Target Detection And Classification In Traffic Video Surveillance

In recent years, intelligent video surveillance is becoming a hot research topic in computer vision, while object detection and classification are two core issues in this topic. Because of numerous factors, such as the complexity of the traffic scene, the diversity of the weather, the shading caused by the vehicles and the change of the light, it is hard for using existing object detection and classification technologies to obtain preferable effects.In this thesis, we study the detection and classification of objects in traffic video surveillance. First, we propose a background modeling approach based on three-frame images to subdivide and extract the objects in the frame. Second, according to the characteristics of traffic scene, we propose an algorithm for vehicles segmentation and mergence based on blob analyzing. Third, according to the shape features of vehicles and pedestrian, we use multi-class classification support vector machine to divide them into four categories, i.e., large car, small car, bicycle and pedestrian. Finally, we design and implement object detection and classification modules in Bupt Surveillance, a prototype system for traffic monitoring. We verify the system in real time traffic monitoring in different time periods. The experiment results show that the object detection and classification algorithms in the thesis have a good accuracy and stability.