VEHICLE DETECTION USING IMAGE PROCESSING FOR TRAFFIC CONTROL AND SURVEILLANCE SYSTEM

Riza Atiq bin O.K. Rahmat, Assoc. Professor, Faculty of Engineering, Universiti Kebangsaan Malaysia, 43600 Bangi, Malaysia. Phone: 60-3-89296227, Fax: 60-3-89252546, e-mail: riza@eng.ukm.my

Kasmiran bin Jumari, Assoc. Professor, Faculty of Engineering, Universiti Kebangsaan Malaysia, 43600 Bangi, Malaysia. Phone: 60-3-89296193, Fax: 60-3-89254675, e-mail:kbj@pkrisc.cc.ukm.my

SUMMARY
Surveillance cameras are normally installed at major road links and intersections in urban area for observation by human operators. Instead of limiting the function of the cameras for these purposes, the images from the cameras can be captured and analysed for further decision making processes. This paper suggests image processing algorithms for traffic counting, queue length, speed measurement and vehicle classification. Traffic counting algorithm adopted in this study is carried out by observing changes in pixels values in the middle of traffic lanes. The queue length measurement, vehicle classification and determination of speed, a single line of pixels placed along a traffic lane are observed. The pattern of these pixels values are used to measure the queue length, length of individual vehicle and to detect the position of a particular vehicle within a short interval of time.

INTRODUCTION
Most traffic control systems in the world use point-based inductive loops as the sensor. During normal condition where traffic flows are not exceeding approach road capacity, the current traffic control systems are performing satisfactorily. However, during traffic congestion periods, it is frequently observed that the queue length at approach roads in most intersections become imbalance. This problem is partly due to the inability of the pointbased inductive loops to measure the queue length. As a way of solving this problem, surveillance cameras have often been installed at major intersections for observation and intervention by human operators. Instead of limiting the usage of the camera for these purposes, the images from the camera can be analysed by employing image processing technology. This paper suggests algorithms for traffic counting and classification as well as queue length and speed measurement.

IMAGE PROCESSING
The basic components of the image processing system used in this study are video camera which produces composite signals, image grabber to digitised the signals and a Windows based computer. In this study, each video frame is digitised into 350 X 286 RGB (Red Green Blue) picture elements or pixels. Each pixel is assigned an integer value for each colour representing the light intensity. The value is between 0 for black and 255 for white (1). Once stored in the memory, the image can be considered as an array of integer numbers. Software can therefore be written to manipulate data to extract information for the image or sequence of images, and thereby realise the image processing function (2). Colour images were used in this study because it is easier and cheaper to get equipment for colour images than that of the grey scale. At the same time it is easier to detect front and rear edges of a high speed vehicle using colour images..

EXISTING IMAGE PROCESSING ALGORITHMS FOR TRAFFIC DETECTION

Digitised video images usually contain enormous amount of data. A stream of more than 7 Mbytes of data per second will be transmitted by a typical 350 X 286 pixel RGB video images at the rate of 15 frames per second (1). This amount of data is far too large for a normal Pentium based computer to handle and therefore impossible to be carried out in realtime. Since the traffic detection system must be capable of working in real time, data reduction techniques are employed by processing the pixels only at the predefined area where vehicles are visible in the image. In the case of the loop emulator developed by Rourke and Bell (3), the algorithm processes only pixels in small defined windows, thereby reducing the data to 1 / 1,500th of the original data. For congestion and queuing detection, Rourke and Bell (2) and Versavel and Boucke (4) have proposed an algorithm that analyse data from a window consisting of a single line of pixels placed along a traffic lane. In the Rourke and Bell, the algorithm processes the data only once every few seconds using the Fast Fourier Transformation. Cypers et al., (5) have developed a queue detection system called Camera and Computer Aided Traffic Sensor (CCATS) which process pixels in programmable windows within the digitised image. Hoose (6) and Hoose et al., (7) developed an incident detection system for highways and urban roads known as IMPACTS by analysing pixels in a number of cells which divided up an image in a cycle of 1 to 5 seconds. Michalopoulos et al., (8) have developed an automatic incident detection system through video image processing by observing pixels on detection lines drawn across traffic lanes. The detection lines exist only on the monitor and not in the pavement and they can be easily removed or adjusted. Every time a vehicle passes through these lines, a detection signal in the form of changes in pixels value, is generated which is similar to the signal produced by loop detectors. Thus, the system with one camera can easily replace many loop detectors. Currently the image processing based incident detection system is gaining acceptance as a more effective tool than the conventional inductive loop detectors Michalopoulos et al., (9)

PROPOSED ALGORITHMS
Image processing in this study is developed for use as traffic control sensor to obtain multilane traffic volume, queue length and down stream congestion that may obstruct traffic flow exiting a particular approach. The algorithm must be sufficiently efficient for real time processing by a Pentium based computer.

Traffic Volume Counting

For traffic counting, the algorithm adopted in this study is developed by observing a few pixels for each traffic lane, which are considered as detectors as shown in Figure 1. The pixel values will surge or shrink drastically every time a vehicle passes the detectors. A computer programme is then written to count this surge or shrinkage.

300 250 pixel value 200 150 100 50 1

A bright coloured vehicle is passing

A dark coloured vehicle is passing

11 16 21 26 31 36 41 46 51 56 Time (second) 61 6

Figure 1: The location of the observed pixels and the fluctuation of a detection pixel value.

Queue Length Measurement

In the queue length measurement, pixels along a single line in the middle of a traffic lane are observed. Pixels which are almost uniform in values correspond to the road surface and hence no vehicle. On the other hand variation in pixel values indicates the presence of objects of different colours. In the first step the algorithm uses this phenomena to determine whether the image at the specific window is the road surface or vehicles. The decision is made that vehicles occupy the road when the standard deviation of the pixel values is more than 30% of the average value. If the road is occupied by vehicles, the algorithm compares pixel values from two frames of 5 second intervals. The comparison is carried out segment by segment of 10 pixels long. Assuming that an ordinary car will take about 6 meters of road length during queuing, the pixels along the detector line can be calibrated with a simple trigonometry technique to get the number of vehicles queuing in term of passenger car units.

Correlation between two sets of pixel values from two frames of the same segment is calculated for every 5 seconds. If the correlation, R2, is more than 0.9 then the algorithm will decide that a queue occurs in that particular segment. Figure 2 shows the comparison of two frames of 5 seconds interval.

250 200 Pixel value 150 100 50 0 1 11 21 31 41 51 Distance (m) 61 71

Almost identical pixel patern

250 200 Pixel value 150 100 50 0 1 11 21 31 41 51 Distance (m) 61 71

Figure 2: Comparison of pixel values from two frame of 5 seconds interval

Classification and Speed

Classification is carried out by measuring the vehicle length. First, a vehicle was detected using the pixels value in the middle of the traffic lane as shown in Figure 3. The vehicle length was measured using the pixels value along the parallel line and a simple trigonometry calculation (1). The speed is determined by detecting the front edge of a particular vehicle in two video frames of known interval. The travel distance between two frames was measured using the pixels value along a parallel line with the traffic flow in the middle of the traffic lane as shown in Figure 4. Time interval between the two frames is 0.25 second and the calculated travel distance is 5.9 meters. Therefore the speed is approximately 85 km/h.

300 250 Pixel Intensity 200 150 100 50 0 1 11 21 31 41 51 61 71 81 D istance (1 unit = 0.2 meter)

4.3 m

Figure 3: Classification of a Vehicle Based on Its Length.

5.9 m

Figure 4: Speed Measurement by comparing two frames of 0.25 second interval

EXPERIMENTAL RESULTS
The experiment for the traffic counting algorithm was carried out at Jalan Raja Laut intersection with Jalan Isfahan, Kuala Lumpur. The result was compared with that of manual counting. Assuming that manual counting is correct, in general the automatic counting is very close to the manual counting. Statistical analysis shows that both counting are correlated closely to each other with R2 = 0.995 during day time and R2 = 0.98 during night time. Closer inspection reveals that the main reason for the error during the day time is due to pedestrians passing through the detectors causing the detector pixels to fluctuate. Besides that, vehicles of the same colour with that of the road surface, will not be detected. It was noted that in the early morning or late afternoon when the sun is very low, the back lighting effect reduces the clarity of the image especially for east-west direction. Steeper camera angles can reduce this effect. During the night, vehicle head lights, especially with high beam, occasionally caused the camera to run out of focus. During this period of time, the detection became irrational. This problem can be solved with higher camera position and steeper camera angle to avoid direct light from the vehicle head lights. In this study the camera was positioned approximately 10 meters from the road surface at an angle of 35o from horizontal and it was found out to be sufficient to avoid sun light and normal vehicle head light. Experiment for queue length measurements was carried out at the intersection of Jalan Sultan Ismail with Jalan Tuanku Abdul Rahman, Kuala Lumpur. Comparison of the number of vehicle queuing obtained automatically with manual counting for 20 cycles of the traffic light shows both data are closely correlated to each other with R2 = 0.92 during day time and R2 = 0.80 during night time. Closer inspection of queues revealed that one or two very slow moving vehicles at the distant end of the queues were recorded by the manual enumerator as queue but the algorithm considered the vehicles as moving objects and therefore were not part of the queues. Long vehicles such as buses and lorries are also a source of error. While the enumerator counts such vehicle as one queued vehicle, the algorithm counts it as more than one since it takes more than 6 meters of road space. Experiment for classification and speed measurement were carried out at the North-South Highway near Kajang Interchange. Classified volume counts obtained by the image processing are compared with that of manual counts. Again both data are closely correlated to each other with R2 = 0.99 during day time and R2 = 0.96 during night time. Most of the error happened when the vehicles travel at the speed greater that 150km/hr, and when the frame grabber rate is set at 8 frames/second, such vehicles sometimes cannot be detected. Experiment to measure travel speed was carried out using a test car. Assuming the speedometer is correct, the speeds of ten passes of various speeds measured by the image processing were comparable with the speeds obtained from the speedometer. Coefficient of determination (R2) between the two data sets are 0.99 during day time and 0.89 during night time.

CONCLUSIONS
The image processing algorithms to obtain multilane traffic volumes and queue length using a single camera at an intersection mentioned above are fast and efficient enough to be executed by a Pentium based computer. A single camera with four algorithms (ie. traffic volume, queue length, classification and speed measurement) can replace many inductive loops, thus making wide area detection cost effective. For traffic control, the algorithms are able to provide two more parameters in addition to the traffic volumes, namely, down stream congestion and the queue length at approach roads, thus a more effective controller can be developed. During the experiment it is found out that the camera position is very crucial in getting accurate data. Higher camera position and steeper camera angle are necessary to avoid direct sun light and vehicles head light. This study is carried out using Pentium II 400 based computer, multimedia video blaster as the image grabber and colour CCTV camera. These equipment are readily available in the market with relatively low priced. In addition, the computer is also used as the traffic controller. There is another advantage when using a computer based system, in which they can be easily networked for co-ordination between intersections as normally required in Urban Traffic Control system.

Acknowledgement
The authors gratefully acknowledge Ministry of Science, Technology and Environment for funding this research and Universiti Kebangsaan Malaysia for generous assistance throughout this research project.

References
(1) R. A. Rahmat, K. Jumari, A. Hassan and H. Basri; Image Processing for Traffic Surveillance System, To be presented at Second International Conference on Advances in Strategic Technologies, Kuala Lumpur, August, 2000. A. Rourke and M.G..H. Bell; Queue detection and Congestion Monitoring Using Image Processing; Traffic Engineering and Control; 32(9); London; 1991; 412 421. A. Rourke and M.G..H. Bell; Application of low-cost image processing technology in transport; Proceeding of 5th World Conference on Transport Research; Yokohama; 1989; 67-77. J. Versavel and B. Boucke; Vedeo for Traffic Data & Incident Detection by Traficon; Proceding of 5th International Conference on Application of Advanced Technologies in Transportation; California ; 1998; 33 40.

(2)

(3)

(4)

(5)

L . Cypers, G. Kolancny, J. M. Poncelet, P. Vervenne, F. Lemaire and F. De Jaegere, CCATS. The image processing-based traffic sensor. Traffic Engineering & control 31(6); London; 1990; 371 375. N. Hoose; IMPACTS: An image Analysis Tool for Motorway Surveillance; Traffic Engineering & control 33(3); London; 1992; 140 147. N. Hoose, M.A. Vicenci, and X. Zhang; Incident Detection in Urban Roads Using Computer Image Processing; Traffic Engineering & control 33(4); London; 1992; 236 244. P. G. Michalopoulos, R. D. Jacobson, C. A. Anderson and T. B. DeBruycker; Automatic Incident Detection Through Video Image Processing; Traffic Engineering & control 34(2); London; 1993; 66 75. P. G. Michalopoulos and D. Panda; Recent Advances in Implementation of Machine Vision Technology in Freeway. Proceding of 5th International Conference on Application of Advanced Technologies in Transportation; California; 1998; 233 246.

(6)

(7)

(8)

(9)