Abstract: To address the issues of low detection accuracy, insufficient feature extraction, and strong background interference in aerial imagery, this paper proposes an improved RT-DETR object detection algorithm to enhance model performance in complex scenarios. A dual-path enhanced feature extraction module is constructed, introducing a dual-path convolution structure and an omni-dimensional dynamic convolution layer to reduce computational cost while improving the perception of small-scale objects. A lightweight and efficient feature fusion module is then designed, which integrates spatial-depth transform convolution and combines the CSP structure with the OmniKernel mechanism to enhance the coupling of multi-scale semantic information, preserve high-resolution features, and improve the model's ability to distinguish small objects under complex backgrounds. Finally, a FMPDIoU loss function is introduced in the bounding box regression process to alleviate the issues of gradient sparsity and class imbalance in small object detection. Experiments conducted on the VisDrone 2019 dataset show that the proposed model outperforms mainstream object detection algorithms in terms of mAP, recall and precision. Compared with the baseline RT-DETR-R18 model, it achieves a 2.1% improvement in mAP and a 1.1% increase in average precision for small objects, demonstrating excellent detection performance and strong transferability.