Traffic sign detection is a critical task in autonomous driving environmental perception. However, models often suffer from degraded detection performance in complex real-world scenarios due to variable target scales, blurred fine-grained features, and complex background interference. This paper proposes an improved YOLOv8n detection model, MSD-YOLO, to address these challenges. First, a Multi-scale Detail Enhancement Module (MDEM) is designed, which achieves targeted enhancement of edge features through high-frequency residual modulation and multi-scale cooperative attention. Second, an enhanced feature pyramid network termed SG-FPN is constructed. It introduces soft nearest neighbor interpolation (SNI) for semantic-spatial aligned feature fusion and employs enhanced lightweight convolution (GSConvE) to improve feature representation. Additionally, the Wise-ShapeIoU optimization loss function is adopted, integrating shape-aware geometric constraints and a dynamic sample weighting strategy to enhance the localization accuracy for traffic signs of different scales and shapes. Experiments on the TT100K dataset show that our method effectively improves detection performance, with mAP@0.5 and mAP@0.5:0.95 increasing by 3.1% and 2.7%, respectively, compared to the baseline YOLOv8n. Moreover, cross-dataset evaluations on CCTSDB and GTSDB show that the model exhibits good generalization capability and robustness. The experimental results indicate that the proposed method can enhance detection accuracy while maintaining efficient real-time inference, offering an effective solution for traffic sign detection in complex scenarios.
Li et al. (Thu,) studied this question.