With The Adaptive Video Tracking Algorithm Occlusion Handling With Template Drift Suppression

Visual object tracking is under intensive research in the field of computer vision and has a very wide range of applications, including human-computer interaction, auto-monitoring, video retrieval, traffic surveillance, vehicle navigation, and etc. Realizing robust and accurate visual object tracking is challenging, because 1) the target appearance keeps changing; 2) template drift frequently occurs in which the template is gradually occupied by the background; 3) the target is often occluded during tracking, and the tracker would have a difficult time deciding if the change in the target appearance is caused by the target itself or by occluders; 4) tracking precision degrades a lot when the target is under partial occlusions; 5) it is very difficult to determine at which frame a completely occluded target re-emerges.Researchers have proposed many algorithms trying to solve the aforementioned difficulties. However, the performances of those algorithms are not satisfactory enough. The tracking algorithm proposed in this thesis provides a much better solution to the problems mentioned above. In order to shield out occluders, the proposed algorithm uses masked template matching to locate the target. As the template mask is generated according to the occlusion situation, it is crucial to effectively analyze the occlusion situation. By combining the information provided by spatiotemporal context, target appearance model, and motion constraints together, the proposed algorithm achieves a much more robust analysis of occlusion situations. As the template mask derived from the previous occlusion situation might be incorrect in the current frame, the masked template matching might yield inaccurate results. The proposed algorithm utilizes the non-occluded part of the target to rectify the erroneous target location. After the target is accurately located, the template is updated, in which the goal is to inhibit template drift to the greatest extent. By explicitly modeling the drift noise in the Kalman appearance filter, the proposed algorithm yields lower updating rate for regions more susceptible to template drift and hence strikes an optimal balance between updating the template and inhibiting drift. To handle complete occlusions, the proposed algorithm exploits the difference between the reappearing target and the foreground/background outliers in terms of their template and backward matching errors, so that the end of a complete occlusion can be detected robustly. To further enhance robustness and reduce computation, the proposed algorithm also applies Kalman filter to the changing rate of the coordinate transformation parameters, where the powers of the noise models in the filter are estimated online, so as to effectively predict the geometric status of the target in the next frame and set the initial searching point accordingly. A large number of qualitative experiments based on real-world test sequences and quantitative experiments based on synthetic test sequences show that the proposed algorithm considerably promotes the robustness and accuracy of visual object tracking.