Particle Filter Based Visual Tracking Algorithm Research

As a major issue in Computer Vision filed, Visual tracking has its broad application prospect in military, monitoring, product control, bio-technology, etc. With soaring demands, visual tracking technology, especially its core algorithms, has been developed rapidly. In various algorithms, particle filter has received special attention because of its unique advantages in dealing with nonlinear problems. Along with rapid descending of computation cost, particle-filter-based algorithm gradually becomes the mainstream. In this dissertation, particle filter is analyzed thoroughly and comprehensively, including some optimization and adjustment to apply it into real application. The main contents include:(1) Firstly, particle filter is successfully implemented in a tracking algorithm which uses RGB color space. However, this elementary algorithm has weakness when the object is blocked, similar interference occurs and the scale of the object is changed. In order to overcome these weaknesses, block determination method, multi-part color model and scale-adaptive particles are employed, which dramatically improve the robustness of tracking algorithm.(2) Traditionally, particle filter uses fixed noise to propagate particles. In this dissertation, an adaptive particle sampling method is presented so that noise can be adjusted in real-time and particles can be better propagated. The employment of this method is aiming for a better particle sample allocation.(3) The sampling part is the major concern of this dissertation. Auxiliary particle filter, Mean-shift embedded particle filter and Mean-shift embedded auxiliary particle filter are introduced to optimize particle allocation. Apart from these exiting methods, a new sampling method, based on auxiliary particle filer and stratified sampling method, is presented in the later part of this dissertation, which leads to better performance in robustness, accuracy and computation efficiency.(4) An experiment platform, constructed by DirectShow and VC++ technology, is used to test all the given algorithms. Extensive video scenarios are performed in this platform to testify the algorithms. In the end of every experiment, detailed analysis can be found to help readers to get a better understanding of the algorithm.