RFHS-RTDETR: Multi-Domain Collaborative Network With Hierarchical Feature Integration for UAV-Based Object Detection
Songtao Tang, Leiming Zhang, Xudong Liu, Rongfu Lv, Ruyi Qin · IEEE Access · 2025
To address UAV-based detection challenges including scale variation, high target density, and hardware limitations, we propose RFHS-RTDETR with four innovations: (1) RMConv, a reparameterized lightweight module for efficient multi-scale feature extraction; (2) FSConv fusing Scharr operator and Fourier transform to enhance edge preservation and robustness; (3) AH module combining HiLo attention for dense target recognition; (4) SOPS Feature Pyramid with hierarchical feature integration of P2 features and dynamic upsampling for small objects. On VisDrone2019, RFHS-RTDETR reduces FLOPs and parameters by 16.2% and 34.3% versus RT-DETR-R18, while improving precision by 2.2%, mAP50 by 2%, and FPS by 17.1%.These advancements demonstrate its practicality for resource-constrained aerial scenarios.