Occluded Multi-target Tracking Based On Local Analysis Of G&E

Target tracking is one of the kernel issues in the field of computer vision research, which has a wide range of applications in many fields, such as security, intelligent transportation, human-computer interaction, having great value and significance. Since the information of image would lost in the process of changing the image from the three-dimensional space to the two-dimensional space, image itself carries a lot of noise, and the complex environment and the move way of the target is changeable, etc. Thence the research of target tracking technology becomes very difficult. Particularly, when occlusion occurs between the targets, the information loss makes the tracking algorithm more difficult.In order to solve the cross occlusion problem in the multi-target tracking, we establish the block model based on G&E features by analyzing common features and the movement of target. Next establishing an overall algorithm framework based on the reasonable multi-target management framework to achieve the multi-target tracking algorithm based on G&E features. Main contents are as follows:(1) Extracting the moving target in the video by Gaussian Mixture Model and Gaussian Mixture Shadow Model, then establishing a multi-target management method for the case of occlusion based on the understanding of single target tracking.(2) After analyzing the target and occlusion in the video, we select the appropriate block model and put forward the G&E features, i.e. E represents the gradient direction of the pixel which is on the band of strong gradient, and G is the gray of inner point get by going along the gradient direction to the target center. And we establish the block model based on the G&E features.(3) Establishing a data link between the detection result and the target in the frame, learning features, developing appropriate matching rules, and designing program to achieve our algorithm proposed in this paper.In this paper, against feature instability and other issues in target tracking, combining with analysis of this problem, we put forward G&E features, screen out valid information by kalman prediction, and establish a rational algorithm framework to achieve the algorithm which we proposed. Finally, we carry out experiments on actual videos and results show that: the algorithm can perform effectively on the cross occlusion.