Research on Dense Small Object Detection Algorithm in Aerial Images Based on Improved RT-DETR

Objects in UAV imagery are typically small and densely distributed, while complex backgrounds, occlusion, and progressive downsampling further complicate detection. To improve the balance between detection accuracy and computational cost, this paper proposes an improved model based on RT-DETR-R18. First, partial convolution (PConv) is introduced into the backbone to reduce computational overhead. Second, efficient multi-scale attention (EMA) is embedded within the backbone residual blocks to enhance feature interaction. Finally, Wise-IoUv3 (WIoUv3) with its non-monotonic focusing mechanism is used to optimize bounding box regression during training. Experiments on VisDrone2019 show that the improved model increases mAP50 from 47.2% to 49.6%, mAP50:95 from 28.1% to 30.2%, and small-object average precision (AP_small) from 17.8% to 19.6%, while reducing parameters from 20.2 M to 18.5 M and computation from 60.1 to 49.2 GFLOPs; measured inference speed improves from 67 to 72.3 FPS. Ablation results show that the effect of each module depends on the combination configuration, and the full combination achieves the highest accuracy in this set of experiments. In addition, when trained and evaluated independently on DOTA v1.0, the improved model also yields higher overall detection accuracy. These results indicate that the proposed method improves overall and small-object detection performance while reducing model size and computation, and maintains real-time inference capability.