EDN-YOLO: Multi-scale traffic sign detection method in complex scenes

Traffic sign detection poses a challenging task for autonomous driving systems, particularly in complex scenarios with multi-scale traffic sign objects. The significant variations of the traffic signs scale leads to the issues of false positives and false negatives in the detection results. Therefore, a detection method named EDN-YOLO based on YOLOv5s is proposed in this study. Firstly, this study uses EfficientVit to replace the original backbone network to improve the global feature perception performance of the model without increasing the complexity of the model. Secondly, an efficient decoupled detection head is employed to handle the classification and regression tasks separately, thereby improving the detection performance of the model on multi-scale targets. Finally, an optimized loss function is introduced, combining CIoU loss with NWD loss as the localization loss, which enhances the sensitivity of model to small traffic signs. Experimental validation on the GTSDB datasets and TT100K datasets demonstrates that the proposed EDN-YOLO model exhibits outstanding performance in the detection of multi-scale traffic signs in complex scenarios. Compared with the original YOLOv5s model, mAP is improved by 3.1% on the GTSDB dataset and by 9.6% on the TT100K dataset. The experimental results indicate that the EDN-YOLO method significantly enhances the capability of detecting multi-scale traffic signs in complex scenarios, holding practical significance for real-time traffic sign detection and the advancement of intelligent transportation systems. Code is available at https://github.com/itouchzh/YOLO_ED.