Environment-Driven Fusion and Optimization Network for RGB-T Person Detection in Underground Mines
Jiajun Xu, Fuming Qu, Zhanliang Niu, Yaming Ji, Lingyu Zhao, Weihua Zhou · Processes · 2026
Reliable personnel detection in underground mines is essential for safe intelligent mining, but uneven illumination, dust occlusion, thermal interference, and cross-modal parallax can degrade RGB-T perception. We propose an Environment-Driven Fusion and Optimization Network (EnvFONet), which uses environmental degradation as a unified prior for feature representation, multimodal fusion, and training optimization. First, Dynamic Receptive Field Relaxation (DRFR) adaptively interpolates original and locally smoothed thermal features to balance fine-detail preservation and contextual robustness. Second, an Env-AdaIN-guided spatially gated fusion module applies environment-conditioned residual affine modulation and high-frequency pixel-level weighting to reduce cross-modal statistical drift and parallax-induced boundary artifacts. Third, Environment-Adaptive Relaxation Loss (EARL) adjusts regression supervision for difficult matched samples according to environmental degradation and localization quality. On the self-collected Anshan underground iron-ore mine dataset, EnvFONet achieves 96.8% mAP50, 63.9% mAP75, and 62.8% mAP50:95. On the public LLVIP benchmark, it achieves 94.4% mAP50 and 59.6% mAP50:95. These results show that environment-conditioned fusion and optimization improve RGB-T personnel detection robustness under the evaluated degraded conditions.