SG-YOLO: A multi-module enhanced target detection method adapted to vehicle–road cooperation

In intelligent transportation and autonomous driving, precise object detection is crucial for vehicle–road coordination. While YOLO algorithms have made strides, issues such as low accuracy and limited robustness persist, especially with small objects, complex backgrounds, and object deformations. This paper proposes SG-YOLO, an enhanced version of YOLOv11, designed to overcome these limitations. It introduces three key modules: the CSPPS (Cross-Stage Partial Prediction Sharing) module, which improves information flow and small object detection; the SCConv module, which strengthens local–global feature fusion for complex backgrounds; and the GAM module, which uses global attention for better target localization. Experimental results show that SG-YOLO outperforms YOLOv11 and other methods in precision and recall. Specifically, on the BDD100K dataset, SG-YOLO achieves a precision of 0.742 and recall of 0.521, while on the NEXET dataset, it reaches 0.681 precision and 0.532 recall. These results demonstrate SG-YOLO’s superiority in handling multi-object overlap and small object detection, proving its robustness for autonomous driving.