Research On Moving Object Tracking Technology Based On Kalman Filter In Visual Surveillance System

As one of the most important parts in intelligent visual surveillance, moving object tracking has attracted many researchers'attention and becomes a very popular research topic in recent years. Because of illumination changing, occlusion and other complex situations, there are a lot of difficult in achieving robust tracking in an actual system. Designing a robust tracking method is still a challenging work.On one hand, the research in this thesis focus on two major problems of tracking, data association and occlusion handling, and solve the tracking problem in the conditions of illumination changing and occlusion based on Kalman filter. On the other hand, this thesis studies on tracking method in people counting using Kalman Filter. Main contents include:(1) Based on an idea which combine motion and appearance information can overcome the influence of illumination changing, a method which uses both Kalman filter and color histogram is proposed in this thesis to deal with the false tracking in this environment. For the problem of occlusion, this thesis divides occlusion into two categories: static and dynamic occlusion. By dynamically changing covariance of process noise and measurement noise in Kalman filter, this method can maintain the tracking of moving objects before, during, and after static occlusion. In the condition of dynamic occlusion, a merge-split strategy is used to track objects continuously. Experiments described on several test sequences of the open PETS2000 and PETS2001 dataset demonstrate the accurateness of this method is better than other traditional method in the case of illumination changing and occlusion.(2) This thesis presents a machine learning HOG header-shoulder detection method which can solve the problem that traditional methods cannot detect correctly when objects are gathering and waiting in people counting. For positive and negative false in the results of the machine learning HOG header-shoulder detection method, this thesis propose to remove the positive false basing on extracting motion regions, use Kalman filter track people continuously and finally realize the processing of counting. Experimental results show that this method increases the robustness of detection and tracking in people counting.