AI City Challenge 2019 -- City-Scale Video Analytics for Smart Transportation

Ming-Ching Chang, Jiayi Wei, Zheng-An Zhu, Yan-Ming Chen, Chan-Shuo Hu, Ming-Xiu Jiang, Chen-Kuo Chiang; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019, pp. 99-108

Understanding large-scale video traffic big data is the new frontier of today's AI smart transportation advancement. The AI City Challenge 2019 is the third sequel of a yearly event that draws significantly growing attention and participation. This paper presents works contributed to the three Challenges Tracks. In Track 1 City-Scale Multi-Camera Vehicle Tracking, we developed a new multi-camera fusion method by extending the state-of-the-art single-camera tracking-by-detection with site calibrations. Our approach jointly optimizes the matching of vehicle image features and geometrical factors including trajectory continuity, vehicle moving directions and travel duration across views, to effectively fuse tracks and identify vehicles across 40+ cameras in a city-wide scale. In Track 2 City-Scale Multi-Camera Vehicle Re-Identification, we propose a Pyramid Granularity Attentive Model (PGAM) for ReID by improving the recent Region-Aware deep Model (RAM) with a pyramid design and training strategy improvements. In Track 3 Traffic Anomaly Detection, we improved the 2nd-best method from AIC2018 with refined event recognizers of stalled vehicles with back-tracking to accurately locate event occurrence. The proposed methods achieve compelling performance in the leaderboard among 80+ world-wide participant teams.