Research On Multi - Template Target Tracking Algorithm Based On MEAN SHIFT

The moving object tracking technology has always been one of the most important research topics in the field of computer vision. Especially in recent years, this technology with fast developing, has been widely used in military, transportation, education or other field. But object sudden change and occlusion is still the focus and difficulty in this field. In this paper, the theoretical development, application and current difficulties of moving object tracking technology have been studied and analyzed, and the improved algorithm based on the existing algorithms is proposed.First of all, because of the advantages of little calculation, fast convergence speed, and being close to real-time tracking, the traditional target tracking algorithm based on MEAN SHIFT has received much attention of researchers. This paper gives the analysis of the object tracking based on MEAN SHIFT tracking algorithm, and summarize the existing problems of this algorithm:(1) if the tracking of the frames ahead is not accurate, the tracking of the following frames will also deviated from the true object position gradually; (2) when the serious frame loss and sudden changing of appearance of object happening, if there are big differences in the color of object between current frame and ahead, the algorithm can hardly recognize the object. (3) when the object has been serious or completely occlusion, accurate tracking will be hard. This paper create and maintain a diverse of template library providing more abundant description information to improve the tacking performance. The data used in the experiment includes videos made in laboratory and international open test data. Experimental results show that the new algorithm can track the moving object better in above situations, even keep a good robustness when completely object occlusion take place.Secondly, considering that the core of improved multi-temple algorithm is still the tracking algorithm based on MEAN SHIFT, which largely dependent on color information and lack spatial information, and because of SIFT descriptors which has the excellent characteristics of being invariant to scale and rotation and light changing, this paper add the SIFT feature space to multi-temple algorithm. When the algorithm can not track the object using the color information, we locate the object by the SIFT feature, which reduce the dependence on color information. The experimental results show that robustness has been further improved.