Multi-Camera Vehicle Tracking System Based on Spatial-Temporal Filtering
Pengfei Ren, Kang Lu, Yu Yang, Yun Yang, Guangze Sun, Wei Wang, Gang Wang, Junliang Cao, Zhifeng Zhao, Wei Liu
Abstract
Multi-Camera multi-target tracking is essential in the research field of urban intelligence traffic. It shows that the task becomes challenging due to differences of illumination, angle, and occlusion under different cameras. In this paper, we propose an efficient multi-camera vehicle tracking system, which contains a model trained with multi-loss to extract appearance features, and a filter with spatial-temporal information between cameras. The proposed system includes three parts. Firstly, we generate tracklets in a single-camera with different views by vehicle detection and multi-target tracking. Secondly, we extract the appearance feature of each tracklet through the trained vehicle ReID model. Thirdly, we innovatively propose a matching strategy that calculates several factors, the similarity of appearance features, the time information, and the space information of target ID between adjacent cameras. The proposed system ranks the sixth place in the City-Scale Multi-Camera Vehicle Tracking of AI City 2021 Challenge (Track 3) with a score of 0.5763.