Object tracking, an intermediate-level vision part, plays an important role in visual analysis and understanding of object motion. The goal of object tracking is to detect, localize and track moving objects in videos. By visual tracking, we obtain the motion parameters, the poses and the trajectories of objects, which lays a solid foundation for the high-level activity recognition and understanding. Although object tracking is really a hot research topic in recent years, there do exist many theoretical and technical problems, especially in the cases of noise, motion blurring, illumination changes, occlusion etc. In this thesis, we aim at tracking object robustly in visual surveillance application. Two topics are investigated in details: (1) occlusion handling under mobile camera in multiple object tracking; (2) objecting tracking based on transfer learning. Experimental results have demonstrated the effectiveness and robustness of our method. The main contribution of this thesis include following issues: (1) We introduce a block-division based appearance model for occlusion handling in multiple object tracking problem. This appearance model divides the object into blocks at the level of space, models each of the block respectively, thus computes each block's likelihood independently. At the same time, the likelihood of an special candidate is calculated by multiplying each block's likelihood between it and its corresponding model, therefore introducing spatial information implicitly. Furthermore, the block-division based appearance model means importance to occlusion reasoning and appearance updating in the following stage. It is just because of division of appearance that we can treat the part of object which is not occluded and the occluded part discriminatively, which facilitates occlusion handling in the following. (2) We propose a selective appearance model updating strategy to deal with the appearance model updating which results from occlusion between multiple objects. Generally speaking, appearance model should learn the variety of object during tracking. However, if this learning process is still carried out when the object is occluded, it will definitely introduce noises. This type of noises would probably lead to failure of tracking. The selective updating strategy we proposed is based on the block-division appearance model. As the object is divided into blocks, we can obtain information about whether the block is occluded or not. Thus, w...
修改评论