In order to tackle the difficulties presented by large – scale fluctuations, densely packed small targets, and category imbalance in traffic sign detection within complex urban traffic situations, this article puts forward an enhanced detection framework for traffic sign feature augmentation. This framework is built upon the YOLOv8n baseline model.When dealing with the CCTSDB dataset, the Star attention module and the ASFF adaptive multi – scale feature fusion module are incorporated. These modules enhance feature representation for occluded and small targets. For the TT100K dataset, coordinate attention and a four-head detection structure are adopted, alongside EMA weight smoothing to boost generalization. Tests on CCTSDB and TT100K show the refined model achieves 0.982 mAP@0.5 and 0.876 mAP@0.5:0.95 on CCTSDB. On TT100K, mAP@0.5 increases from 0.775 to 0.808, with the best F1 score of 0.75. Its performance approaches Faster R-CNN while preserving YOLO’s real-time advantage. This method provides an efficient and reliable traffic sign detection solution for vehicle-assisted driving.
