With the rapid development of digital video technique and Internet technique, computer vision attracts much attentions in recent years and develops to be one of the most important branches in the information science field. Visual tracking is the foundation of many high-level visual applications, which directly influences the performance of action recognition, video analysis, human-interaction, etc. Recently, although numerous researches have been done in the visual tracking field, the information of the target in motion is not fully exploited. In this thesis, we focus on visual tracking problem, including target representation, target appearance learning, and data association to improve the performance of the tracker. Graph based model can describe the appearance of the objects and the relationships between different objects effectively. This thesis focuses on using the graph model to solve difficult problems existing in visual tracking field. The main contributions of the thesis are summarized as follows. 1. Most of previous methods focus on the variations of the target appearance only, while ignoring the relationships between the target and its surroundings, which is one of the main problems weakening robustness of trackers in unconstrained environments. To solve that problem, we propose a spatio-temporal context model based tracker, which incorporates both the variations of the target appearance and the relationships between the target and its surroundings. A subspace model is used to describe the appearance of the target, which is updated with another subspace constructed by several sequential positive samples. Meanwhile, we exploit numerous weak contextual supports around the target to form a strong supporting field to improve the tracking performance. 2. To solve the problems existing in the previous bounding box or part based methods that are ineffective in handling large deformation and occlusion challenges, we proposed a dynamic structure graph based tracker, which use the geometric structure graph of the target to describe the deformation of the target to enhance the robustness. Each node of the graph corresponds to the target part, while the edges describe the relationships between the parts. Thus, the tracking problem is formulated as the graph matching problem between the target geometric structure graph and the candidate graph, which can be solved by spectral matching algorithm effectively. 3. In some applications, we are required to...
修改评论