Abstract
Ensuring road safety and advancing autonomous driving technology necessitate the accurate and timely detection of traffic signs. Traffic sign detection faces challenges from complex backgrounds and varying sizes, and substantial computational demands. To tackle these challenges, this study presents YOLO-SAL, a lightweight model for traffic sign detection. Initially, the model adopts and innovates on the SCConv concept by introducing the SCC2f design. This innovation optimizes the conventional convolutional residual block using spatial and channel mechanisms, drastically cut-ting the model's parameters and computational load. Furthermore, the model enhances the Adaptive Feature Pyramid Network (AFPN) by fostering improved interactions across different layers through a progressive feature pyramid network. This enhancement boosts the accuracy of detecting traffic signs across a range of sizes. Additionally, the model incorporates the Long-Sequence Knowledge Attention (LSKA) at the detection layer, leveraging attention principles to refine the model's focus on crucial information. Experimental results demonstrate that, compared with the baseline YOLOv8n model, the proposed YOLO-SAL improves mean Average Precision (mAP) by 4.9%, while reducing model parameters by 13.3% and computational load (FLOPs) by 8.6%, respectively. These advancements highlight the model's ability to meet the needs for fast and accurate traffic sign detection.