Abstract: The underground coal mine environment, characterized by strong electromagnetic interference (EMI), low illumination, dust, and dynamic occlusions from large equipment, poses significant challenges. Consequently, achieving continuous and precise personnel localization using a single technology is difficult. To address this challenge, this paper proposes a precise underground coal mine personnel localization method based on the deep fusion of Ultra-Wideband (UWB) and machine vision. The method first constructs a system architecture comprising a UWB localization subsystem, a visual perception subsystem, an edge computing terminal, and a cloud data center. At the algorithmic level, the core innovations are as follows: To address the robustness issues of the visual module under occlusion and low illumination, a masked semantic self-supervised model is proposed. By introducing random masks during the training of an improved YOLOv5 network, the model is compelled to learn deep semantic features of personnel rather than shallow geometric contours, significantly enhancing detection accuracy in partially occluded scenarios. A cross-frame spatio-temporal clustering tracking algorithm is designed to stably associate target identities and suppress drift in continuous video sequences. To achieve optimal fusion of heterogeneous data, a dynamic adaptive weighted fusion strategy based on the Kalman filter is employed, which effectively suppresses the abrupt errors of UWB signals in Non-Line-of-Sight (NLOS) environments. Experimental results demonstrate that in a simulated mine roadway environment, the proposed method achieves an average localization error of 0.48 meters and reduces the maximum error to 0.65 meters. Compared to a standalone UWB system using Double-Sided Two-Way Ranging (DS−TWR), the localization accuracy is improved by 51.9%, and the path tracking smoothness is significantly enhanced. This method provides reliable technical support for dynamic personnel management and emergency rescue in the construction of smart mines.