Research on Multispectral Human Target Detection Algorithm in Underground Mining

Abstract To address the instability of single-modal detection caused by weak illumination, dust occlusion, and thermal-source interference in underground coal mines, a dual-stream detection framework integrating visible-light and infrared images was developed. Independent preprocessing pipelines were designed for the two modalities, and early fusion was performed along the channel dimension to construct a unified bimodal input representation. On this basis, a gated feature enhancement mechanism was introduced into the You Only Look Once version 5s network to adaptively select informative modal features and improve target perception in complex underground environments. A public paired visible-light and infrared dataset collected under low-light conditions was used to simulate underground operating scenarios, and both ablation and comparative experiments were conducted. The results show that the infrared unimodal model outperformed the visible-light unimodal model, with the accuracy increasing from 94.33% to 96.11% and the mean average precision improving by 2.61%, indicating the higher reliability of infrared imaging in low-light conditions. In addition, the proposed bimodal fusion scheme outperformed all unimodal baselines, and the split-stream four-channel fusion configuration achieved a mean average precision of 98.41%. Although the proposed model showed slightly lower localization performance than the comparison model at stricter evaluation thresholds, it reduced the number of parameters, computational cost, and memory consumption by 43.3%, 38.8%, and 44.7%, respectively, demonstrating a better balance between detection accuracy and model complexity.