Real-time UAV small object detection: An efficient approach using SGFNet and dynamic loss optimization

In recent years, unmanned aerial vehicles (UAVs) have become core tools in traffic monitoring applications due to their flexible perspective control and efficient data acquisition capabilities. However, UAV images commonly exhibit small objects (dimensions <16×16 pixels) and motion blur phenomena, which leads to insufficient semantic representation of shallow features and significantly constrains detection accuracy. Meanwhile, the contradiction between existing networks' high computational complexity and practical scenarios' demand for precise real-time detection makes balancing model accuracy and inference efficiency a critical challenge. To address these issues, this paper proposes SGF-YOLO, a real-time efficient small object detection algorithm. We enhance shallow semantic representation and optimize cross-level feature transfer efficiency through integrating a horizontal attention feature fusion (HAFF) module with a multi-scale feature encoder (MFE). Our work innovatively combines auxiliary bounding boxes with a dynamic non-monotonic focusing mechanism to achieve more robust localization optimization in dense scenes. The proposed method is evaluated on three distinct urban UAV image datasets (VisDrone2021-DET, CARPK, and HazyDet) dedicated to civil applications. Compared with the YOLOv8m baseline model, our approach achieves a 10.3% improvement in mAP50 and 18.6% increase in $A{P}_{small}$ on the VisDrone2021 dataset, while significantly reducing model size and computational cost (FLOPS). When compared to YOLOv8x, the proposed model demonstrates 68.8% fewer parameters and 40.21% faster inference speed. This study provides a practical solution for UAV-based traffic monitoring systems that effectively balances accuracy, speed, and deployment costs.