Research On Vehicle Detection And Segmentation Methods In Video Based Traffic Surveillance

Getting accurate real-time traffic parameters is a prerequisite for intelligent traffic management. By analyzing the traffic image sequence, video surveillance method carries out the detection, recognition and tracking of traffic targets, so as to identify and judge their behaviors. Compared with other means, video surveillance can simultaneously collect various important traffic parameters, which profits to achieve intelligent traffic management, so it has a great potential.In intelligent video based traffic surveillance, before tracking and recognition, it is a common requirement for different solutions to detect and segment vehicles in the video images. Therefore, vehicle detection and segmentation methods are the key infrastructure in the video surveillance system.The common way for vehicle detection is the background difference method, but its performance must be ensured by the stability of video images, and the robustness of background estimation. Shadows in the scene and occlusion between different vehicles are two main challenges to the segmentation of vehicles. This paper did in-depth research on these issues and presents the following innovative works:(1) To guarantee the effectiveness of background difference method, this thesis proposed a three-parameter model to correct the camera motion, which achieves rapid image stabilization; and proposed a mixed multi-modal background model, based on local gray-scale distribution, which can adaptive to the multi-mode dynamics in the background of traffic scene.(2) Night images' quality is poor because of under exposure, lack of even brightness distributing, and with more red pixels .In this thesis a retinex based algorithm combined with discrete wavelet transform fusion algorithm is proposed to enhance the night image. Compare to others, the image visual effects has been improved significantly by this method.(3) For the detection of shadows, this thesis utilized the gradient and geometric characteristics to improve the traditional HSV method; the new algorithm can effectively reduce mistakes.(4) For the occlusion problem, by improving a 2.5-D description model in occlusion resolvability, and integrating 2-D convex hull segmentation algorithm, this thesis proposed a new solution, and presents an new strategy for occlusion processing in object tracking.(5) A novel algorithm is proposed to detect the vehicle's position violations. Using the 2.5-D description model to analyze self occlusion and mutual occlusion, we propose the rules of occlusion resolvability. Assuming that the vehicle's bottom and the road marks were coplanar on the road, analyzing the positional relativity of vehicle's bottom edges and the road marks, this thesis gains two detecting criterions.