A Region-and-Trajectory Movement Matching for Multiple Turn-Counts at Road Intersection on Edge Device

Duong Nguyen-Ngoc Tran, Long Hoang Pham, Huy-Hung Nguyen, Tai Huu-Phuong Tran, Hyung-Joon Jeon, Jae Wook Jeon; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2021, pp. 4087-4094

In intelligent traffic systems, vehicle detection and counting have become an important task. The counting information is essential for reducing traffic congestion and improving traffic signal capability. Traditional methods have been focusing on counting vehicles in a single frame or consecutive frames. However, they have not yet considered the movement of interest (MOI) of the vehicles moving in different lanes and directions. This paper proposes a region-and-trajectory movement matching method that aims to detect and count vehicles for each movement on the road. First, the YOLOv5 detection model is used to detect candidate vehicles in the region of interest (ROI). Second, the SORT tracking method associates vehicles of the same instance in consecutive images to create tracked trajectories. Then, the counting method using the combination of MOI regions and predefined movement tracks. Each tracked trajectory is assigned to the corresponding movement id and is outputted to the result file. The efficiency and effectiveness of the proposed method have been evaluated and ranked 3rd on AI City Challenge 2021 Track 1 leaderboard. Further experiments showed that the method could achieve around 120 fps on an NVIDIA Quadro RTX 8000 and 20 fps on an NVIDIA Jetson Xavier AGX.