The Research On Target Tracking Algorithm Based On Particle Filter

Video target tracking is a core issue in computer vision. It has been widely applied in thefield of visual surveillance, human-computer interaction, robot visual navigation. A usefultracking algorithm can tackle with the difficulties of complex background, light, interferenceof similar object, target occlusion and so on. For this reason, researchers have proposed manytracking algorithms, but to design a robust universal tracking algorithm is still a challengingtask. Therefore, this article mainly aims at researching and improving object tracking basedon the particle filter framework.1. Introduces the basic theory of Bayesian state estimation in detail, leads algorithmsbased on Monte Carlo particle filter. And then describes the general process of sequentialimportance sampling and standard particle filter algorithm. Analyses the system model andobservation model of the particle filter algorithm. Analyses the advantages of particle filteralgorithm in target tracking based on experiments.2. Improves object representation method in the particle filter algorithm.The using ofnuclear-based gradient direction color histogram to make better use of the colorcharacteristics. At the same time, a combination of nuclear-based gradient direction histogramcan represent the target. Meanwhile, use a new method to mix multi-feature adaptively, it canadaptively weight different characteristics in different scenes. In the process of tracking, it isnecessary to update the model of the target, and adjust the speed of the target template updatethrough a learning rate. If the target occluded, we should stop the update of the template.3. For Target variable speed, this paper proposed a hybrid motion model. This method notonly is able to track the normal movement of the target, but also be able to capture thevariable-speed goals, however the calculation is simple and fast. And for target occlusion,judge the occlusion time and degree using another occlusion factor. If the target is partiallyocclusion, use sub-block to vote the result, to track persistently by using unblock area. If itoccluded seriously, then stop the update of the target and keep the target information, attemptto regain it relying on detection techniques. 4.Combines the improved particle filter algorithm with the visual background extractionalgorithm for multi-target tracking. This method can quickly find new targets and recovertargets that completely blocked. Relying on detection sets and track sets, it can achievegeneral inference based on the correlation matrix, at the same time, it can well achieve thetarget association,which has improved the accuracy and robustness of multi-target tracking.