AI City Challenge 2020 - Computer Vision for Smart Transportation Applications

Ming-Ching Chang, Chen-Kuo Chiang, Chun-Ming Tsai, Yun-Kai Chang, Hsuan-Lun Chiang, Yu-An Wang, Shih-Ya Chang, Yun-Lun Li, Ming-Shuin Tsai, Hung-Yu Tseng; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 620-621

We present methods developed in our participation of the AI City 2020 Challenge (AIC20) and report evaluation results in this contest. With the blooming of AI computer vision techniques, vehicle detection, tracking, identification, and counting all have advanced significantly. However, whether these technologies are ready for real-world smart transportation usage is still a open question. The goal of this work is to apply and integrate state-of-the-art techniques for solving the challenge problems under a standardized setup and evaluation. We participated all 4 AIC20 challenge tracks (T1 to T4). In T1 challenge, we perform vehicle counting by associating deep features extracted from Mask-RCNN detections and tracklets, followed by vehicle movement zone matching. In T2 challenge, we perform vehicle type and color classification and then rank matching vehicles using a PGAM re-id network. In T3 challenge, we proposed a new Multi-Camera Tracking Network (MTCN) that takes single-camera vehicle tracking as input, and performs multi-camera tracklet fusion and linking, by jointly optimizing the matching of vehicle appearance and physical features. In T4 challenge, we adopt a leading method based on perspective detection and spatial-temporal matrix discriminating, and improve it with background modeling for traffic anomaly detection. We achieved top-6 and top-4 performance for T3 and T4 challenges respectively in the AIC20 leaderboard.