<p>Accurate traffic sign recognition under complex weather conditions is critical for ensuring road safety and enabling robust perception modules in all-weather intelligent driving systems. Adverse environments, such as rain, snow, and low-light nighttime conditions, often lead to occlusion, texture degradation, and increased image noise, which significantly reduce detection accuracy and hinder a vehicle’s ability to respond in real time to key information such as speed limits and directional instructions. To address these challenges, we propose WAC-YOLO, an improved traffic sign recognition model based on YOLOv8. First, we integrate a Bidirectional Weighted Feature Pyramid Network (BiFPN) to enhance multi-scale feature fusion, and replace traditional downsampling and upsampling operations with WaveletPool to improve accuracy while maintaining a lightweight design. Second, a novel C2f-WT module is introduced in the backbone to enhance textural extraction under low visibility, and a C2f-AC module is introduced to enhance the model’s perception of local details in partially occluded traffic signs. Finally, Inner-ShapeIoU is employed as the loss function, integrating auxiliary box constraints and geometry-aware weighting to enhance the accuracy of bounding box regression. Experiments on the enhanced TT100K dataset show that WAC-YOLO outperforms YOLOv8s by 3.3% in mAP50 and 3.9% in mAP50:95, while reducing parameters by 43.6% and computational cost by 25.2%, achieving real-time performance at 125 FPS.</p>