Multi-person Multi-Camera Tracking for Live Stream Videos Based on Improved Motion Model and Matching Cascade
Yundong Guo,Xinjie Wang,Hao Luo,Huijie Pu,Zhenyu Liu,Jianrong Tan
DOI: https://doi.org/10.1016/j.neucom.2021.12.047
IF: 6
2021-01-01
Neurocomputing
Abstract:In a small-scale distributed multi-camera system like video surveillance system of a museum, shopping mall, plaza, etc., or advanced driving assistance system, real-time multi-person tracking is essential for public and pedestrian safety consideration in the smart security system. In this paper, a real-time multi-person multi-camera tracking framework is presented, which is compatible with both overlapping and non-overlapping views. Since cameras have different orientations and exposures, false matching occurs frequently when people cross the camera boundaries or reenter the same camera. To deal with this challenge, an improved multi-person multi-camera matching cascade scheme is proposed, which can increase the accuracy of inter-camera person re-identification (Re-ID) by taking advantage of association priorities of targets and features. Besides, the proposed method can deal with the occlusion of people and variation of appearance features. Experiments are implemented with overlapping and non-overlapping videos, and results show that the proposed method has robust performance in different situations.