Object Tracking Algorithm Based On SVM

Object tracking is a very important work within the field of computer vision, and it incorporates advanced achievements in different areas. At present the object tracking is widely used to man-machine interactive, video monitoring, military security and so on. The classical mean shift (MS) kernel-based tracking algorithm has the advantages of real-time and accuracy, and it easily was merged into other algorithms. But the candidate template inevitably contains quite a lot of background information in the case of complex background, the information led directly to decline the tracking precision. Meanwhile, MS algorithm lacks the necessary template updating mechanism, so it is prone to track failure in the course of object shelter from the background or the object shape changing.In view of defect of traditional MS algorithm, this paper tries to combine SVM and MS to overcome these problems. Firstly, we use SVM to forecast the pixels including the target in certain limited boundary, it divides the pixels into the target and background points, and then achieves the tracking on the basis of the detection by the help of MS. This method improves the tracking accuracy for complicated background in certain real-time conditions. At the same time, According to the classification results, we establish a new template updating mechanism, and it can effectively prevent the template ill-posed problems. This article mainly gets such results as below:(1) Propose a new tracking algorithm which incorporates SVM and weighted MS object tracking algorithm. Firstly, the algorithm uses SVM classifier based on color character to forecast pixels, and then combines the new weighted MS algorithm which endows with different weight values for the object points and background points to realize the tracking. In the case of the video sequences of complicated background, the new weighted MS algorithm can reduce the interference caused by the background factors, and protrude target features. This method not only reduces the tracking of iterations, but also improves the accuracy.(2) Establish a template updating mechanism according to the classification of SVM, it can update the template in real time, and effectively prevent tracking failure occurred in the situation of object shelter from the background or the object shape changing.(3) In view of the video sequences which are the same color between the targets and background are hard to process. Research on how to employ the co-training theory in multiple SVM training based on different features to realize object tracking.