A Vehicle Counts by Class Framework Using Distinguished Regions Tracking at Multiple Intersections

Nam Bui, Hongsuk Yi, Jiho Cho; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2020, pp. 578-579

Turning movement counting plays an important step for traffic analysis at complex areas (e.g. intersections). Specifically, accurate and detailed traffic flow information enables the traffic control system to be more efficient and valuable. Recently, with the successful development of Deep Learning for vehicle detection and tracking, the current research focuses on video-based traffic analysis which is regarded as an emergent approach to monitoring vehicle movements. In this study, we present a comprehensive vehicle counting framework by integrating state-of-the-art techniques of object detection and tracking such as Yolo and DeepSort. Furthermore, in order to improve the vehicle counting problem, we propose a distinguished region tracking approach for the vehicle trajectory monitoring, which is able to work well with various scenarios, especially in complex areas with complicated movements. Regarding the experiment, the proposed framework is evaluated on the CVPR AI City Challenge 2020 dataset. Accordingly, our method is able to achieve around 85% of the accuracy which places to the top 10 of the leaderboard in Track 1 of the Challenge.