Research On Video Based Vehicle Tracking Algorithm In Complex Environment

Traffic information collection is the basis of traffic management, control and guidance. Compared with traditional loop vehicle detector, video-based vehicle tracking technology has the advantages of more convenient installation and maintenance, larger detection zone and more detectable parameters. Though video-based vehicle tracking technology has huge potential for traffic information collection, its reliability and accuracy may be influenced a lot by complex environments such as changing illumination and occlusion between vehicles. So the research on vehicle tracking in complex environment has important significance for traffic information collection. This paper studies the vehicle tracking algorithm under changing illumination and occlusion, and the main research contents include the following aspects:(1) To detect moving vehicles under changing illumination, a self-adapting ViBe(visual background extractor)background model has been proposed. Firstly, it is analyzed that if the ViBe background model's threshold is too small, some background region may be detected mistakenly as vehicle and if the threshold is too large, some vehicle region may be detected mistakenly as background. Then, the error functions of vehicle detection are defined. The first kind of error function measures the small threshold'influence, and is defined by the ratio of the area of the background region which detected mistakenly as the vehicle to the area of the whole image. The true vehicle region is expressed approximatively by the minimum convex hull which contains the detected vehicle region. Taking the convex hull as a reference, the average relative error of all the vehicles in the image is calculated, and defined as the second kind of error function to measure the large threshold'influence. Finally, based on the criterion that a small error of the second kind is more needed than a small error of the first kind, a set of conditions is set to judge whether the current threshold is reasonable or not, and the threshold is adjusted according to a certain step length. Experimental results show that the self-adapting ViBe background model can improve the vehicle detection accuracy under changing illumination.(2) To improve the durability and accuracy of vehicle tracking, a motion vector based vehicle tracking algorithm under the framework of non-grid blocking has been proposed. Firstly, it's analyzed that there are two shortcomings for the motion vector based vehicle tracking algorithm under the framework of grid blocking. One is that in the grid blocking the small vehicle contains little blocks, and the influence of noise motion vectors may increase and reduce the durability of tracking. Another is that in the step of block extension, the noise region and nearby vehicle may influence the object vehicle a lot and decrease the accuracy of tracking. Aiming at these shortcomings, firstly the vehicle is blocked in a non-grid manner rather than in a grid manner. So the blocks can locate at any position of the image, and they can overlap with each other. In the proposed non-grid blocking, the small vehicle region can contain more enough blocks and be influenced little by the noise motion vectors. This will increase the durability of tracking. Then in the step of block moving, each block is moved by its own motion vector. To do this, the noise motion vertors are detected and adjusted according to the spatial continuity of motion vectors. This manner of block moving makes the block extension omissble, decreases the influence of noise region and nearby vehicle, and improves the accuracy of vehicle tracking. Experimental results show that compared with the motion vector based vehicle tracking algorithm under the framework of grid blocking, the proposed algorithm has better durability and higher accuracy of tracking.(3) A Markov random field based occlusion handling algorithm under the framework of non-grid blocking has been proposed. When the motion vector of the occluding vehicle and that of the occluded vehicle is close, the Markov random field based occlusion handling algorithm under the framework of grid blocking can't segement occlusion region well. Aimming at this problem, an irregular spatio-temporal neighbourhood system is defined using Euclidean distance to describe the geometrical relationship between blocks. Then based on the neighbourhood system, a Markov random field under the framework of non-grid blocking is built. A two-dimensional gaussian vector which consists of the gray level difference of motion and the distance between color histograms is used to define the new energy function. The energy function mixes the vehicles'motion information and color information together, and can express the difference between the occluding vehicle and the occluded vehicle more completely. Finally the energy function is optimized by simulated annealing method to segment the occlusion region. Experimental results show that when the vehicles'motion vectors are close, the proposed algorithm can segment occlusion region accurately by using the color difference between vehicles. It can extend the usable range and improve the accuracy of occlusion handling.(4) To handle the occlusion between two vehicles which have both just entered the traffic scene, an ellipse fitting algorithm with a fixed direction of major axis is proposed to segment the occlusion region initially. Firstly, the occlusion's existence is judged using the ratio of the foreground to its convex hull, and the foreground contour is divided into certain groups of sub-contours by a set of straight lines which are parallel to the lane line. Then, by adding the constraint that the major axis'direction is fixed, the least square ellipse fitting algorithm with a fixed direction of major axis is proposed and used to fit each group of sub-contours. Finally, the evaluation function of ellipse fitting is defined by Newton iteration method and minimized to find the optimal segmentation line. As a result, the vehicles' initial locations can be got. Experimental results show that the proposed algorithm has good robustness to the irregular foreground contours and can get accurate initial locations of the vehicles.(5) To verify the effectiveness of the proposed algorithms, a vehicle tracking software is designed and realized in Matlab2011 environment. The software can detect and track vehicles in traffic scene. Furthermore, using the vehicles' trajactories it can also detect traffic volume, and traffic behaviors such as parking and lane changing.