Research On Sequential Monte Carlo Method With Applications In Visual Tracking

As a challenging research topic in computer vision, visual tracking has a lot of potential applications in security surveillance, intelligent traffic system, video compression and index, etc. Image sequences are the input of visual tracking system, and the system returns some attributes (size, position, velocity and so on) of the object. Ideally, all the information should be outputted in real-time and accurately. But in real world, it is very hard to get the ideal output because of noises. The fundamental building block of a tracking syetem is a filter for recursive target state estimation. Recently there has been a surge of interest in nonlinear and non-Gaussian filtering algorithms. In this thesis, we present a visual tracking algorithm which can resist noises in real world.First, some filtering algorithms such as the Kalman filtering, the extended Kalman filtering and the Unscented Kalman filtering were reviewed in this thesis, and the main focus were the tools of sequential Monte Carlo estimation, reffered to as particle filter. Secondly, the Unscented transformation strategies in the Unscented transformation and the resampling strategis in particle filter were studied in this thesis, and an algorithm called Unscented kalman particle filtering algorithm based on minimal skew sampling was proposed. The proposed sampling strategies were then used in the UKF algorithm by which a proposal distribution is generated and draws samples from it. By doing this, problems such as particle degeneracy problem which were caused by the general particle filter using transition prior density function as proposal distribution were solved. At last, a new visual tracking algorithm, whose frame is particle filter, was proposed. The color information and motion information would be fused in our algorithm. The theory analysis and simulation results show that the improved UPF algorithm improves the stability and accuracy of filtering, and the operating efficiency increases by 30%; the new visual tracking algorithm can improve the robustness and the accuracy.