An Improved YOLOv8 Model for Small Traffic Sign Detection Using P2 Feature Fusion and CBAM

ABSTRACT

Small traffic sign detection in complex road scenes is still challenging. Traffic signs in the image are usually small in size, weak in visual features, and susceptible to interference from complex backgrounds. Although the YOLO‐based model performs well in small target detection tasks, its repeated downsampling process may weaken some fine‐grained spatial information, thus affecting the detection accuracy. In order to improve this problem, this study proposes P2C‐YOLOv8. The model introduces a P2 high‐resolution feature layer to retain more location information related to small traffic signs. At the same time, a CBAM module is embedded before prediction to refine multi‐scale features and suppress irrelevant background responses. The experiment was carried out on the TT100K2016 subset. Compared with Faster R‐CNN, RetinaNet, YOLOv8 and P2‐YOLOv8, P2C‐YOLOv8 achieved the highest or joint‐highest Precision across the three traffic sign categories, reaching 0.899, 0.839, and 0.901 for indication, prohibitory, and warning signs, respectively. Its F1 score is also higher than YOLOv8. The confusion matrix and visualization results show that the model can reduce the false detection caused by background interference and improve the positioning effect of small traffic signs, so as to achieve more stable recognition performance in a complex road environment.