Design of an efficient object detection algorithm for traffic accident recognition

Abstract

Road traffic accidents pose a serious threat to public safety and require rapid and reliable detection for intelligent transportation systems. However, conventional deep learning-based accident detection models often involve substantial computational costs, limiting their deployment on resource-constrained edge devices. In addition, complex traffic scenes involving occlusion, illumination variations, and nighttime conditions can further challenge the robustness of accident detection. To address these issues, this study proposes a lightweight traffic accident detection model, SDD-YOLO11n, designed to achieve an effective balance between detection accuracy and computational efficiency. First, ShuffleNetV2 is introduced as the backbone to reduce redundant computation and construct a compact feature extraction network. Second, a DW_ADown module based on depthwise separable convolution is developed to reduce the computational cost of feature downsampling while preserving useful spatial information. Third, the original SPPF module is replaced with DS_SENetV2, which enhances discriminative feature representation through lightweight channel-wise feature recalibration. Experiments are conducted on a traffic accident dataset containing vehicle and accident instances under diverse traffic conditions. The proposed SDD-YOLO11n achieves an mAP@0.5 of 0.824 with only 0.80 M parameters, 1.91 MB model size, and 2.2 GFLOPs, while maintaining an inference speed of 69.9 FPS. Compared with the baseline YOLO11n, the proposed model substantially reduces model complexity while maintaining competitive detection accuracy, demonstrating its suitability for real-time and resource-constrained edge deployment. The compact architecture and efficient feature extraction also provide a practical basis for maintaining reliable accident recognition in challenging traffic scenes with occlusion and illumination variations. These results demonstrate that SDD-YOLO11n provides a favorable trade-off among detection accuracy, computational complexity, model size, and inference efficiency, making it a promising solution for real-time traffic accident monitoring on edge-oriented intelligent transportation platforms.