Spatio-temporal Consistency and Hierarchical Matching for Multi-Target Multi-Camera Vehicle Tracking

Peilun Li, Guozhen Li, Zhangxi Yan, Youzeng Li, Meiqi Lu, Pengfei Xu, Yang Gu, Bing Bai, Yifei Zhang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 222-230

Recently, many approaches have been addressed to realize Multi-Target Multi-Camera(MTMC) vehicle tracking, which is critical in intelligent transportation system (ITS). Continuous improvements of MTMC have been limited by two modules - trajectory feature representation and feature metric in the city-scale camera condition. In this paper, we propose a spatio-temporal consistency and hierarchical matching method to overcome the challenges. As first step, a popular object detection and object tracking method are implemented to detect vehicles and track them in single camera, thus achieved high performance. The smoothness of trajectory and slice direction of movement make spatio-temporal consistency more confident. As second step, a bottom-up hierarchical match strategy is used to match targets in different cameras. Top performance in City-Scale Multi-Camera Vehicle Tracking task at the NVIDIA AI City Challenge 2019 demonstrated the advantage of our methods.