Study On Detection And Tracking Of Multiple Human Body Targets In Video Surveillance

Video surveillance is one of the most active researches in computer vision .It has a very widely application. It is mainly composed of object detection, object classification, object tracking and object behavior understanding .In this paper, only the human tracking is studied. The main task of the human tracking is to establish the human body corresponding relations among successive frames. It involves human detection, human body feature description and feature match. In this paper, the latter two aspects are emphatically discussed.In the process of human tracking, two situations are considered. On the one hand, Kalman filter is used to track when human body is not occluded, on another hand, besides position prediction, color feature is also used.How to create human feature is a critical problem in human tracking. In this paper, cluster analysis of ISODATA is used to get the color block of the human body, and each block is regarded as a standard sample set. For the pixels of the corresponding region in the next frame, mahalanobis distance is used to measures whether the pixel is belong to one of the color blocks. According to the number of the pixels in the specific region, whether the region is matched can be judged, furthermore, whether the human body is being tracked accurately can also be judged.When the human body is occluded, the color blocks can not be got accurately from successive frames. To get the tracking steadily, updating the color feature based on prediction is proposed, in the same time, mahalanobis distance threshold also be updated in the same way.Considering more time consuming with the cluster algorithm, normalization of the human body is designed. In this way, not only the time is saved, but also the accuracy of the tracking is improved.