Research On Moving Target Tracking Algorithms Under Video Surveillance Environment

With the development of computer and monitoring equipment’s resolution, the research of intelligent video surveillance systems is getting more and more interest and attention as a hot research in the field of computer vision. How to completely achieve the liberation of the human in video surveillance system is a common goal of all researchers. The process of video sequences in intelligent video surveillance system includes image preprocessing, target detection and classification, target tracking and target behavior understanding. Target tracking is an important technology in the entire process. Target tracking is above the target detection, and it can provide assistance for higher-level information processing. Target tracking algorithm is the main topic of this thesis.The background and foreground are complex and changeable under video surveillance environment. So finding an algorithm which can adapt to all situations is difficult. But there have been some algorithms which can solve part of the problems. The classification of tracking algorithms is introduced in this thesis. Some mainstream tracking algorithms, such as Kalman filter, particle filter, MCMC particle filter, RJ-MCMC particle filter and Mean Shift are described in detail. The advantages and disadvantages of these algorithms are analyzed and compared.A fusion algorithm that is the combination of Kalman filter and Mean Shift is introduced. The tracking performance of Mean Shift algorithm is not satisfactory. Tracking will fail if the target’s speed is high or the target is occluded. The fusion algorithm does not need that there are overlaps between target areas of two adjacent images. It can be seen from the simulation results that, the fusion algorithm has good performance. In addition, it can track the target which has occlusion through the judgement of whether occlusion occurred. Inspiration for the following research is provided by the idea of this fusion algorithm.AMCMC, an improved MCMC algorithm, is introduced. Aiming at the problem that MCMC can not track variable number of targets and that the computational complexity of RJ-MCMC is high, it is proposed that the detection of the entering and leaving targets is independent of the MCMC sampling.It is shown by the experiment results that large number of particles needed to be set in AMCMC algorithm to ensure Markov chain’s convergence. This is because many bad proposed samples are discarded. Based on the Mean Shift’s characteristics that it can quickly find the peak, the bad proposed samples are optimized by Mean Shift first, and then the optimal states are judged to accept or refuse. The utilization of the particles is thus greatly improved.Experiment results show that, the optimized algorithm can track targets better than the original when smaller number of particles are set.The AMCMC’s ability to deal with object interaction and occlusion is poor because of the defect of color model. So Kalman prediction is combined with AMCMC algorithm. The result of Kalman prediction is adopted when the targets are severely blocked. It is shown by the experiment results that the problem of severe occlusion can be dealt with well by the improved algorithm.