Real-Time Traffic Flow Analysis without Background Modeling

Transcription

1 Real-Time Traffic Flow Analysis without Background Modeling Cheng-Chang Lien and Ming-Hsiu Tsai Department of Computer Science and Information Engineering, Chung Hua University, Hsinchu, Taiwan, R.O.C. ABSTRACT In this study, the research goal is to analyze the traffic flow with the vision-based method. However, the vision-based methods may face the problems of serious illumination variation, shadows, moving clouds, or swaying trees. Here, we propose a novel vehicle detection method without background modeling to overcome the aforementioned problems. First, a modified block-based frame differential method is established to quickly detect the moving targets without the influences of rapid illumination changes. Second, the precise targets regions are extracted by the dual foregrounds fusion method. Third, a texture-based object segmentation method is proposed to segment each vehicle from the merged foreground image blob and remove the shadows. Fourth, a false foreground filtering method is developed based on the concept of motion entropy to remove the false object regions caused by the swaying trees or moving clouds. Finally, the texture-based target tracking method is proposed to track each detected target and then apply the virtual-loop detector to compute the traffic flow. Furthermore, the traffic jam event is also detected in the proposed system. Experimental results show that our proposed system can work in real time with the computing rate above 20 fps and the average of accuracy of vehicle counting can approach 86%. Keywords: traffic flow analysis, dual foregrounds fusion, texture-based target tracking. * Corresponding author. 1

2 1. INTRODUCTION Traffic flow analyses for detecting traffic jam or traffic accidents are the important researches in the intelligent transportation systems (ITS). Traffic flow information can be obtained from different kinds of sensors e.g., loop detectors [1], ultrasonic sensors [2], or cameras [3-7]. Generally, the sensor-based methods [1-2] can not accurately analyze the traffic flow and detect some abnormal traffic events. The sensor-based method has restricted abilities to detect the speed and the number of vehicles. In addition, the sensor-based methods are very costly and hard to maintain. On the contrary, the vision-based system [3-7] can detect and track each moving vehicle and then the traffic flow can be analyzed accurately. Furthermore, some abnormal traffic events, e.g., illegally parked vehicles, traffic jams, and accidents can be monitored by the vision-based methods. However the vision-based systems are difficult to extract the accurate foregrounds on the serious illumination changing scene. The rapid illumination variation can make the conventional object detection systems with background modeling [9-14] unable to update its background model in time and cause the false object detections. In order to retrieve the efficient and accurate traffic information, these problems need to be tackled. Recently, many vision-based researches on the vehicle or human detection [3-15] were proposed. In general, the moving objects could be detected by three kinds of methods: motion-based approaches [8], background modeling [9-14], and temporal difference [15]. In the motion-based approaches, the optical flow method [8] utilizes the motion flow segmentation to separate the background and foreground regions. By applying the optical flow method [8], the moving objects can be extracted even in the presence of camera motion. However, the high computation complexity makes the real-time implementation difficult. For the background modeling methods, the construction and updating of background models [9-14] often is time-consuming. For example, in [7,14], the Gaussian Mixture Model (GMM) is frequently adopted to model the intensity variation for each pixel over a time period and need high computing cost to calculate the GMM parameters. Furthermore, the foreground detection with background modeling method is extremely sensitive to the rapid illumination variation or the dynamic background changing. In [10], the kalman filter is used to update the background model with less computational complexity. But this method can t solve the problem of serious scene change which can make the system unable to update the background model accurately. The temporal difference method doesn t need to establish the background model instead of subtracting the two adjacent frames and detecting the scene change introduced by the moving objects. The advantage of this method is less susceptible to the scene change, i.e., it has capability to detect the moving objects in dynamic environments. For example, in [15], the temporal difference method is resilient to the dynamic illumination variation, but the regions of the moving objects can t be extracted completely when the objects move slowly. Besides, if the moving object is static for some time periods, the frame difference method couldn t find the regions of objects. Based on the detection and tracking of moving vehicles, the vision-based traffic flow analysis system can be developed. The conventional target tracking systems [20, 21] are often developed on a less crowded scene. However, multiple targets tracking on a crowded scene are highly demanded for the video surveillance applications. In general, multiple targets tracking become difficult on a crowded scene because the split and merge or occlusions among the tracked targets occur frequently and irregularly. Hence, the robust target segmentation method is crucial for the tracking in the crowded traffic scene. The information of vehicle counting may be obtained from several kinds of vision-based sensor [3-7]. In [3], the authors defined a set of regions of interest on each lane in the roadway, and then the vehicles are detected and counted only in the regions of interest. In [4], the adaptive background updating and shadow elimination methods are applied to detect vehicles. The traffic flow is analyzed by computing the mean of virtual detector. In [5], background difference and virtual-loop sensor are used to detect vehicle flow. This method set one virtual-loop sensor across multiple lanes on the roadway. The non-vehicle foreground is removed according to the size of the foreground region and then the vehicles are counted with the rule-based methods. However, these methods are difficult to judge whether the vehicle appears in the virtual detector region or not under different lighting conditions. Furthermore, the lack of tracking information will make the vehicle counting inaccurate. In addition, some researchers used base lines to count the vehicles. In [6], a blob-based analysis is developed to discriminate the bike and car objects and two- lines vehicle crossing scheme is used to calculate the vehicle s velocity. In [7], the methods of adaptive background subtraction and kalman filtering are employed to detect and track the vehicles. According to the positions and trajectories of vehicles the mask of the road can be extracted to reduce computing complexity. A two-line vehicle crossing scheme is also applied to count vehicles in this study. In this paper, we propose a novel vehicle detection method without background modeling to overcome the aforementioned problems. In the proposed system, several novel methods are developed. First, a modified block-based frame differential method is established to quickly detect the moving targets accurately without the influences of rapid illumination changes. Second, the precise targets regions are extracted via the dual foregrounds 2

3 fusion method in which the short-term foreground and the long-term foregrounds are matched to extract more complete foreground region than traditional frame difference method. Third, a texture-based object segmentation method is proposed to segment each vehicle from the merged image blob. Fourth, a false foreground filtering method is developed based on the concept of motion entropy to remove the foreground regions caused by the swaying trees or moving clouds. Finally, the texture-based target tracking method is proposed to track each detected target and then the traffic flow can be analyzed. Instead of using the color feature that may change in outdoor environment to develop the target tracking, we adopted the distance transform [24] of the gradient features to serve as the object measurement in the Karman filter tracking scheme. After the vehicle detection/tracking process, some traffic analyses, e.g., vehicle counting and traffic jam detection, can be performed. The system block diagram is shown in Fig. 1. Our system consists of the following modules: foreground extraction (foreground detection with dual foreground, foreground segmentation, and true object verification), vehicle tracking (kalman filter tracking), traffic jam detection, and vehicle counting. detection method to extract more complete foregrounds. Firstly, each frame will be divided into non-overlapped blocks. The block size can be adjusted according to the object size. The foreground detection algorithm is described in the sequel. 1. We transform the RGB color space into the YC r C b color space and detect the moving object in Y channel, C r channel, and C b channel separately. 2. In each image block, if the number of detected foreground pixels exceeds a specified threshold, then we categorize this block into foreground block. 3. By fusing the foreground regions detected from Y channel, C r channel, and C b channel with voting rule, we can obtain a more complete object region. Video Input Foreground Extraction Foreground Segmentation True Object Verification Traffic Jam Detection Vehicle Tracking Vehicle Counting Fig. 1. The block diagram of the traffic flow analysis without background modeling. 2. FOREGROUND EXTRACTION WITHOUT BACKGROUND MODELING In general, foreground detection can t be accurate on the light variation or cluster scenes. In this section, we propose a novel foreground extraction method without the background modeling to segment the foregrounds from the light variation or cluster scenes. Furthermore, to detect the foregrounds efficiently and robustly, the motion-based false object filtering method is proposed Adaptive Block-Based Foreground Detection Traditional pixel-based frame difference method can cause many fragmented foreground regions because the spatial relationships among neighboring pixels are not considered. Hence, if we detect the moving objects using the frame difference method both local and global properties should be considered. Thus, we proposed the block-based foreground Fig. 2. The difference images for each color channel. Y channel. C r channel. C b channel. The reason why we adopt the voting method is that Y, C r, and C b channels have different property in foreground detection. In Fig. 2-, the objects are detected with grayscale s variation in the Y channel; while in Fig. 2-, the objects are detected with color s variations in the C r and C b channels. The relationship between grayscale and color information are complementary. If the region changes in Y channel but doesn t change in the C r and C b channels, then it may be the shadow of an object. On the contrary, the region changes in the C r and C b channels but doesn t 3

4 change in the Y channel, and then this region may be a noise region. Hence, in our system, the object detection within both the grayscale and color channels is considered. The fusing rule is designed by the voting method. We adopted the rule-based method to determine whether the pixel/block belongs to foreground or not. 1. In each image block, if the difference of the Y channel exceed 1.5 times threshold of the Y channel, we classify this block into foreground directly. 2. Otherwise, we consider the Y channel and the C r and C b channels together with the rule: (Y>T Y )&&((C r >T Cr ) (C b >T Cb )). If the pixel/block changes in the grayscale and color channels obviously together, then we regard this pixel/block as foreground. Through the careful observation, we found that the block-based method can extract more complete foreground than the pixel-based method. When the block size becomes bigger, the foreground becomes more complete and the computing time is also faster. However, the condition of wrong connection between different vehicles occurs more frequently. Fig. 3 shows the comparison for the block-based and pixel-based method. It is obviously, the moving vehicles extracted with the block-based shown in Fig. 3- are more complete than the pixel-based method shown in Fig. 3-. In our experiment, we select the block size as 2 2 to fit the size of moving objects. So far, the block-based foregrounds belong to the short-term foregrounds that will be combined with long-term foreground to generate a precise foreground that will be described later. 2.2.Orientation Adaptive Morphological Operations We apply the morphological operations to remove the noise regions and make the foreground region more complete and smooth. However, the vehicles on the highway can move along a slant angle and the symmetry structure elements in the morphological operations can merge the adjacent cars regions on neighbor lanes shown in Fig. 4. Here, we propose an orientation adaptive morphological method to overcome this problem. According to the vehicle moving direction, we apply the oriented structure elements shown in Fig. 5 to remove the noise region and make the object region more complete. Some experimental results are shown in Fig. 4. It is obvious that the morphological operation with the structure element in Fig. 5- make the image blob in Fig. 4- having the wrong connection between vehicles. The morphological operation with the structure element in Fig. 5- won t have wrong connected image blobs shown in Fig. 4-. Fig. 3. Block-based foreground extraction. Pixel-based. Block-based (2 2). Block-based (4 4). Block-based (8 8). 2.3.Precise Object Region Extraction with Dual Foregrounds The traditional frame difference method often generate strong response in the boundary of moving object, but lower response in the inside region of a moving object shown in Fig. 6-. When the object becomes larger, the incomplete object detection will be more serious. To tackle this problem we apply both the characteristic of short-term and long-term foregrounds to make the foreground more complete. The short-term foreground can define precise object boundary and the long-term foreground can extract a more complete object region shown in Fig. 6- but with motion history information. 4

5 structure element. Slant structure element with angle 135 o. Slant structure element with angle 45 o. Vertical structure element. Horizontal structure element. The long-term foreground is constructed by accumulating successive short-term foregrounds (In our experiment, the long-term foreground is constructed by accumulating 4 short-term foregrounds). By projecting the short-term foreground onto the long-term foreground and searching the precise object boundary based on the short-term foreground shown in Fig. 6-(f), and preserving all information about short-term and long-term foregrounds between the boundaries of short-term foreground, we can extract the precise region of a moving vehicle shown in Fig. 6-(g). (d 1 ) (d 2 ) (e 1 ) (e 2 ) (f 1 ) (f 2 ) Fig. 4. Morphologic operations for different structure elements. Initially detected foregrounds. After the morphologic operation using the structure element of Fig. 5-. After the morphologic operation using the structure element of Fig. 5-. Zooming of partial regions in. Zooming of partial regions in (f) Zooming of partial regions in. (f) Fig. 5. Oriented structure elements. Symmetric (g) (h) Fig. 6. Procedures for precise object region extraction with dual foregrounds. Original object. Short-term foreground. Long-term foreground. Morphologic operation. Short-term (gray region) and long-term (black region). (f) Combine with dual foreground (Red: upper bound; green: lower bound). (g) After combining with the dual foregrounds. (h) The result of dual foreground method Foreground Segmentation 5

6 In the outdoor environment, the serious shadow problem can introduce the inaccurate and false foreground detection shown as the object #4 in Fig. 7- and improper merging with neighboring objects shown as the object #2 in Fig. 7-. Based on the careful observation, the texture is not obvious in the shadow region illustrated as object #0 in Fig. 7-. The texture analysis is then used to eliminate the object shadows. Here, the texture analysis is performed by analyzing the gradient content obtained from the Sobel and Canny operations [16]. By projecting the edge information along horizontal and vertical axes, we can find the proper segmentation position (green bins) using the Ostu s method [17], which are shown in Figures 7- and 7-. Once the merged vehicle region is partitioned, each precise vehicle region shown in Fig. 7-(f) is labeled via the labeling algorithm [18]. Fig. 8. Removal of the shadow region with the property of texture. Original frame. The texture of foreground. The texture density distribution for the 6 th object region. The texture density distribution for the 7 th object region. The result of shadow removal. (f) Fig. 7. The procedure of foreground segmentation. Original image. Canny edge detection and object labeling. Edge projection to x-axis of object #2. Object labeling after vertical segmentation. Edge projection to y-axis of object #0. (f) Outcome of texture segmentation. In general, the distribution of texture density of a moving object is higher than the distribution of its shadow. Hence, we can utilize the texture densities to separate the moving object and shadow. By removing the boundary vertical regions with small texture density shown in Fig. 8-, the shadow region can be separated from the object region shown as the green regions in Fig True Object Verification For a real moving object, the direction distribution of motion vectors should be consistent, i.e., the motion entropy in the object region will be low. Based on this observation, the value of motion entropy for each detected foreground region can be used to distinguish whether the detected object is a true object or not. First, we apply the three-step block matching method [19] to find the motion vectors for each foreground region and compute the orientation histogram of the motion vectors. In Fig. 9-, the red lines represent the motion vectors of a detected foreground region. The blue points in Fig. 9- represent a static foreground region in which the values of motion vectors are very small. Second, we compute the motion entropy for each detected foreground with Eq. (1). E m K i 1 p i m log p, (1) where m is object index, i is the index of i-th bin in the orientation histogram, K is the number of total bins in the orientation histogram, and p m i is the probability of the i-th bin in the orientation histogram. In Fig. 9, the objects labeled as numbers #0 and #11 are incorrectly detected by the color noise in the 2 i m 6

7 C r and C b channels. It is obvious that the motion entropies in these false detected regions are very large and then these false detected regions can be removed with the motion entropy filtering process. Fig. 9. True object verification with motion entropy. Original Frame. Object #9 with consistent motions. Object #11 with cluttered motion. The orientation histogram of moving object #9. The orientation histogram of static object # TRAFFIC FLOW ANALYSIS The traffic flow analysis is based on the precise vehicle tracking. Some typical target tracking methods, e.g., kalman filter [20-22] and particle filter [23], need a reliable feature measurements to track the targets continuously. For example, the color histogram [12] is frequently used as the feature measurement to track the detected targets. However, the variation of color histogram of an object often occurs in an outdoor environment and makes the color feature unreliable for the target tracking. In this study, the texture feature that is not easily influenced by the illumination variation is utilized as the measurement in the target tracking algorithm Vehicle Tracking In general, the detected vehicles can be tracked with the methods of kalman filter or particle filters. With the efficiency consideration, we apply the kalman filter [22] to track each detected vehicle on the roadway. Each detected vehicle is tracked with the constant-velocity motion model. The state and measurement equations of kalman filter are defined in Eq. (2). x Ax w where, z k k Hx k 1 k 1 k v k, (2) px, k py, k px, k xk 1, zk v, x, k py, k vy, k A, H. In Eq. (2), x k-1 denotes the state at frame k-1, z k denotes the measurement at frame k, p x,k and p y,k represent the center position of the observed vehicle in frame k, and the v x,k and v y,k represent the velocities along x and y directions from frames k-1 to k. The random variables w k and v k represent process and measurement noise. They are assumed to be independent, white, and normal distributed and defined as Eq. (3). p( w) ~ N(0, Q) p( v) ~ N(0, R), (3) where, Q denotes the process noise covariance matrix and R denotes the measurement noise covariance matrix. In order to determine whether an observed vehicle belongs to a new incoming vehicle or a previously existed vehicle, we propose an efficient matching algorithm to calculate the matching distance. Here we use the texture-based features to find the reliable feature for the target matching process. For the texture matching, the texture feature can be generated from the canny edge [16] and the distance transform [24]. With the canny edge detection method we can retrieve the rough contours of the detected moving vehicles that is hard to be influenced by the illumination variation and shadow shown in Fig. 10. Then, the distant transform is applied to the canny edge context to retrieve the texture content for the target matching. 7

8 Frame #3 Frame #5 Frame #7 filter to track or update the kalman filter with observed vehicle position Vehicle Counting The conventional vision-based vehicle counting method [4] count the vehicles on the predefined virtual region on a lane that only can monitor the vehicle in that lane. Another method in [5] sets the detecting region on entire road which can be solve the problem of lane changing, but this method count the vehicle with rule-based methods. However, the moving vehicles are detected by the conventional background subtracting method such that the vehicle counting could be inaccurate under different lighting condition. Our method depends on the robustness of kalman filter tracking and then the performance of vehicle counting can be improved by the trajectories of vehicles. Here, we set a virtual detecting region on entire road to monitor whether the vehicle reach the detecting region or not. When a tracked vehicle touches the region, we start to monitor its trajectory within the detecting region. If the vehicle s trajectory satisfies the following two conditions, then the vehicle is counted. First, the vehicle is detected in the virtual detecting region. Second, the length of the trajectory of a tracked vehicle must larger than a specified length threshold (100 pixels). Some examples of vehicle counting are shown in Fig. 11. In Fig. 11, the red line is used to separate different direction of traffic flow and the green region is the detection region for counting. The black text represents the current number of passed vehicles. Frame #9 Frame #11 Fig. 10. The extractions of the texture features for each tracked vehicle shown from to. Distance transform matching is a technique for finding the matching distance between two different edge features by minimizing shifted template matching defined in Eq. (4). M O N P Tem ( m, n) Ref ( o q, p r) min,(4) q 0 r 0 where Tem is the template feature of size M N and Ref is the reference feature of size O P. Tem(x, y) denotes the pixel value in the template feature and Ref(x, y) denotes the pixel value in the reference feature at the position (x, y). If the matching value is less than the specified threshold, we can initialize a new kalman (f) Fig. 11. The procedure of vehicle counting. The tracked vehicle. The tracked vehicle is driving through the detection region. (f) The tracked vehicle passes the detection region. 8

9 3.3. Traffic Jam Detection The conventional frame difference method [15] can t keep tracking the moving object when the vehicles moving slowly because the difference of moving object will be not obvious. Here, with assistance of the target tracking scheme all the targets can be tracked continuously even though the vehicle moves slowly or stops. Some examples are shown in Fig. 12. Our work is suit for detecting event of breakdown vehicle. If there are too many cars that are stopping or moving slowly initially, the temporal difference can fail to detect the moving vehicles. The left side in Fig. 12 represents the result of vehicle stopping detection and the right side represents the result of foreground extraction. The dark red line in Fig. 12 represents the trajectory of vehicle and the black rectangular region represents predicted region. 4. EXPERIMENTAL RESULTS In this section, the performance analyses of the proposed system with traffic flow analysis are shown. The video sequence of PetsD2TeC2 [25] and the real traffic videos captured from the web site of Taiwan Area National Freeway Bureau ( are used as the test videos. First, the proposed foreground extraction without background modeling and the foreground segmentation with the texture features are used to extract the moving vehicles. Second, the kalman filter tracking is used to track the moving vehicles. Finally, a vehicle counting method and traffic jam detection are used to analyze the traffic flow The Extraction of Moving Vehicles Fig. 13- shows the video sequence of traffic scene on the freeway. Fig. 13- shows the detected foreground using the adaptive block-based foreground detection. Fig. 13- shows the results of the precise object region extraction using the method of dual foregrounds fusion. The long-term foreground is constructed by accumulating 4 short-term foregrounds. Fig. 13- shows that the bounding boxes on the detected vehicles. Frame #7 #17 #27 #37 #47 #57 #67 #77 Fig. 12. The video sequence of vehicle stopping. The driving vehicle is shown from to. Slow moving vehicle is detected. Stopping vehicle is detected. #87 #97 9

10 #7 #17 #7 #17 #27 #37 #27 #37 #47 #57 #47 #57 #67 #77 #67 #77 #87 #97 #87 #97 #7 #17 #7 #17 #27 #37 #27 #37 #47 #57 #47 #57 #67 #77 #67 #77 #87 #97 #87 #97 Fig. 13. Moving object detection. A traffic flow 10

11 scene. The foregrounds are detected by adaptive block-based foreground extraction. The detected foregrounds are improved with dual foreground and noise removing processing. The foregrounds are extracted with foreground segmentation. The regions of precise moving objects are labeled with red rectangular box. (i) (j) Here we use the PetsD2TeC2 video sequence to evaluate the performance of foreground detection for our method, traditional frame difference method, and the method of background subtracting with Gaussian Mixture Model method. The block size is chosen as 2 2 to fit the size of moving object in our method. The lower bound of object size is set to 10 blocks for the human detection. The foreground detections in Fig. 14 (f)(g)(h) show the scene adaptability and fragmented region for the frame difference method. The fragment often occurs on large moving object in Fig. 14 (g). The foreground detections in Fig. 14 (i)(j)(k)(l) show the object completeness of the background modeling method [25]. Fig. 14-(l) shows a noisy image that is generated by the inefficient background updating process. The foreground detections in Fig. 14 (m)(n)(o)(p) show the scene adaptability and detection completeness of our proposed method. It is obvious that our method outperforms the other typical methods in terms of scene adaptability and detection completeness. (k) (m) (l) (n) (o) (p) Fig. 14. The examples of foreground detection at frames #325, #683, #1023 and #2517. The original frames are shown from to. The foreground detection using the conventional frame difference method is shown from to (h). The foreground detection using the GMM background modeling method is shown from (i) to (l). The foreground detection using our proposed method is show from (m) to (p). (g) (f) (h) 4.2. Traffic Flow Analysis The vehicle tracking is illustrated in Fig. 15. In Fig. 15-, the moving vehicles start to be tracked. Fig. 15-(f) show the tracking of a new vehicle after foreground segmentation. Fig. 15-(g)(h) show the correct label of track after merge and split of the two foreground. The red rectangular represents the untracked object and the yellow rectangular represent the tracked object. The red number denotes the number of the tracked vehicle. After vehicle tracking, we set a detecting region on the roadway to count the moving vehicles. The detecting region across multiple lanes is used to count vehicles. Some example of vehicle counting is shown in Fig. 16. Fig. 16- show a vehicle passing through the detecting region and the counter is added by one. Fig. 16- show that two vehicles pass through the detecting region in the same time and the counter is accurately added by two. 11

12 (f) (g) (h) (f) (i) (j) (g) (h) Fig. 15. Vehicle tracking on crowded vehicle scene. In Fig. 16-(f)(g)(h), three vehicles are moving side by side and our foreground extraction method can separate them well. In Fig. 16-(i)(j)(k), two overlapped vehicles can be count accurately. The performance of vehicle counting is illustrated in Table 1. In the first and second video clips, the density of traffic flow is low. Hence, the performance of the vehicle tracking is satisfied. In the third video clip, the performance is reduced because a few buses introduce the serious occlusions and wrong segmentation. (k) Fig. 16. Some examples of vehicle counting. The traffic jam detection is shown in Fig. 17. In the left side of scene has a gateway, the vehicle sometimes stops there. Our method can detect stopping vehicle with the result of vehicle tracking. (f) Fig. 17. Traffic jam detection in the left side of scene. The vehicle #7 in driving is shown in and. The vehicle #7 stops about 4 seconds. Another vehicle #0 is coming. The detected vehicles are identified as moving slowly in shown in and (f). 12

13 1 st video clip 2 nd video clip 3 rd video clip Table 1. The accuracy analysis of the vehicle counting. Actual number Detected number 5. CONCLUSION Our research goal is to analyze the traffic flow automatically with the vision-based methods. However, the vision-based methods may face the problems of serious illumination variation, moving shadows and influences of moving clouds or swaying trees. Here, we propose a novel vehicle detection method without background modeling to overcome the aforementioned problems. The modified block-based frame differential method, the precise object region extraction with dual foregrounds, the foreground segmentation, and the true object verification show the scene adaptability and detection completeness. The texture-based target tracking method is proposed to track each detected target and then apply the virtual-loop detector to analyze the traffic flow. Furthermore, the traffic jam event is also detected in the proposed system. Experimental results show that our proposed system can work in real time with the rate above 20 fps and the accuracy of vehicle counting can approach 86%. REFERENCES [1] Y. Zhang and Z. Ye, A Derivative-free Nonlinear Algorithm for Speed Estimation using Data from Single Loop Detectors, IEEE Intelligent Transportation Systems Conference, pp , September [2] M. Yamada, K. Ueda, I. Horiba, and N. Sugie, Discrimination of the Road Condition Toward Understanding of Vehicle Driving Environments, IEEE Transactions on Intelligent Transportation Systems, Vol. 2, No. 1, pp , March [3] K. F. Wang, Z. J. Li, Q. M. Yao, W. L. Huang, and F. Y. Wang, An Automated Vehicle Counting System for Traffic Surveillance, IEEE International Conference on Vehicular Electronics and Safety (ICVES), pp. 1-6, [4] M. Lei, D. Lefloch, P. Gouton and K. Madani, A Video-based Real-time Vehicle Counting System Using Adaptive Background Method, IEEE International Conference on Signal Image Technology and Internet Based Systems, pp , False positive False negative accuracy % % % [5] B. Qin, M. Zhang, S. Wang, The Research on Vehicle Flow Detection in Complex Scenes, IEEE International Symposium on Information Science and Engineering, Vol. 1, pp , [6] T. H. Chen, Y. F. Lin, and T. Y. Chen, Intelligent Vehicle Counting Method Based on Blob Analysis in Traffic Surveillance, Second International Conference on Innovative Computing, Information and Control, pp , [7] E. Baş, A. M. Tekalp, and F. S. Salman, Automatic Vehicle Counting from Video for Traffic Flow Analysis, IEEE Intelligent Vehicles Symposium, pp , June [8] X. Dai and S. Khorram, Performance of Optical Flow Techniques, IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp , June [9] S. Kamijo, Y. Matsushita, K. Ikeuchi, and M. Sakauchi, Traffic Monitoring and Accident Detection at Intersections, IEEE Transactions on Intelligent Transportation Systems, Vol. 1, No. 2, pp , June [10] J. Zhou, D. Gao, and D. Zhang Moving Vehicle Detection for Automatic Traffic Monitoring, IEEE Transactions On Vehicular Technology, Vol. 56, No. 1, pp , January [11] R. Khoshabeh, T. Gandhi, and M. M. Trivedi, Multi-camera Based Traffic Flow Characterization & Classification, IEEE Intelligent Transportation Systems Conference, pp , [12] B. T. Morris and M. M. Trivedi, Learning, Modeling, and Classification of Vehicle Track Patterns from Live Video, IEEE Transactions on Intelligent Transportation Systems, Vol. 9, No. 3, pp , September [13] P. Negri, X. Clady, S. M. Hanif, and L. Prevost, A Cascade of Boosted Generative and Discriminative Classifiers for Vehicle Detection EURASIP Journal Advances in Signal Process, Vol. 2008, No. 2, 2008, pp

14 [14] C. C. Lien, J. C. Wang and Y. M. Jiang Multi-Mode Target Tracking on a Crowd Scene, Proceedings of the Third International Conference on International Information Hiding and Multimedia Signal Processing, Vol. 02, pp , [15] R. Jain, W. Martin and J. Aggarwal, Segmentation through the Detection of Changes due to Motion, Compute Graph and Image Processing, Vol. 11, pp , [16] J. Canny, A Computational Approach to Edge Detection, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 8, No. 6, pp , [17] P. S. Liao, T. S. Chen and P. C. Chung, A Fast Algorithm for Multilevel Thresholding, Journal of Information Science and Engineering, Vol. 17, pp , [18] S. G. Dellepiane, F. Fontana, and G. L. Vemazza, Nonlinear Image Labeling for Multivalued Segmentation, IEEE Transactions on Image Processing, Vol. 5, No. 3, pp , March [19] T. Koga, K. Iinuma, A. Hirano, Y. Iijima, and T. Ishiguro, Motion Compensated Interframe Coding for Video Conferencing, in Proc. Nat. Telecommunication Conf., Nov. 29 Dec. 3, pp. G5.3.1 G5.3.5, [20] C. Wren, A. Azarbayejani, T. Darrell, and A. Pentland, Pfinder: Real-time Tracking of the Human Body, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 19, No. 7, pp , July [21] M. Xu, J. Orwell, L. Lowey and D. Thirde, Architecture and Algorithms for Tracking Football Players with Multiple Cameras, Image and Signal Processing, IEE Proceedings, Vol. 152, No. 2, pp , April [22] G. Welch and G. Bishop, An Introduction to the Kalman Filter, [23] K. Nummiaro, E. K. Meier, L. J. V. Gool, An Adaptive Color-based Particle Filter, Image Vision Computing, Vol. 21, No. 1, pp , [24] A. Ghafoor, R. N. Iqbal and S. Khan, Image Matching Using Distance Transform, SCIA 2003, LNCS 2749, pp , [25] T. Huang, J. Qiu, T. Sakayori, S. Goto and T. Ikenaga Motion Detection Based On Background Modeling And Performance Analysis For Outdoor Surveillance, IEEE International Conference on Computer Modeling and Simulation, pp , Journal of Information Technology and Applications 14