ST-DETR: an improved small object detection algorithm for unmanned aerial vehicles

The increasing deployment of unmanned aerial vehicles (UAVs) in complex environments has pressed the need for accurate and efficient detection of small objects. To enhance the detection of small objects in complex UAV scenarios, this study investigates the real-time detection transformer model, identifies its limitations in practical applications, and proposes spatio-temporal object traces attention detection transformer (ST-DETR) as an improved framework. In the backbone and neck networks, a HyperScale-ARF module is introduced to enhance the collection and fusion of small-scale features, whereas a global–local feature fusion module is designed to incorporate a global self-attention mechanism and a dual-branch dynamic gating structure, enabling the model to adapt to detection tasks in diverse environments. The experimental results demonstrate that the improved ST-DETR model achieves mean average precision improvements of 1.7% and 3.1% on the VisDrone2019 and VOC2012 datasets, respectively, verifying its effectiveness and robustness in small object detection tasks.