Video-based road accident detection on highways: A less complex YOLOv5 approach

Abstract

Advancement of Artificial Intelligence (AI) technologies and the availability of high-end computing devices create scope for the implementation of intelligent transport infrastructure for road safety. This paper proposes an intelligent model for accident detection on highways using more robust, less complex, and more accurate YOLOv5 and StrongSort to locate and track vehicles. It provides a reliable approach for detecting accidents based on vehicle speed, acceleration, trajectory anomalies, and area anomalies. The methodology follows three major steps. The first stage performs vehicle detection, the next stage performs vehicle tracking and feature extraction, and the last stage does crash detection. In this study, only vehicle detection and tracking are addressed using a deep learning-based methodology. However, the accident prediction is done by an algorithm using threshold levels of various parameters such as the speed of vehicles, acceleration anomalies, etc. The model shows good performance when evaluated using the test crash videos under different ambient conditions such as daylight and night