Multiple Targets Tracking System Based On Multiple Collaborative Cameras

Video based multi-target tracking technology is a challenging and attractiveresearch area in the field of computer vision. Multi-target tracking technology canserve as a basis for the video analysis technology and be widely used in the field ofvideo surveillance, intelligent transportation and intelligent robots. There is no doubtthat multi-target tracking has broad application prospects.In practical applications, single camera multi-target tracking system, due to itshorizon limitations, it is hard for it to track the targets all the time and it is difficult tosolve the problem of target occlusion. How to take advantage of multi-camera to solvethese problems has becoming a hot topic. This paper proposes an algorithmframework for multi-target tracking via multiple collaborative cameras andimplements a prototype system based on the algorithm framework.In the paper, a tracking-by-detection approach is used for multi-target tracking.With the results of pedestrian detection, tracking firstly be done in every singlecamera, then a multi-camera fusion will be make to track the targets in multiplecollaborative cameras. A multi-thread strategy is used in the system to achieve it. Inthe single camera independent tracking stage, particle filter is used for multi-targettracking and Hungarian algorithm is used for data-association. Traditional cameracalibration based multi-camera tracking approach is not easy to use in practicebecause the multi-camera calibration process is so complex. This paper use a pureimage approach based on planar homography, epipolar line constraint and field ofview constraint for multi-camera fusion and collaborative tracking. This approach iseasy to implement and useful in practice. In addition, in order to improve the speed ofpedestrian detection, foreground connected regions are extracted through codebookbackground modeling algorithm firstly, then use the HOG pedestrian detection algorithm to make more rapid and accurate detection on the foreground connectedregions.The multi-target tracking system has been used as an experiment system deployedin our lab. Experiments show that the system can make automatic pedestrian detectionand tracking in both indoor and outdoor scenarios.