A Deep Learning Framework for Traffic Accident Detection Based on Improved YOLO11

The automatic detection of traffic accidents plays an increasingly vital role in advancing intelligent traffic monitoring systems and improving road safety. Leveraging computer vision techniques offers a promising solution, enabling rapid, reliable, and automated identification of accidents, thereby significantly reducing emergency response times. This study proposes an enhanced version of the YOLO11 architecture, termed YOLO11-AMF. The proposed model integrates a Mamba-Like Linear Attention (MLLA) mechanism, an Asymptotic Feature Pyramid Network (AFPN), and a novel Focaler-IoU loss function to optimize traffic accident detection performance under complex and diverse conditions. The MLLA module introduces efficient linear attention to improve contextual representation, while the AFPN adopts an asymptotic feature fusion strategy to enhance the expressiveness of the detection head. The Focaler-IoU further refines bounding box regression for improved localization accuracy. To evaluate the proposed model, a custom dataset of traffic accident images was constructed. Experimental results demonstrate that the enhanced model achieves precision, recall, mAP50, and mAP50–95 scores of 96.5%, 82.9%, 90.0%, and 66.0%, respectively, surpassing the baseline YOLO11n by 6.5%, 6.0%, 6.3%, and 6.3% on these metrics. These findings demonstrate the effectiveness of the proposed enhancements and suggest the model’s potential for robust and accurate traffic accident detection within real-world conditions.