Object tracking is one of the hottest research topics in the field of computer vision. It has broad application prospects in video surveillance, human-computer interface, intelligent transportation, and video retrieval, etc. Although, with the efforts on theresearch of object tracking for several decades, the technology of object tracking has made great progress, it is still difficult to track an arbitrary object in complex environments. There are still many challenges and difficulties on the tracking under complex scenes, for example, occlusion, clutter background, complex motion and object appearance and shape changes, etc. In this dissertation, we focus on the key techniques of object tracking to solve the problems above. We have proposed some robust tracking methods from the view of object appearance, Bayesian tracking and discriminative tracking. The main contributions of the thesis include the following issues: 1、We propose a visual attention based mean shift tracking method. We combine the KL static attention with motion attention to obtain the object visual attention region. Since object representation using color histogram cannot be adapted to complex scenes, the weight of object representation based on visual attention is calculated for kernel density estimation mean shift tracker. To handle the occlusion, we propose an interactive Bayesian filter to combine the mean shift tracking method, which improve the performance of tracking. 2、We propose a robust fish swarm optimized Bayesian tracking method with Riemannian manifold metric. To solve the problems of covariance Bayesian tracking,we embed the fish swarm algorithm into particle filter, which enables most particles move into the region with high likelihood to avoid both degeneracy and impoverishment problems of particles. The calculation of similarity of object uses Riemannian manifold metric. The object representation is encoded in the proposed methods to handle difficult background and appearance changes. The fragment-based representation effectively utilizes the spatial distribution of object for partial occlusion tracking. All parts are integrated to form a complete framework of fish swarm optimized Bayesian tracking methods with Riemannian manifold metric. 3、We propose a max-confidence boosting learning framework based on object tracking method. The max-confidence boosting replaces the determinate labels with indeterministic labels as well as handing unlabel data, and it can be seen...
修改评论