Multi-target Tracking Algorithm For Multi-camera Network Environment

Single-camera visual tracking system has many unresolved issues,including the target occlusion, the limitation of camera field of vision, can not be all-round tracking and so on, but multi-camera tracking system can well overcome these problems. The traditional tracking methods often ignore problems such as inaccurate geometric constraints or violate the collection restrictions, but requires more sophisticated methods to deal with this problem. Based on previous studies, the paper focuses on how to improve the multi-camera multiple target detection accuracy and the problems of target fusion and tracking after detection, to realize the multiple target tracking under multi-camera environment The main work of this paper are as follows:(1) First of all, the paper analyzes target detection algorithm based on human principal axis, according to the characteristics that the human body principal axis almost completely symmetrical on both sides, we detect the targets over multi-cameras by appling the Least Median of Squares to determine the principal axis of an isolated person,and acquire the corresponding position of each target in different cameras. Based on single-camera tracking, our method first acquire the human body’s binary image by background subtraction,and then get the axis information of the target by the Least Median of Squares,finally we realize the dectection of the target under multi-camera environment by calculating the homography between cameras and the target’s corresponding position in different cameras acquired by the homography.(2) Secondly, studied a multi-camera multi-target tracking algorithm based on MCMC particle filter. This algorithm first established autoregressive motion model and observation model based on the target color histogram and motion histogram, using particle filter for tracking in each single camera, and then fuse the data from cameras in central processor and tracking the targets over multi-cameras based on the MCMC particle filter. This method contrasted the laboratory data and the public data sets of CARIAR and PETS, and compared with other classical methods. The experiment results show that the method has more advantages on the recall and precision than traditional methods.(3) Finally, we established a small multi-target tracking system under a laboratory environment by three cameras.Based on certain hardware platform, the system embeds the detection and tracking algorithms into software system and can track targets under three cameras.The system has a simple interface that can accurately show the track effects. Experimental results show that the system can track3-4person through the interface in a reliable way.