【摘要】 针对在雾天环境下的道路目标检测存在特征提取难度大、精度与效率难以平衡的问题，文中在YOLOv5n的基础上创新性地提出一种高效且轻量的YOLOv5n-BGF变体，该变体结合了双向特征金字塔网络（BiFPN）模型，利用双向连接的结构特点，更加有效地结合不同尺度的特征；其次，引入GELAN模块代替颈部网络中的C3结构，在减少计算量的同时增强了有效特征的提取；最后，考虑不同样本的边界框回归问题，采用Focaler-IoU来进一步提高检测性能。在本地平台针对非公开雾天道路目标检测数据集D-8800进行验证，实验结果表明，相较于基础模型YOLOv5n，改进后的YOLOv5n-BGF的mAP@0.5提升了5.3%，参数量减少了25%,GFLOPs仅为3.5,YOLOv5n-BGF凭借其卓越的性能，在雾天道路目标检测数据集D-8800上的表现优于其他目标检测模型，为雾天道路目标检测提供了高效的解决方案。更多还原

【Abstract】 There are difficulties in the road object detection on foggy days, such as feature extraction and balancing accuracy and efficiency. In view of this, the study innovatively proposes an efficient and lightweight variant based on YOLOv5n, and the variant is named YOLOv5n-BGF. This variant incorporates a bi-directional feature pyramid network(BiFPN) model, which leverages the structural characteristics of bidirectional connections to effectively integrate features of different scales. Furthermore, a GELAN module is introduced to replace the C3 structure in the neck network, which enhances the extraction of valid features while reducing computational load. Finally, on the basis of taking account of the bounding box regression of different samples, Focaler-IoU is employed to improve the detection performance. The proposed model has been validated on a local platform on the private foggy road object detection dataset D-8800. Experimental results indicate that in comparison with the base model YOLOv5n, the mAP@0.5 of the YOLOv5n-BGF is increased by 5.3%, its parameter count is reduced by 25%, yet its GFLOPs is only 3.5. With its exceptional performance, the YOLOv5n-BGF is superior to the other object detection models on the foggy road object detection dataset D-8800. To sum up, it provides an efficient solution for foggy road object detection.更多还原