MSDGFM-YOLO11n-OBB: Selective High-Resolution Difference-Guided RGB-Infrared Fusion for Aerial Oriented Vehicle Detection.

Abstract

Aerial vehicle detection using red-green-blue (RGB) and infrared (IR) imagery supports unmanned aerial vehicle (UAV) perception, but oriented localization remains difficult because vehicles are small, dense, arbitrarily oriented, and affected by illumination. Fusion must decide both how cross-modal differences are used and where adaptive fusion is placed. We propose MSDGFM-YOLO11n-OBB, a dual-branch oriented bounding box (OBB) detector with a Modality-Selective Difference-Guided Fusion Module (MSDGFM). Independent backbones preserve modality-specific features. MSDGFM learns sample-dependent RGB/IR weights and injects absolute cross-modal differences through a spatial gate and learnable channel-wise residual; zero initialization makes equal-weight Mean Fusion the exact starting solution. Adaptive fusion is applied only at high-resolution P3, while P4/P5 retain parameter-free Mean Fusion. On DroneVehicle Shared-label data, the three-seed results reach 0.7853 ± 0.0027 mean average precision (mAP)@0.50 and 0.6472 ± 0.0013 mAP@0.50:0.95, exceeding the strongest evaluated single-modal nano-scale OBB baseline by 3.88 and 6.82 percentage points. Relative to matched Mean Fusion, the gain averages 0.83 points in mAP@0.50:0.95 and is positive on all three seeds. In seed-0 capacity control, it exceeds YOLO11s-OBB (IR) by 4.32 points with 58.1% fewer parameters and 54.0% fewer FLOPs. These results support compact, selective difference-guided fusion for aerial RGB-IR OBB detection.