D2G-Mamba: A Distribution-to-Geometry Alignment Paradigm for Efficient Cross-Modal Object Detection

Reliable perception in adverse illumination and weather is essential for applications such as autonomous driving. Visible-infrared fusion mitigates single-sensor failure modes, yet semantic misalignment between modalities and the quadratic cost of Transformer-style fusion remain practical obstacles. We present a dual-stream detector on the YOLOv11 backbone in which cross-modal cues are aligned under a Distribution-to-Geometry (D2G) formulation: shared invariant decoding and moment-based distribution matching are followed by differentiable sampling from predicted offsets so that RGB features are geometrically reconciled with infrared references before fusion. A dual-path neck then integrates standard FPN–PAN fusion with a Mamba-based branch that mixes modalities after state-space channel weaving and uncertainty-aware gating, yielding long-range context at linear complexity in sequence length. On LLVIP the network reaches 88.73% mAP@0.5, 50.5% mAP@0.5:0.95, and 49.18% mAP@0.75 at 234.72 FPS; zero-shot tests on CVC-14 and M3FD report 46.8% and 45.3% mAP@0.5:0.95, indicating useful transfer without dataset-specific retraining.