The Research And Implementation Of Target Tracking Algorithm Based On Mean-shift

The automatic target detection and tracking technology based on video images is an important issue which is in urgent need of solving in accurate strike weapon systems, and also it is a research focus in computer vision. Mean-shift is a non-parametric estimation of the density gradient, which was first used for cluster analysis in pattern recognition, and has been widely used in target tracking in recent years.The algorithm of Mean-shift is mainly used for target tracking, and the capture process mainly relies on artificial aids or other methods. The paper used improved accumulative difference images (ADI) to locate moving targets roughly, which could eliminate the impact of slow changes of background. Then segmentation based on double-window histogram was used to obtain the accurate scale and location of the target, which effectively overcame the difficulty of incomplete segmentation especially when the target has complex texture or similar gray to the background.The paper focused on several major factors: Firstly, the selection of target template is directly related to convergence speed and location accuracy. Using background suppression histogram as the template, center weighted histogram as a candidate, not only accelerated the convergence, but also enhanced anti-interference ability. Secondly, the bandwidth updating of kernel function is so important that only when bandwidth and scale of the target are matched can make tracking more accurate and much steadier. The paper used edge weighted histogram with background suppression histogram to compute Bhattacharyya coefficient. According to the coefficient changes between different kernel bandwidth, combined with test method, it achieved good effect on tracking target with variant scale. Then shift to the problem of template update. It requires that the template can reflect latest features of the target, while at the same time retain inherent characteristics. The paper achieved this goal by comparing Bhattacharyya coefficients. And the final problem is the prediction mechanism. A good prediction mechanism can greatly reduce the iteration time and solve the blocking problem. The paper checked whether the target is blocked by the value of Bhattacharyya coefficient, and used linear prediction mechanism to track target steadily.The author designed an embedded video processing platform with DM642, transplanted the tracking algorithm based on mean-shift onto it, and ultimately realized a small electro-optical tracking system by debugging and testing.