Video Target Tracking Based On MCMC Partical Filtering

Moving targets tracking in video sequences is a new research subject and has become more and more active recently. It contains many technologies from different subjects such as computer vision, pattern recognition, artificial intelligence and even some knowledge from biology. So it is an interdisciplinary subject. Precisely tracking targets is the prerequisite of recognition and decision and the basis of advanced computer vision. As a primary subject, targets tracking can be found in many areas from daily life to military. Especially it is used extensively in monitoring systems and traffic systems.Through these years of effort, many methods for tracking targets have been put forwarded, and practically applied. Common tracking algorithms contain Kalman filter, Mean shift algorithm, particle filter and the algorithm based on particle filter: MCMC particle filter, RJMCMC particle filter and AMCMC particle filter. These algorithms can perform well in tracking objects under some conditions. But in dynamic scenes, it is very hard to get the satisfactory results because the movement of background affects the normal performance of the algorithms. Especially under front or back camera translation movement, many problems about tracking targets appear. Fortunately, we noticed that a position called FOE also exists in such case. The movement of a camera causes the optical flow field of background to merge in FOE. Because FOE can represent the direction of camera movement, we can estimate the position of objects with the help of FOE. In this way, by combining MCMC and FOE, satisfactory result is got after the problem caused by camera movement is solved.MCMC algorithm and the improved MCMC algorithm provide us some useful methods to track targets, and it can track objects steadily in diffferent scenarios. While it still has some problems which remain to be solved, high computational complexity limilts its application. For example, Metropolis Hastings algorithm is a common method to carry out the MCMC Particle Filter. The key reason of high time cost is that M-H usually needs hundreds of sample processing. And inefficient sample is decided by the M-H itself. In order to satisfy the requirement of real-time application, we studied the M-H algorithm deeply and an improved algorithm based on M-H is worked out in this paper. Making use of the greedy algorithm, our algorithm adopts piecewise kernel functions to produce samples to deal with this problem and obtains satisfactory results.Following the tracking methods based on parametric estimation, we also studied an algorithm based on nonparametric estimation called Mean shift. Different from parametric estimation, Mean shift does not need any other parameters once the target is defined. And it also can compute the gradient of probability distribution. So Mean shift is combined with MCMC to optimize paticle and better results are achieved.