Lightweight P2CL-YOLO with targeted feature enhancement for traffic sign detection of small and occluded targets

Abstract

To address the challenges of traffic sign detection in complex road environments, particularly the limited discriminability of small objects and localization errors caused by blur and occlusion, this paper proposes P2CL-YOLO, a lightweight detection framework built upon YOLOv11s. The proposed method introduces a unified P2CL design to enhance the representation of small and occluded traffic signs. First, a dedicated P2 detection layer is designed to preserve high-resolution shallow features for small-object perception. On this basis, a lightweight CBAM attention module is embedded to refine channel-spatial features, suppress background interference, and enhance target saliency. The refined features are further exploited through a dual-path strategy: they are fed back into the multi-scale feature fusion pathway to strengthen hierarchical representations, while simultaneously being delivered to the LSDECD detection head for accurate localization of occluded targets via dynamic feature calibration. In addition, the Wise-IoU loss is adopted to improve bounding box regression stability and convergence behavior. Experimental results on the TT100K dataset show that the proposed method achieves 88.3% mAP@0.5, outperforming YOLOv11s by 5.8%. The model contains 10.05 M parameters with 23.7 GFLOPs. Further evaluation on the German Traffic Sign Detection Benchmark dataset demonstrates strong cross-domain generalization ability. Overall, P2CL-YOLO achieves a favorable balance between detection accuracy and computational efficiency, making it suitable for real-time traffic sign detection in resource-constrained environments.