The Research Base On Mean-shift Tracking Algorithm

With the rapid development of intelligent computing technology, computer vision which gains attention from the security and researches has become the frontier of the IT industry and high technology field. However, as an important part of computer vision, visual tracking is trying to get the location of the moving subject interested from every frame of image sequence, which means to mark the object that is tracked from all the frames in a video.Mean-shift method is a rapid estimation way in a gradient direction. This method is originally used in pattern recognition for cluster analysis, but now it is widely used in target tracking.Traditional mean-shift algorithm mainly has some important things that can be researched:choosing the reasonable target characteristics; choosing the appropriate similarity measure model; the problem of scale adaptive about algorithm; the problem of algorithm converges to a local extreme.In some complex situations, the process of tracking may raise some questions like target occlusion, light effect and the changing of the object’s shape. This thesis is focus on these problems, trying to make a progress on the basis of the traditional mean-shift algorithm. First, the traditional mean-shift algorithm only deals with the color information of the target and ignores the light effect, which usually causing the failure of the target. This thesis brings the description of target’s texture feature into consideration, thus can increase robustness of light scenes. Secondly, as to the the problem of target occlusion, this thesis raises an opinion that in the traditional mean-shift algorithm, in the process of search for the best point by iterative searching,search field should be increased without losing real-time requirements. At the same time, finding out the best point of convergence by using texture features. Thirdly, with regard to the problem that the mix of many characters may increase the characteristics dimensions, this thesis has already set the character weight of the target, which not only reduce the characteristic dimensions, but also can use each characteristics of the target according to the the complex circumstances. What’s more, on the basis of the problems solved before, this thesis changes the traditional B coefficient matching measure, uses the mentioned EMD model, aiming at solving the adaptation of complex situation and the real time of algorithm.Through the comparison and demonstration of experiment simulation, our algorithm can track the target accurately. And when compared with the traditional algorithm and tracing algorithm based on single feature, the new algorithm has good advantages on robustness and real time.