Real-Time Traffic Monitoring and Status Detection with a Multi-vehicle Tracking System

With live street videos posted online, the Macao Government provides means to the general public to assess the latest road traffic conditions. After reviewing over these videos, a person may decide to change the travel route from the one he or she initially plans to take. To let road users make decisions better and faster, it would be desirable to design an automated software, being a component of an Intelligent Transport System, which offers proper suggestions to the users instantly upon analyzing all available live videos. In this paper, we propose to create a real-time road traffic condition estimation system. Its design is based on a combination of deep learning algorithms: the YOLOv5, DeepSORT, and the Non-Maximum Suppression algorithms. Putting together the YOLOv5 with our proposed two-stage NMS strategy, the improvement on the efficiency of object detection on live videos is noticeable. Our two-stage strategy removes the requirement to manually tune the NMS parameters continuously. With DeepSORT, we are able to track moving vehicles, and create motion trajectories, which we can use filtering strategy to assess the latest road traffic conditions. Since different lanes on a road may have different traffic situations, we separate the lanes based on angles and propose to use a lane status score independently for each lane. Through the experimental results, our system design could estimate the traffic status in real-time without requiring any manual parametric adjustments.

Automated summary

This paper proposes a real-time road traffic condition estimation system based on deep learning algorithms. By analyzing live videos and using improved object detection and tracking algorithms, the system can provide instant suggestions to road users and independently assess the traffic conditions of each lane.