面向非机动车闯红灯识别的交通信号灯检测算法优化

Optimization of Traffic Signal Detection Algorithm for Non-Motorized Vehicle Red Light Running Recognition

摘要: 针对电动自行车闯红灯检测过程中交通信号灯存在着小目标漏检、复杂背景误检以及目标定位精度低等问题。提出了一种经过改进的YOLOv8检测方法，即YOLOv8n-BLI。首先在Head层构建BiFPN加权双向特征金字塔，强化对多尺度目标的感知能力。其次在Head的C2f模块之后引入了LSKA注意力机制，提升整个模型的鲁棒性。最后采用InnerMPDIoU取代原本的CIoU损失，提升定位的精度。实验表明：YOLOv8n-BLI检测模型其精确率P、召回率R以及平均精度mAP@0.5分别达到94.8%、92.5%以及94.6%，与近年主流轻量级模型YOLOX-s、PP-YOLOE-s、NanoDet-Plus对比，在保持154.7 FPS推理速度的同时，mAP@0.5分别提高3.9%、2.5%、4.3%。

Abstract: For the detection of red light violations by electric bicycles, there are issues such as small target missed detection, complex background false detection, and low target localization accuracy in traffic signals. A modified YOLOv8 detection method, namely YOLOv8n-BLI, is proposed. First, a weighted bidirectional feature pyramid is constructed in the Head layer to enhance the perception of multi-scale targets. Second, an LSKA attention mechanism is introduced after the C2f module in the Head to improve the robustness of the entire model. Finally, InnerMPDIoU is used instead of the original CIoU loss to enhance localization accuracy. Experiments show that the YOLOv8n-BLI detection model achieves an accuracy rate P, recall rate R, and mean average precision mAP@0.5 of 94.8%, 92.5%, and 94.6%, respectively. Compared with mainstream lightweight models such as YOLOX-s, PP-YOLOE-s, and NanoDet-Plus, while maintaining 154.7 FPS inference speed, mAP@0.5 increases by 3.9%, 2.5%, and 4.3%, respectively.