Real-time Traffic Incident Detection and Tracking System using YOLOv10 and ByteTrack Algorithm

Resumen

Many large cities throughout the world struggle with traffic congestion, and one of the main causes of it is traffic incidents, which can result in a high number of injuries and fatalities. The automatic, real-time identification of traffic anomalies—such as collisions, stopped cars, debris, or congestion—using sensors, AI, and security cameras is known as traffic event detection. These technologies, which are a component of intelligent transport systems (ITS), reduce secondary incidents by analyzing traffic flow and alerting operators more quickly than through manual reporting. Numerous research have been conducted about the detection or prevention of incidents. The majority of them feature accident statistics or sensing objects that could cause accidents. In this research, a system which detects happening incidents will be studied. The system will collect necessary information from neighbor vehicles and process that information using deep learning tools to detect possible accidents.  The aim of this research is to develop an algorithm to detect car accident based on hybrid byte track algorithm with YOLOv10. The integration of a tracking system to accurately track incidents. The proposed system for traffic incident detection consists of three stages. Frames were elicited from the input video, and then resizing, normalization and reduction are applied in the first stage.  In the second stage, the System applied the YOLOv10 paradigm to detect the class type of the event and to reduce latency. A byte track algorithm was utilized in the third stage to maximizes tracking accuracy by utilizing nearly all detection boxes, including low-confidence ones. Based on visual patterns and characteristics, the system Indicate that you used YOLOv10 to reduce latency. The integration of the tracking system to accurately track incidents. The model achieved promising results in classifying 4 different categories (Human Incident, Vehicle Incident, Human Normal, and Vehicle Normal). The experimental results show for example of vehicle incident that the precision 84%, a recall 79.5%, mAP50 85.4% and mAP50-95 61.2%.