With the development of smart cities and the increase in vehicles, traffic accident detection has become a crucial component of road safety monitoring in smart cities. Existing traffic accident detection methods suffer from an imbalance between speed and accuracy, poor adaptability to diverse scenarios, and insufficient generalization capabilities, making it challenging to meet real-time monitoring requirements. In this study, a YOLOv5-based traffic accident detection method is proposed to enhance system efficiency in complex road environments. The research leverages the deep learning object detection framework YOLOv5 and optimizes model parameters and training strategies to efficiently identify and localize traffic accidents. The experiments conducted on a self-constructed traffic accident dataset demonstrate that the proposed model achieves superior performance in accuracy and speed, with a of 90.93%, while maintaining high inference speed and meeting real-time detection requirements. The study provides an effective solution for traffic accident detection. It validates the feasibility of applying the YOLOv5 model in road scenarios, offering a significant reference for traffic safety monitoring systems in smart cities.