Research On The Theory Of Particle Filtering And Its Applications In Video Processing

With the development of the computer network and communication as well as multimedia technologies, video processing has been widely used in many fields. Research contents of video processing are very rich, including moving target detection, tracking, classification, and behavior understanding, and technologies relate to computer vision, pattern recognition and artificial intelligence, which are very challenge issues. Particle filter can accurately represent the posterior distribution based on measurement and control variables, which has been widely used in non-linear, non-Gaussian systems. From the perspective of particle filtering, this paper carried out work in four areas related to video processing:Firstly, the particle filter theory is studied. Particles degerneration and sample depletion problem are the major issues affecting the performance of particle filter methods. This chapter firstly presents a new iterative particle filter IPF, using the extended Kalman filter and unscented Kalman filter state to estimate system states iteratively. Simulation results show that the accuracy and stability of IPF is superior to other contrast filter algorithms, however, very time consuming. On the basis of analyzing the performance of IPF, an iteratively selective unscented particle filter SIPF is proposed, which can eliminate invalid operation, while maintaining the accuracy and reducing running time. The experimental testing results on two different non-linear/non-Guassian systems have shown that, the accuracy of PF-ISUKF is better than other that of other compared algorithms, while keeping relatively low running time.Secondly, video stabilization is studied. Video stabilization typically includes several parts such as motion estimation, motion compensation and image compensation. Compared to other types of video stabilization algorithms, video stabilization algorithms based on feature points are with higher robustness. However, the accuracy of video stabilization algorithms based on feature points depends on the accuracy of feature point location and matching errors. To solve these issues, this chapter presents a video stabilization algorithm based on particle filter. First, SURF is used to extract feature points in two adjacent frames, followed by the use of RANSAC algorithm excluding local motion vectors. When carrying out motion estimation, firstly feature points are matched based on least squares method to obtain the motion estimation of the initial value; Secondly, based on the initial estimates, PF-ISUKF is used to obtain accurate values of the motion estimation. Experiment results proved the algorithm effectiveness.Thirdly, object contour segmentation based on active contour is studied. Firstly, for the issue that current shape prior based active contour methods can only obtain single segmented region, this chapter proposed a new multi-target segmentation method based on active contour using shape degeneration. In this proposed model, the image firstly is segmented by a non-shape piror active contour method. After that a novel shape degeneration model is proposed to judge the matching situation of the segmented areas and shape priors, and then invalid regions will degenerate automatically. Finally multi-valid-regions have been segmented. Secondly, as curve evolution of active contour models is easy to fall into local optimal, as well as the blindness of results of the active contour segmentation model based on particle filters, this chapter proposed a novel model combining particle filtering and geometric active contour with shape priors. Experiments show that the proposed algorithm can effectively make a video object contour segmentation.Finally, for the issue of the high computational cost of the camshaft based particle filter, this chapter proposed a novel adptive Camshift based particle filter for video object tracking. The proposed method firstly establishs the target state model, and then uses existing tracking results to estimate model parameters. After obtaining the motion state variance, the search area with a minimum cumulative Camshift algorithm is used to dynamically adjust the search window, to update the weight of each particle in the particle filter. Experimental results show that, compared to CGPF, ACGPF does not only can reduce the computation of the algorithm, while the background noise is better able to overcome the influence of the tracking process.