For traffic safety and efficiency, the surveillance system based on image analysis has become popular research project in recent years. Intelligent surveillance system using pattern recognition and image processing technologies have been developed and improved intensively in this decade. In the thesis, we propose a method to automatically segment the regions of interest in the surveillance videos via analysis of vehicle trajectories. The system aims at dealing with wide-range surveillance, which is an important complement to ground-plane surveillance. The traffic scenes in experimental videos are taken from high buildings. Therefore, the vehicles in the scene are not stereoscopic. The pedestrians are just like black spots, and there are many noises in the scenes. First, we do inter-frame differencing and dilation to get the motion region. Then, we perform tracking using the ORB feature, HOG and the color histogram in the motion region. And the tracking results are stored as the trajectories. Afterwards, we use spectral method to automatically determine the number of clusters, and use K-means with modified Hausdorff distance to cluster the trajectories. Then the clusters are segmented with the orientations of trajectories. Assuming that most of vehicles move along the lanes, and lane changing seldom appears. We apply the point rotation and projection to get the valley in the distribution range of each cluster. The position of valley implies the position of lane. After the segmentation of lanes, we merge the overlapping cluster. Finally, to get the correct road map, we do connection with neighbor cluster and orientation. We conduct experiments with different challenging surveillance scenes to validate the proposed method.
