Research On Video Object Tracking Based On Mean Shift Algorithm

Object tracking is the process that to search the most similar parts to targets in video sequences, which is an important research topic in computer vision. Object tracking technology has been widely used in many fields such as smart surveillance, human-computer interaction, object based video compression, and national defense industry. Among the various tracking algorithms, mean shift has become popular due to its simplicity, efficiency, better real-time and good performance.Mean shift is a gradient increased mechanism based on statistical probability density function. It deals with the object matching problem between two successive frames and runs both fast and effective. In the tracking process, the target area and the corresponding kernel color histogram are usually established in the first frame by user. In the subsequent frames, the best target candidate areas are searched iteratively by mean shift algorithm based on the Bhattacharyya similarity function.However, mean shift use color to model target, when the target was partially or completely occlusion, or there is no object overlap between two sequence frames, the algorithm will converge to a local optimum, leading to fail in tracking object. For the defects of mean shift, mean shift generally should not be a single tracking algorithm, but combined with other tracking algorithms.In this paper, it combines the mean shift with the particle filter which is a non-linear, non-gaussian and multi-peak tracking algorithm. Mean shift and particle filter samples some particles respectively and determine the final state of target. If particles sampled by mean shift are closer to final state, the particles will replace equal particles sampled by particle filter, otherwise discard the particles. The improved algorithm not only overcome the defect of mean shift when target was partially or completely occlusion, but also introduces current observation to particle filter, and realizes the complementary advantages of both. In addition, for the size of target will enlarge or shrink over time in the process of tracking, the adaptive window update strategy was introduced. That is, the search window and tracking window will do some adjustments with exchange of target size. This ensures the validity of the extracted features to enhance the robustness of target tracking. Various representative video sequences are utilized in experiments to analyze and verify the performance of improved algorithm. Experimental results show that improved algorithm not only maintain the high robustness and better performance of resisting occlusion, but also the lower computational cost meets the need of real-time disposal.