LAA-YOLO: a traffic sign detection method in foggy weather based on asymptotic feature fusion and attention
Haixiang Li, Yixin Sui, Daohui Zheng, Zhenyan Chu, Xuelian Sun · 2025
Traffic sign detection in foggy weather is a challenging task for unmanned driving systems. Fog introduces noise pollution to the images, which will affect the detection accuracy. In this paper, a traffic sign detection method based on asymptotic feature fusion and attention is proposed, to address the problem of large parameters and poor noise resistance of existing models in foggy weather. Firstly, a atrous selective kernel attention is proposed to improve the ability to extract target features by dynamically adjusting the receptive field range. Secondly, a lightweight asymptotic feature pyramid network is proposed to enhance the interaction of non-adjacent layer information in feature fusion. Thirdly, a novel lightweight convolution, GSConv, is integrated into the backbone network to reduce the model's parameters. The experimental results show that mAP50 achieves 98% on the CCTSDB 2021 dataset, respectively, which is 2.7% higher than the original model, and the parameters are decreased by about 40%. It shows that our method achieves detection results with fewer parameters and higher accuracy.