A Multi-Camera Vehicle Tracking System Based on City-Scale Vehicle Re-ID and Spatial-Temporal Information

Minghu Wu, Yeqiang Qian, Chunxiang Wang, Ming Yang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 4077-4086

With the demands of the intelligent city and city-scale traffic management, city-scale multi-camera vehicle tracking (MCVT) has become a vital problem. The MCVT is challenging due to frequent occlusion, similar vehicle models, significant feature variation by different lighting conditions, and viewing perspective in different cameras. This paper proposes an MCVT system composed of single-camera tracking (SCT), vehicle re-identification (Re-ID), and multi-camera tracks matching (MCTM). In the SCT phase, we designed a tracker update strategy and used the Re-ID model in advance. We also adopted a template matching method to re-associate the discontinuous tracklets. As for vehicle Re-ID, we implemented a spatial attention mechanism based on the background model. Then we fully leveraged the labels of synthetic data to train attributes Re-ID models as the attributes features extractor. Finally, we proposed an MCTM method to leverage tracklets representation and spatial-temporal information efficiently. Our system is evaluated both on the City-Scale Multi-Camera Vehicle Re-Identification task (Track 2) and City-Scale Multi-Camera Vehicle Tracking task (Track 3) at the AI City Challenge. Our vehicle Re-ID method has achieved 3rd place of Track 2, with an mAP score of 66.50%, and achieved state-of-the-art results on the VeRi776 dataset. Our MCVT system has achieved 3rd place, yielding 76.51% IDF1 of Track 3. Experimental results demonstrate that our system has achieved competitive performance for city-scale traffic management.