Cluster Analysis And Anomaly Detection Of Video-based Vehicle Trajectory

Vehicle trajectory represents the vehicle behavior intuitively. We can extract the traffic scene and monitor the vehicle behavior by analyzing vehicle trajectory. This thesis described and analyzed vehicle behavior by trajectory extracting, trajectory clustering and anomaly detection. Direction feature of trajectory, trajectory similarity and anomaly detection are emphatically researched. Related concepts and algorithms are proposed. Verifying test with real vehicle trajectory data indicated that these algorithms are reasonable and effective.The main works and contributions in this thesis are followed:Firstly, we realized vehicle video-based detection and tracking using background subtraction and Kalman filtering. In this way we can extract vehicle trajectory providing trajectory data for trajectory analysis later. Secondly, we proposed two new methods to measure the similarity between vehicle trajectories based on studying trajectory direction feature. the effectiveness and performance of similarity are validated by cluster the trajectory data using real trajectory data. Thirdly, we extracted the typical vehicle trajectory based on trajectory clustering and created the statistical model of trajectory based on preceding clustering result. Finally, realized anomaly detection of vehicle trajectory using single-point detection and multi-point detection based on the statistical model. Single-point method only detect position abnormality, but multi-point method detect direction abnormality else.