UAV Dense Target Detection Algorithm Based on the Improved RT-DETR

To tackle the challenges of UAV image target detection, where diverse backgrounds and highly dynamic environments often lead to reduced detection accuracy, and the computational complexity of existing algorithms results in performance degradation, this paper introduces an improved UAV dense target detection algorithm based on RT-DETR, termed MRT-DETR. The proposed approach optimizes model performance through three key innovations. First, a lightweight backbone network, CSP-PMFA, is designed to reduce computational load and parameter size while enhancing the model's feature extraction capability. Second, to address the high computational cost, limited model capacity, and efficiency bottlenecks in the original intra-scale feature interaction module, a cascaded group attention (CGA) mechanism is introduced to refine feature representation. Third, a cross-scale feature fusion module, EMFPN, is developed to enhance receptive field adaptability and improve feature processing. Furthermore, the EUCB upsampling module is integrated into the decoding stage, ensuring both feature retention and fusion without compromising accuracy. Experimental results on the VisDrone2019 UAV target detection dataset demonstrate that MRT-DETR model significantly outperforms the original model, achieving a 4.4% improvement in mAP50 and an increase of 9.3 frames per second (FPS), while simultaneously reducing computational cost and parameter size by 14.4% and 39.2%, respectively. These improvements make MRT-DETR a highly efficient and practical solution for real-time UAV-based dense target detection.