Head Detection And Tracking In Video Surveillance

In some specific scene, such as the building corridor, the school library or other public places, the monitoring system is used more and more widely. But most of the monitoring systems are still semi-artificial style, and these systems cost a large count of manpower and material resources, which is a great drawback for the popularity of video surveillance in public places. Therefore, to research more intelligent visual surveillance system has become the main direction in this field. This paper mainly researches on the detection and tracking of the multi moving pedestrian targets in the video surveillance system. The main work includes the following aspects:In practice, the surveillance camera monitors the surveillance scene statically during most of the time. Thus, the background of the region is relatively static. Based on this characteristic, this paper does the pretreatment on the surveillance video at the beginning, by adopting Gaussians Mixture Model (GMM) extract the moving foreground. There are two advantages of the pretreatment:Firstly, it could narrow the detection area, speed up the detection process and reduce the computational complexity. Secondly, the impact of background on the test results can be eliminated effectively, while the false detection rate is reduced. In this paper, an adaptive update coefficient for Gaussians Mixture Model is proposed to further improve the quality of the extracted foreground.As the reason that most of the surveillance videos are captured from a downward-slope view, thus, the head compared with the full-body, has much lower probability to be completely blocked by each other in the dense crowd. Therefore, after the foreground extraction processing on the video sequence, a cascade AdaBoost classifier is used to search the head in the moving areas. This paper introduces the Forward Feature Select (FFS) algorithm to reduce the training time. The final head classifier can be achieved after the sample selection and feature calculation. Then, the human head detection is realized by using the trained classifier.Last not the least, the head in the video does not has enough characteristic information than the body to be used in the multi-target tracking, which means the issue that there are blocked between each other can not be solved by simply using the RGB image based methods. Hence, in this paper, the depth information is taken as one of the features of head. Depth information corresponded to each detected head and pedestrian, which is obtained by Kinect camera, can be tracked and estimated using the cost function, as well as location and speed information. Experiments show that the method proposed in this paper can not only meet the actual requirements in real-time, but also overcome the occlusion situation, especially the serious occluded situation. Actually, the occluded multiple targets can be tracked effectively by the proposed approach.