Research On Target Tracking Based-on Particle Filter Algorithm

As the key technology of the computer vision, video target tracking become one of theimportant research topics in computer vision area at present. In a number of areas, such asvideo surveillance, image coding, image indexing, medical diagnostics, human-computerinteraction and imaging guidance, video target tracking has a broad application prospect. Atpresent, numerous approaches for video target tracking have been proposed. However, due tothe complexity of the video tracking, existed video tracking algorithms still have manyproblems. Based on Bayesian estimation theory, particle filter could provide a general andoptimal solution to nonlinear/non-Gaussian dynamic state estimation problems, so widelyused in the area of target tracking. This paper mainly studies the target tracking based onparticle filter. The major works in this thesis include.1. The traditional particle filter estimation method is improved. A improved particlefilter algorithm based on weight selected is proposed. The method uses the particle with largerweight for state estimation, so that this method is better than traditional methods. A newweight fused with distance factor is proposed, this could make up for the deficiencies of onlyusing the color histogram as the target feature. Experiments show that the method couldimprove the accuracy of the state estimation.2. A research for adaptively adjusting the tracking window size based on particlefiltering tracking frameworks is made, and a self-updating tracking window method isproposed. This method adopts different strategies for single color target and non-single colortarget to update the tracking window according to the number and distribution of the particlewith larger weight. Experiments show that the method could adaptively adjust the trackingwindow size with the scale of target changing.