基于改进YOLOv8的恶劣天气交通标志检测算法

Improved YOLOv8 for Traffic Sign Detection Algorithm in Adverse Weather Conditions

摘要: 交通标志检测是自动驾驶环境感知系统的重要组成部分。针对恶劣天气条件下图像对比度降低、特征模糊和细节丢失导致的漏检、误检问题，以及车载边缘平台对模型轻量化和实时性的要求，本文提出了一种改进YOLOv8的恶劣天气交通标志检测算法。首先，在骨干网络中引入GCConv模块，以增强复杂场景下的特征提取能力；其次，在LSCD检测头的基础上改进设计SA-LSCD检测头，以降低模型参数量和计算量并增强多尺度特征表达能力；最后，采用Focaler-PIoU2损失函数替代CIoU损失函数，以提升小目标边界框回归精度。实验结果表明，相比原始YOLOv8n，改进模型的参数量由3.01 M降低至2.36 M，FLOPs由8.1 G降低至6.5 G，FPS由125提升至141，mAP50由81.7%提升至84.4%，mAP50-95由53.2%提升至55.6%。该方法在保持模型轻量化和实时性的同时，有效提升了恶劣天气下交通标志检测的精度与鲁棒性。

Abstract: Traffic sign detection is an essential component of environmental perception systems for autonomous driving. To address the problems of missed detections and false detections caused by reduced image contrast, blurred features, and loss of detail under adverse weather conditions, as well as the requirements for lightweight design and real-time performance on vehicular edge platforms, this paper proposes an improved YOLOv8-based algorithm for traffic sign detection in adverse weather. First, the GCConv module is introduced into the backbone network to enhance feature extraction capability in complex scenarios. Second, the SA-LSCD detection head is developed based on the LSCD head to reduce the number of model parameters and computational cost while strengthening multi-scale feature representation. Finally, the Focaler-PIoU2 loss function is adopted to replace the original CIoU loss so as to improve the bounding box regression accuracy for small targets. Experimental results show that, compared with the original YOLOv8n, the improved model reduces the number of parameters from 3.01 M to 2.36 M, the FLOPs from 8.1 G to 6.5 G, and increases the FPS from 125 to 141, while improving mAP50 from 81.7% to 84.4% and mAP50-95 from 53.2% to 55.6%. The proposed method effectively improves the accuracy and robustness of traffic sign detection in adverse weather while maintaining a lightweight model design and real-time performance.