Research On Occlusion-Adaptive Object Tracking Algorithm

Tracking moving targets in a video is one of the most challenging issues in the field of computer vision, which has wide applications such as vision navigation, security surveillance and so on. Developing robust tracking algorithms has also important theoretical significances. Considering the complexity of environment in reality and multi-status of target, it is a challenging task to extract the target features fast and accurately, identify and track the target correctly. And, the occlusion inference in target tracking, due to its theoretical complexty and importance influence on the performance of the tracker, has become the hotspot research problem in the area of target tracking.There are three mainstream tracking algorithms at present: Meanshift, Kalman Filter and Particle Filter developed basing on the latter. However, the Meanshift is liable to converge to local extreme point, which limits the robustness of the algorithm under occlusions despite it can support real-time processing. And the Kalman Filter requires Gaussian and linear system to support its optimistic estimation, which always impossible in reality. Particle filters is very effective in estimating the state in nonlinear and non-Gaussian dynamic systems in a recursive manner, and shows better performance in complex environment compared with Kalman Filter, and achieves multiple applications in target tracking field.This paper focuses on the occlusion inference problem in target tracking. We discussed the processes of occlusion and corresponding solutions including the judegement condition of occlusion, classical occlusion inference algorithms and their advantages and disadvantages. Based on theories discussed above, an innovative occlusion inference algorithm has been proposed. The Particle Filiter is the framework for the algorithm, the color histogram is extracted as the target feature, and the idea of region growing is introduced in particle relocating mechanism. We developed occlusion detection, particle relocation and adaptive number of particles. When occlusion happens, the particle will stop searching and relocate around the contour of the occlusions. This method makes the particle can grasp the target as soon as the occlusion ends, avoids the potential situations like incorrect searching zone or disturbed by the target-similar information from background, which assures the robustness of algorithm. A variety of experiments verify the effectiveness of the proposed algorithms when dealing with partial and full occlusion. In some extent, the algorithm achieves robust and real-time requirements,it also improves the stability and accuracy of tracking. The work presented in this paper opened a new research direction for occlusion inference in target tracking.