Multi-object tracking via particle sampling and enhanced probabilistic data association for event detection in intelligent video systems

This dissertation presents a system framework that can automatically and robustly track multiple targets in video-based event detection applications. Combining the advantages of both the flexibility of particle sampling and mathematical tractability of Kalman filter, the tracking framework achieves both high tracking accuracy and computational simplicity. The system can distinguish occlusion and segmentation error cases and resolve those cases by constructing measurement candidates via particle sampling and an enhanced version of Probabilistic Data Association (PDA) for updating of filters. Despite the misleading information extracted from the segmentation mask when occlusion or segmentation error occurs, our proposed measurement candidate list construction procedure and enhanced PDA are still able to accurately associate correct measurements to each target at the update phase of filtering. Also, we integrate the initial occlusion handling module in the tracking system to backtrack and correct the object trajectories when initial occlusion cases are detected. The reliable tracking results can serve as the foundation for automatic event detection in intelligent surveillance systems. We also demonstrate some examples of event detection in both traffic monitoring and human surveillance applications by classifying the trajectories of the tracked objects with Hidden Markov Models (HMMs). The experimental results have shown that the proposed tracking mechanism can solve the occlusion and segmentation error problems effectively and the events can be detected with high accuracy.