Vehicle Detection And Tracking Under Occlusion

The result of vehicle detection, tracking and behavioral state is directly related to the core functions of ITS system implementation. Particularly, as the motor vehicles increase in the current days, the road block is much more serious, and the result of block partition in vehicle detection and tracking procession plays a decisive role.Detection and tracking of vehicle and behavior analysis technology is widely used, such as parking control, road congestion detection, monitoring, toll scenarios and other situations.This paper focuses on vehicle detection, vehicle occlusion detection and segmentation, vehicle tracking under occlusion and the analysis of vehicle behavior state, and the approaches to solve the problems. The main contributions in this paper are as follows:(1) For the problem that it is hard to extract background and the foreground usually contains cavities, the paper proposes an improved histogram-mean model background reconstruction algorithm. After extracting the moving vehicle by background subtraction algorithm, it proposes a hole-filling algorithm for the extracted vehicle based on foreground-scan-and-statistic method, and finally obtain the whole vehicle foreground.(2) There are issues that it is hard to detect the vehicle because of the occlusion and the results of segmentation are not good. For the above issues, the paper proposes a statistical model based vehicle occlusion detection and segmentation algorithm, which distinguish the blocked vehicle and the blocking vehicle.(3) Considering that the tracking loss frequently happens when the vehicle is blocked and it is unable to track the vehicle when it is severely blocked, the paper proposes a vehicle-region-matching tracking method. It is based on kalman-filter forecasting algorithm and track after distinguishing the blocking situation. (4) Analyze the behavior state, introduce finite state automata based on the tracking of vehicles, do micro-expressing for the vehicle trajectory information, determine the vehicle state and identify the simple action of the vehicle.Moreover, this paper also designs and performs several experiments on the proposed methods. The experimental results show that these methods are feasible and effective.