MCFusion: A Lightweight RGB-T Pedestrian Detection Method with Progressive Thermal Compensation

RGB-T pedestrian detection remains challenging under low-light, occluded, crowded, and complex-background conditions. To improve cross-modal feature fusion while maintaining model efficiency, this paper proposes MCFusion, a lightweight RGB-T pedestrian detection method with progressive thermal compensation. MCFusion adopts a dual-branch RGB–thermal feature extraction structure and introduces a Modality-Compensated Gated Fusion (MCGF) module at the P4 and P5 semantic stages, which is implemented as a zero-initialized residual compensation mechanism. MCGF uses RGB features as the primary stream and progressively compensates them with thermal auxiliary features through a zero-initialized convolutional gate, reducing the interference caused by direct fusion. In addition, a Lightweight Shared Convolutional Detection Head (LSCD) is adopted to reduce redundant computation in multi-scale prediction. On the LLVIP dataset, MCFusion achieves 95.30% mAP${}_{50}$ and 60.10% mAP${}_{50:95}$ with 5.21 M parameters and 10.50 GFLOPs. Compared with the YOLOv11n RGB baseline, it improves mAP${}_{50}$ and mAP${}_{50:95}$ by 7.50 and 10.70 percentage points, respectively. Experiments on KAIST, ablation studies, and visualization results further demonstrate the effectiveness of the proposed method.