Research Of Video Object Tracking Algorithms Based On Subspace Representation

Visual tracking has long been an important topic in computer vision field, especially for application of surveillance, vehicle navigation and human computer interface. Although many tracking methods have been proposed, it remains a challenging problem due to factors such as partial occlusions, illumination changes, pose changes, back-ground clutter and viewpoint variation.Video sequences are selected randomly in the paper as the research object, and the main work is as follows:(1) Summarizing the research background and present situation of object tracking, introducing the present research situation of object tracking, and analyzing classification, technical difficulties and performance evaluation of object tracking algorithm.(2) Introducing the basic theory of object tracking, Principal component analysis, kahnan filter, particle filter and incremental subspace method.(3)Because of the poor efficiency and effectiveness of current visual tracking algorithms, this paper proposes a real-time objective tracking algorithm based on subspace learning. Under the framework of particle filtering, this paper uses the incremental PCA subspace method to learn an orthogonal subspace, and then get the linear representation of target appearance. In order to avoid the tracking drift produced by complicated interference, such as occlusions, motion blur and so on, this paper builds an observation model and a template update scheme, which consider the complicated interference especially occlusions, to solve the drift problem of the traditional observation model based on minimum mean square error. The experimental results show that the algorithm in complicated conditions can be well implemented compared with several state-of-the-art algorithms.(4) In order to represent object more effectively, we propose a new object tracking algorithm based on square template and two-stage method. Considering some partial occlusion problems, we use the orthogonal subspace and square template to linearly represent the target appearance. In tracking stage, we adopt a two stage sampling method for object tracking, which increased the real-time performance effectively.Finally, we summarize the research contents, analysis the results and performance of research, do some reasonable suggestions for improving and perfection, and making the discussion prospect of development in the future of object tracking.