Research Of Target Tracking Based On Local Information

Automatic target tracking is the basic and key technology in many fields, such as intelligent monitoring, weapon guiding, intelligent transportation system, video compression, image searching and medical image system. It is also one of the hot and difficult subjects in the field of computer vision. The algorithm for tracking targets in complex moving background is discussed here. The issues which affect the algorithm's stability are in deep investigation especially, for example, the shelter of the target, the change of the size in the target and so on. The stability of the tracking algorithm would be decreased by many problems, especially when the target is occluded or target size is changed. In order to solve these two problems, a new algorithm based on the local information of the target is proposed in this paper, and different local information is used in different period of the tracking process. In this paper, the local information is relative to the whole information of the target, and the local information can be one pixel or the whole area of the target in the extreme environment. In this thesis, works are done as follows:1) During tracking process, the occlusion is always happened from the edge of the target. Based on this fact, the normalized cross-correlation algorithm is improved in this paper. In stead of dividing the template into several parts and assigning a fixed threshold to each part, the new algorithm, named as weighted normalized cross-correlation algorithm, set different thresholds2) In order to track the target in complex and moving background, a new algorithm using the target's local information is proposed, which can update the template adaptively. When the template matching failed, the segmentation for new target area was introduced. Therefore, the concrete reason for template matching failure could be concluded according to the similarity comparison between the newtarget area and the template, and different strategies would be used to solve the problem. In this way, the target can be tracked even when it is occluded severely or its size changes.3) In order to improve the efficiency of the algorithm, the kalman filter was used to forecast the target's trajectory in the process of tracking. So the template matching carried out in a smaller area and the algorithm's efficiency was improved. 4) The strategy to solve the occlusion problem is put forward in this thesis. First, the occluded position is judged, and then the left area will be used to track the target. When the occluded area occupies more than 90% of the whole target, the target will be tracked by the kalman filter, which can forecast the comparatively accurate trajectoryof the target. And when the occluded area occupies less than 5% of the whole target, the algorithm exits the occlusion processing. The experimental results proved the stability and efficiency of this new strategy.