Research On Object Tracking Using Meah Shift Combined With Kalman Filtering

Moving object tracking in video sequence is an important and necessary technology within the filed of computer vision. Mean Shift is an excellent one of many tracking algorithms. This thesis focuses on the Mean Shift theory, the traditional Mean Shit tracking and improvement measures taken to overcome its weakness that it usually fails in tracking when the target is moving fast or severely hided.Mean Shift chooses the kernel weighted color histogram as the description of target model. It has a good real-time performance and the unimodality of the kernel keeps its robustness when coping with partial occlusions and target rotation. However, Mean Shift has some inherent flaws. For example, it often fails to track a object which is moving very fast or which is severely or even completely occluded, due to a convergence to the nearby background region. In this situation, the background region often has a similar color appearance to the object. To solve the fast moving object tracking problem, this thesis proposed a new effective approach combining Mean Shift and Kalman filter. Also, a method to decide the presence of severe occlusion and a corresponding solution is proposed.Related theory about Mean Shift is introduced, such as, kernel based estimation, interpretation of the Mean Shift concept and proof of the convergence of Mean Shift is discussed. This thesis describes in detail how Mean Shift theory is applied to object tracking. A adaptive scale strategy is proposed to cope with the target scale variation. The experiment results of different video are also shown.To solve the fast moving object tracking problem, a new effective approach combining Mean Shift and Kalman filter is proposed. After a introduction to Kalman filter, the modeling process which takes advantage of its ability of prediction and improves the tracking performance, is discussed in detail. The experiment results of the improved method at a situation of significant occlusion are given. Compared with the traditional Mean Shift and the improved algorithm, we conclude that when coping with rapidly moving target tracking and severe occlusion, the improved method can obtain a continuous and stable tracking results. That is, the ability of the improved algorithm has been significantly improved compared with the traditional Mean Shift.