Hardware-Aware Decoupled Modulation for RGBT object detection in UAVs

Day-and-night Unmanned Aerial Vehicle (UAV) object detection faces immense challenges in complex environments, particularly the cross-modal spatial parallax inherent in dual-camera systems and the high-frequency edge loss of minuscule targets during multi-scale feature upsampling. To surmount these distinct physical bottlenecks, this paper proposes a highly efficient RGB-Thermal (RGBT) dual-stream detection framework based on YOLO26s, introducing two decoupled spatial modulation mechanisms. First, the Cross-Modal Residual Calibration (CMRC) module performs difference-aware spatial attention weighting, effectively suppressing parallax-induced ghosting in unaligned dual-lens systems. Second, the Soft-Gated Dynamic Alignment Upsampling (SG-DAU) module performs context-aware high-frequency edge reconstruction for strictly aligned coaxial sensors. Rather than naively concatenating these modules, we propose a novel Hardware-Aware Decoupled Modulation (HADM) strategy. By utilizing deterministic hardware priors to isolate CMRC for non-coaxial parallax hardware and SG-DAU for strictly aligned sensors, HADM effectively prevents spatial geometric interference. Extensive orthogonal cross-scenario validation confirms the superiority of this strategy. On the parallax-heavy DroneVehicle dataset, the CMRC-driven configuration achieves a state-of-the-art mAP@0.5 of 82.64% with merely 22.5M parameters. Conversely, on the strictly aligned Low-Light Vision (LLVIP) dataset, the SG-DAU-driven architecture secures a remarkable 59.75% on the stringent mAP@0.5:0.95 metric, substantiating the necessity and robust generalizability of our hardware-specific deployment paradigm.