Occlusion Problem In Target Tracking

Moving target tracking is a crucial subject in computer vision, and it has very important practical value and wide developmental prosperity in military visual missile guidance, robot vision navigation, industrial product detection, medical diagnosing, traffic surveillance, etc. Target tracking is a challenging subject because of the variety of the target feature and the complexity of the background, especially the occlusion problem in the course of tracking is the crucial restrictive factor to the robustness of tracking algorithm.This thesis mainly aims to study and solve occlusion problem in target tracking. It detailedly discusses the occlusion handling methods in several algorithms such as intensity-correlation-matching and multi-block-voting, mean shift, particle filtering, etc, making effective improvement to increase the robustness of those algorithms under occlusion. Also it studies multi-target tracking based on moving detection, in which the occlusion problem is analyzed and handled. The contributions are summarized as follows:1,To the occlusion problem in rigid moving target tracking, we propose the tracking algorithm of combining intensity-correlation-matching with multi-block-voting. Firstly, the target is adaptively divided, making each block with distinct feature to increase the accuracy degree of occlusion estimating and block matching. Then the occlusion region is estimated accurately through the blocks. Under occlusion use the reminder unoccluded blocks to match and vote, and the both results of the two algorithms are valued and integrated by the Kalman prediction. Under full occlusion regard the Kalman prediction as the real state of the target to track. The algorithm can track the rigid target accurately and real-time, but has limitation to the tracking of non-rigid target which is small and deformed greatly.2,Aiming at the more general tracking of rigid and non-rigid targets in complicated background especially under occlusion, the tracking algorithm of combining mean shift and paticle filtering is proposed. Firstly, the mean shift algorithm is improved, making the target density peak and increasing its immunity to...