Research On Mean Shift Algorithm And Its Application In Video Target Tracking

It is an information world now. The rapid development of computer technology has brought great convenience to our work and life. As an interdisciplinary field based on computer and information technology, video target tracking is now widely used in the areas of video surveillance, robot vision, intelligent transportation, national defense construction, aerospace, national economic construction, and so on. The major point of video target tracking is to split the target reasonably, extract target features accurately and identify the target quickly. Meanwhile, the factors of the speed, accuracy and stability of the algorithm are also in consideration. The essence of target tracking is the state estimation problem of automatic control system, that is, the procedure of filtering continuous random noise, observing and estimating the target state, and finally finding out the movement elements of the objects.In the subject research, by means of learning and exploring the issue of target tracking in video surveillance, we have made some progress. Due to the complexity of the background and the video object tracking environment, it is difficult to find an algorithm that can fit all situations. So variety of algorithms are proposed. This thesis describes several classic algorithms in the field of target tracking, that is Mean Shift, Particle Filter and Least Square algorithm. We analyze the advantages and disadvantages of various algorithms, and then proposed different improved methods according to the lack of the algorithm.Mean Shift searches the target from the initial point, which may result in poor performance when the target moves fast. We propose to use Least Square prediction to solve this problem. Least Square is first used to predict the position of the target, and then Mean Shift is then applied to search the target from the predicted position and finally find out the target. This algorithm has a good real-time quality via reducing the distance of the vector from the convergence point in every frame. The improved algorithm is called Least Square Mean Shift.Experimental results show that, the improved algorithm has a better performance in situation of fast moving target or the occurrence of occlusion. Meanwhile, we get the number of iterations and processing time from both the original Mean Shift and our proposed algorithm. It is shown that, the improved algorithm has a better real-time quality.Least Square Mean Shift has smaller calculation amount. And it also has a good real-time quality in the situation of a simple background of the video image. But when the target is sheltered largely, the result is not satisfactory. In contrast, Particle Filter has a good performance in the case of complex background and occlusion. However, the calculation of the Particle Filter is larger and it has a bad real-time quality.According to different advantages and disadvantages of these two algorithms, we proposed a combination algorithm of Least Squares Mean Shift and Particle Filter in target tracking. When the target is not sheltered or has a small proportion of occlusion, we use Least Squares Mean Shift algorithm. Otherwise, when the target is sheltered heavyly, we use Particle Filter algorithm. The improved algorithm has a good performance via switching between two different algorithms according to the variational similarity coefficient. Experimental results show that, the combined algorithm has both better real-time quality and robustness.