Automatic Traffic Accident Detection Using LSTM-Based Time Series Analysis

In this study, we propose a YOLO-LSTM-based model for real-time traffic accident detection, verified using the CarCrashDataset (CCD), which contains 4,500 curated videos covering a diverse range of driving conditions. Our approach integrates YOLOv5 for rapid object detection—focusing on vehicles, motorcycles, and pedestrians—with an LSTM module that captures critical time-series patterns indicating imminent collisions. By leveraging both pre-extracted features from CCD and dynamically generated YOLO outputs, the model achieves robust performance in challenging scenarios such as night driving, inclement weather, and partial occlusions. Experimental results demonstrate high accuracy and recall, confirming the system’s ability to reliably identify accidents, often before collisions occur. Furthermore, the fusion of YOLO-based detections with advanced feature representations enhances detection precision while minimizing false alarms. These findings highlight the system’s potential for deployment in intelligent transportation infrastructures, where proactive warnings can significantly mitigate accident-related damage. Future work will explore transformer architectures and multi-sensor data fusion, thereby offering a strong foundation for next-generation traffic safety applications demanding reliable, real-time accident prevention.