Research On Object Tracking Technology In Intelligent Video Surveillance

Intelligent Video Surveillance (IVS) system can detect and process the abnormal events automaticly. It has drawn many researchers’attention as it meets the requirements of "smart city" and "safe city". As one of the core technologies in IVS. object tracking is the foundation of many advanced applications (e.g. object recognition, object classification, abnormal behavior detection, etc.), and it has very important research value. However, developing a robust and real-time tracker is still a challenging and promising problem due to the dynamic change of the surveillance environment (e.g. illumination change, camera motion, etc.) and dynamic change of the tracked objects (e.g. pose variation, scale change, occlusion, etc.).This dissertation mainly focuses on the study of object tracking technologies in the Intelligence Video Surveillance, including the online single object tracking algorithm based on static single camera, the single object tracking algorithm based on moving single camera, the multi-object tracking algorithm based on static single camera, and the multi-object tracking algorithm based on multi-camera. The main contributions of the dissertation are summarized as follows:Firstly, the online single-object tracking algorithm based on the single camera is studied in the dissertation. The sparse coding and linear subspace learning algorithms are introduced and analyzed. In order to overcome the existing problems of the traditional online object tracking algorithms, we present two visual object tracking algorithms:one algorithm based on the Maximum Likelihood Estimation (MLE) and L2-norm; and the other algorithm based on feature grouping. Compared with other state-of-the-art methods, the proposed methods can track the objects stably and robustly under some abnormal conditions (e.g. occlusion, rotation, scale change, illumination variation, etc.).Secondly, the single-object tracking algorithm with cameras mounted on moving platforms is studied. Feature points are selected. And then the global motion of the camera is estimated by using optical flows on the rest feature points. To achieve robust tracking on unstable video sequences, the proposed algorithm modifies the particle filtering method according to the global motion of camera and adopts the block-based color histogram as the appearance model. Experiments demonstrate that our algorithm could track moving object robustly on challenging videos captured by moving cameras. Meanwhile, our algorithm overcomes the disadvantages of the traditional tracking algorithms that require processing video stabilization and tracking separately. Thus, it reduces the computational complexity and therefore improves the processing speed. Thirdly, the multi-object tracking algorithm based on static single camera is studied in this dissertation. We put emphasis on introducing the graph cuts theory and discussing how to establish the relationship between object tracking and graph theory. Based on the discussions, a multi-object tracking algorithm based on the graph cuts theory is proposed. The proposed algorithm combines the color and motion information of the pixels to establish the energy function and construct the graph. Then the algorithm minimizes the energy function using the max-flow/min-cut method and therefore achieves multiple objects tracking. The experimental results demonstrate that our multi-object tracking method is robust to occlusion and variation of the number of objects.Finally, the multi-object tracking algorithm based on static multi-camera is studied. Based on moving objects detection using the Codebook model, we propose two multi-object tracking algorithms. The first algorithm computes the homography matrix by using the information of vanishing points. Multiple objects are located at multiple planes and then tracked by using the graph cuts theory. The second algorithm computes the view-to-view homography matrix using several landmarks on different planes and performs multi-objects tracking by the shortest paths optimization algorithm. The experimental results demonstrate that our tracking methods have strong robustness to illumination change, occlusion and complex motion. In addition, the second algorithm achieves real-time performance.In terms of robustness and real-time performance, the above-mentioned works enrich the theory of object tracking, and make some valuable exploration on the applications of the theory in the related fields.