Small-target traffic sign detection based on improved YOLOv8

To address the challenges of high proportions of small traffic signs and significant environmental interference in road traffic scenarios, an improved YOLOv8 model specifically designed for small traffic sign detection is proposed. We first integrated Bi-level Routing Attention (BRA) based on the Transformer architecture into the C2f network structure of YOLOv8n. This integration aims to reduce missed detections of small objects and improve the network’s perception of small targets. In addition, Content-Aware Reassembly of Features (CARAFE) is employed to retain complex image feature information and reduce the loss of target information without significantly increasing model complexity and parameters. This approach enhances the model’s accuracy and generalization capability. Inner-IoU loss function based on auxiliary bounding boxes is also introduced. The experiments are conducted on the CCTSDB2021 and TT100K datasets. The results show that the proposed model achieve improvements of 1.8% in mAP@0.5 and 1.7% in mAP@0.5:0.95 compared to the baseline model on the CCTSDB2021 dataset. On the TT100K dataset, the proposed model achieve a 6.8% increase in precision, an 8.9% increase in recall, and a 7.9% improvement in mAP@0.5 compare to the baseline model. Experimental results indicate that the improved YOLOv8 model effectively detects traffic signs in complex road scenarios.