Object Tracking Algorithm Research Based On Particle Filter

Moving object tracking is one of the most important fields in computer vision. The main task of object tracking is to detect moving object accurate from each image frame, and position those object in the subsequent frames. Then we could get full trajectory of moving object. Moving object tracking is of great practical value in the field of intelligent monitoring, traffic monitoring, human-computer interface and military defense etc.Kalman Filter and Extended Kalman Filter are the most typical filter algorithms in object tracking domain. Kalman Filter could get pretty good filtering performance when system and noise are linear. The Extended Kalman Filter could make the nonlinear system local linearization, and use the principle of Kalman Filter indirectly. But it only adapt to the situation when the filtering error and forecast error are smaller. Because vision procedure is actually nonlinear and non-Gaussian, using the corresponding techniques become a kind of important research tend in object tracking fields. Particle Filter has been proposed based on the Monte Carlo method. Particle Filter algorithm is the main content of this paper and it has distinguishing features in the nonlinear and non-Gaussian system.The research content and innovation are as follows:(1) The author analyzes the basic theory of vision object tracking at first, and introduces the basic principles of Particle Filter detailed. (2) The moving object is susceptible to the background interference when tracking under the natural environment. Then an improved color distribution model is proposed. The proposed method is based on Particle Filter, integrated with improved color distribution model in the measurement model. The improved color distribution model could describe the feature of the object better. So we model each pixel of the object areas using the improved color distribution model. (3) At first we analyze the problem of occlusion when tracking. Then an improved occlusion detecting method is proposed, and the improved method could detect exactly when the occlusion happens and when the occlusion is over. When the object under occlusion, we update the object template in the time using the strategy that the paper introduction. Meanwhile, we recover the update of object template when the occlusion is over. (4) Compared to the kinds of similarity measurement methods, the paper use Bhattacharyya distance to evaluate the likelihood.(5) Kalman prediction algorithm is used to track the moving object when full occlusion happens.Finally, the paper's method is applied to the video image of natural environment. The experimental results show that the paper's method could hand the occlusion effectively and track the moving object very well when the object under occlusion. And it also show that our method achieve a high robustness.