Research And Application On Target Tracking Algorithms Based On Mean Shift

With the continuous improvement of computer processing power in recent years, video tracking, which is based on objective analysis and processing, has become a new hot spot areas of computer vision research. Target tracking technology in video surveillance, medical imaging, military defense and other fields have a wide range of applications.Mean Shift tracking algorithm which has low computational complexity, simple iterative process and less prior knowledge of study attracts many research scholars’ attention. The research on the improved Mean Shift algorithm is endless. The paper firstly elaborates the Mean Shift algorithm and points out the existed two shortcomings such as the kernel bandwidth keeps stable and color histograms are too single. Then, the thesis makes a comparison between the popular improved algorithms based on Mean Shift. Finally, the article proposes algorithms to solve the aforesaid two shortcomings.In the process of object tracking, the kernel bandwidth keeps stable in the Mean Shift algorithm so that it may lose object caused by scale changing. In order to solve the problem, this paper proposes the Mean Shift algorithm with regularize Hellinger distance. In the algorithm, two term regularizations are defined to modify the kernel bandwidth, which can not only completely describe the target information, but reduce the mixed background information. However, the improved algorithm may lead to wrong scale estimation and tracking failure in completely blocked target and more complex background. To solve the difficulties further, the article proposes the Mean Shift algorithm with backward consistency check, which use reverse tracking from position obtained by forward tracking and validating the estimated scaling factors from step t-1 to t and t to t-1. This validation ensures that in the presence of background clutter the scale estimation does not “explode” and enables tracker to recover. Experimental results show the effectiveness of these two proposed algorithms.Color histogram is commonly used as an object tracking feature of the Mean Shift algorithm, which is likely to be affected by similar backgrounds and lost the object. To solve this problem, the Mean Shift algorithm based on improved particle swarm optimization is proposed, which uses improved particle swarm optimization algorithm to optimize the different features combination to generate the best tracking feature. The experiments show that the method represented in this paper achieves better results.