symposium on Computational Intelligence in Image and Signal Processing, 2007, pp. 226-231. 18. J. G. Sun, J. Liu, and L. Y. Zhao, Clustering algorithms research, Journal of Software, Vol. 19, 2008, pp. 48-61. 19. L. G. Wu and Y. Ma, Design of BP neural networks classifier based on modified fuzzy c-means clustering, Control and Automation, Vol. 22, 2006, pp. 191-193. 20. G. Z. Tan, W. J. Yuan, and H. Ding, Traffic flow prediction based on generalized neural network, in Proceedings of IEEE Intelligent Transportation Systems Conference, 2004, pp. 406-409. 21. A. Gorban, B. Kegl, D. Wunsch, et al., Principal Manifolds for Data Visualization and Dimension Reduction, Springer-Verlag, Berlin, Heidelberg, 2008. 22. S. Peter, J. Hector, T. Sameer, et al., Automatic vehicle type classification using strain gauge sensors, in Proceeding of the 5th Annual IEEE International Conference on Pervasive Computing and Communications Workshops, 2007, pp. 425-428. 23. I. Urazghildiiev and R. Ragnarsson, Vehicle classification based on the radar measurement of height profiles, IEEE Transactions on Intelligent Transportation Systems, Vol. 8, 2007, pp. 245-253. 24. P. Ji, L. W. Jin, and X. T. Li, Vision-based vehicle type classification using partial gabor filter bank, in Proceedings of IEEE International Conference on Automation and Logistics, 2007, pp. 1037-1040. 25. J. Leitloff, S. Hinz, and U. Stilla, Estimation of traffic parameters in urban areas from satellite images, in Proceedings of IEEE-ISPRS Workshop URBAN, 2007, pp.

DISTRIBUTED THRESHOLD ALGORITHM FOR VECHICLE CLASSIFICATION

783

1-6. 26. H. W. Wu and M. M. Jerry, Classification of battlefield ground vehicles using acoustic features and fuzzy logic rule-based classifiers, IEEE Transactions on Fuzzy Systems, Vol. 15, 2007, pp. 56-72.
Wei Zhang () received the M.S. degrees in Software Engineering and the B.S. degree in Telecommunication Engineering from Jilin University, Changchun, P.R.C., in 2005 and 2002 respectively. He is currently working towards the Ph.D. degree in the Department of Computer Science and Engineering at Dalian University of Technology, Dalian, P.R.C. His research interests include wireless sensor networks, real-time traffic surveillance, optimal traffic control and multi-agent system, etc.

Guo-Zhen Tan () received the M.S. and Ph.D. degree in Computer Engineering from Harbin Institute of Technology, Harbin, P.R.C. and Dalian University of Technology, Dalian, P.R.C., in 1998 and 2002 respectively. He is a professor with the Department of Computer Science and Engineering, Dalian University of Technology, Dalian, P.R.C. He was a visiting scholar with the Department of Electrical and Computer Engineering of University of Illinois at Urbana-Champaign, IL, U.S., from Jan 2007 to Jan 2008. His research areas include network optimization, intelligent transportation system, and wireless sensor networks, etc. Hui-Min Shi () received the B.S. degree in Computer Science and Technology from Dalian University of Technology, Dalian, P.R.C., in 2007. She is a M.S. candidate with the Department of Computer Science and Engineering, Dalian University of Technology, Dalian, P.R.C. Her research interests include large scale traffic flow predication, intelligent algorithms including artificial neural networks, Bayesian networks, support vector machine, etc.

Ming-Wen Lin () received the B.S. degree in Computer Science and Information Engineering from Northeastern University, Shenyang, P.R.C., in 2007. He is a M.S. candidate with the Department of Computer Science and Engineering, Dalian University of Technology, Dalian, P.R.C. His research interests include wireless sensor networks and data mining, etc.

You might also like