Occlusion pedestrian tracking is still a difficult problem in video surveillance, while traditional mean shift tracking algorithms fail to track these kinds of targets. Proposed is an improved mean shift tracking approach to solve this problem. Two aspects are improved for the traditional mean shift tracking algorithm. First, occlusion layers are used to represent pedestrian occlusion relation and the non-occlusion part of each pedestrian which is obtained according to occlusion relation is used for the mean shift tracking algorithm. Secondly, the states of the related occlusion pedestrians are gradually adjusted one by one to eliminate the occlusion effect, during the tracking process. The contrast experiment results show that the improved algorithm is real time for well tracking the occlusion pedestrians which cannot be tracked by the traditional mean shift tracking algorithm.