Automatic Selection Window For Object Tracking Algorithm

Video object tracking is the critical technology of computer vision and the hotspot of the certain research domain. As a very complicated problem, object tracking problem relates to many aspects of computer vision research. In recent years, object tracking is used in many applications such as surveillance, perceptual user interfaces, object-based video compression, and so on. Among the various tracking algorithms, Mean-Shift tracking algorithm has become popular due to its simplicity, efficiency and good performance.This thesis focuses on Mean-Shift tracking algorithm, which is a modeling mechanism based on statistical probability density function. In the tracking process, the target area and the corresponding histogram are usually established in the first frame. In the subsequent frames, the best target candidate areas are searched iteratively by Mean-Shift algorithm based on the Bhattacharyya similarity function. Mean-Shift tracking algorithm can be performed in real time and is robust for partial occlusion and distortion. However, the traditional fixed bandwidth Mean-Shift tracking algorithm can not have an effective tracking for changes in targets.In order to improve this shortcoming of Mean-Shift algorithm,an novel method was proposed that was multi-scale space theory combined with UKF filter. UKF filter was introduced to predict the information in the tracking window which was calculated by the multi-scale space theory. Then, the proportion of the target image area was got by the modified information. Finally,it was implemented by the combination of the Mean-Shift Tracking algorithm and the UKF filter to track targets. The tracking experiments confirmed the effectiveness of the improved algorithm.