GEDSR-DETR: A UAV Object Detection Network With Gaussian-Enhanced Features and Dynamic Sparse Relation
Abstract
Object detection in uncrewed aerial vehicle (UAV) imagery remains challenging due to small object sizes, dense target distributions, complex backgrounds, and large scale variations. Although recent real-time detectors have shown promising efficiency, their performance is still limited by insufficient feature robustness, detail loss in multiscale fusion, and redundant query interactions in crowded aerial scenes. To address these issues, this article proposes GEDSR-DETR, an efficient UAV object detection network built upon RT-DETR. Specifically, a cross stage partial-based Gaussian feature enhancement module is introduced to improve feature discrimination through learnable Gaussian aggregation and lightweight channel interaction. In addition, a multiscale fidelity aggregation module is designed to preserve fine-grained spatial details and enhance semantic consistency across feature levels. Furthermore, a dynamic sparse relation decoder is developed to model informative query relations according to predicted box overlap and spatial proximity, thereby reducing redundant interactions and improving localization in dense scenes. Extensive experiments on the VisDrone2019 benchmark, together with cross-dataset comparisons on UAVDT and HIT-UAV datasets, validate the effectiveness of the proposed method. In particular, on the VisDrone2019 dataset, GEDSR-DETR achieves 33.0% mean average precision (mAP)@0.5:0.95 and 53.1% mAP@0.5, surpassing the baseline RT-DETR by 3.5% and 5.1%, respectively. These results demonstrate the superiority of the proposed method for complex UAV scenes, especially in detecting small and densely distributed objects, while maintaining practical inference efficiency.