Real-time traffic accident detection using yolov8

Urban traffic plays an essential role in daily activities, but it carries many safety risks. A challenge that must be addressed is the need for an immediate response when an accident happens. It is a fact that traffic manager often takes a significant amount of time to recognize emergencies and make a decision regarding assistance activities. With the widespread application of artificial intelligence, particularly deep learning, it is possible to detect accidents in real time and promptly initiate rescue operations to minimize casualties and property damage. Among these deep learning algorithms, the YOLO algorithm has emerged as a potential solution. Recently, the latest version of YOLO, known as YOLOv8, has shown superior performance regarding accuracy and precision across various fields such as security, e-commerce, education, and entertainment. However, its application in the traffic safety field, such as addressing traffic accidents, has yet to be fully explored. Consequently, this paper focuses on implementing a series of YOLOv8 e.g. n, s and m, for accident detection. Some hyperparameters e.g. batch size and number of epochs are investigated to gain insights into their impact on the performance of YOLOv8. Accident images under consideration of normal and extreme weather are collected to train YOLOv8. Then the trained model is utilized to identify accidents using information extracted from cameras. When using a batch size of 16 and running for 100 epochs, YOLOv8s proved to be superior performance in terms of computational time (0.27 hours) and precision when employed to identify traffic conditions using videos from CCTV cameras.