Robust Kernel-Based Tracking with Multiple Subtemplates in Vision Guidance System

The mean shift algorithm has achieved considerable success in target tracking due to its simplicity and robustness. However, the lack of spatial information may result in its failure to get high tracking precision. This might be even worse when the target is scale variant and the sequences are gray-levels. This paper presents a novel multiple subtemplates based tracking algorithm for the terminal guidance application. By applying a separate tracker to each subtemplate, it can handle more complicated situations such as rotation, scaling, and partial coverage of the target. The innovations include: (1) an optimal subtemplates selection algorithm is designed, which ensures that the selected subtemplates maximally represent the information of the entire template while having the least mutual redundancy; (2) based on the serial tracking results and the spatial constraint prior to those subtemplates, a Gaussian weighted voting method is proposed to locate the target center; (3) the optimal scale factor is determined by maximizing the voting results among the scale searching layers, which avoids the complicated threshold setting problem. Experiments on some videos with static scenes show that the proposed method greatly improves the tracking accuracy compared to the original mean shift algorithm.