The Research On Intelligent Multiple Video Targets Tracking Based On Improved Particle Filter

Video target tracking is an important topic in the field of intelligent video processing. The main content is to estimate the attractive targets' motion state or motion trajectory from the video sequences. As the broad application prospects, intelligent video tracking successfully attracted world wide scholars' attention.,Recent years, since the great improvement of computer technology, on the one hand makes large-scale computing speed greatly improve, on the other hand, make the mathematical theory of video tracking be applied practically. All of these makes the intelligent video tracking technology be widely used. Particle filter is a probabilistic tracking method based on Bayesian estimation. Because of the effective treatment of non-linear, non-Gaussian systems, it was more and more used in video tracking. But due to the characteristics of viedo itself and disadvantages of particle filter some inherent problems still need researchers' great efforts. In response to these problems, my main contributions in the thesis are listed as follows:Based on the theoretical framework of particle filter used in video tracking, we summarized the common used state transition models and visual features used to establish likelihood function. We verify the particle filter though simulation experiments, and analysis the failure reasons—degeneration and "sample impoverishment". We combine the particle filter with immune genetic algorithm, and describe how the immune genetic algorithm optimize the particle set. then propose the immune genetic particle filter algorithm of video tracking. At last, we verify the new proposed algorithm by experiment, and gain the conclusion that the new proposed algorithm is better than standard particle filter in correction and robustness.In order to improve the performance of particle filter from the perspective of multi-feature integration, we first summarize several common image texture feature extraction methods, especially the rotating composite wavelet texture. Then we proposed the immune particle filter based on color and wavelet texture in video tracking. At last, though experiments we compare the tracking results of standard particle filter, immune genetic particle filter and the proposed algorithm, and verify that the new proposed algorithm's performance is better.We analysised the framework of multi-target tracking based on particle filter, and point out the strengths and weakness. We introduced particle filter into two different video scene to verify its performance.We analysis the characteristics of most dynamic jump Markov nonlinear systems, and briefly introduce the Kalman filter of discrete-time jump Markov systems with certain noise and the model of discrete-time jump Markov systems with uncertain noise.Then summarize the Bayesian estimation of jump Markov non-linear systems, analysis the state estimation of such system. Since its prediction equation satisfied Fokker-Planck-Kolmogorov Equation (FPKE), and most FPKE are difficult to obtain analytical solution. We use particle filter to gain its numerical solution, and proposed the particle filter framework of jump Markov non-linear systems. And we verify the proposed algorithm's accuracy and effectiveness by multiple targets tracking.At the end, we summarize of content, advantage and the deficiency of the thesis, and narrate further development of the study.