Most existing fusion positioning methods for ultra-wideband (UWB) and pedestrian dead reckoning (PDR) ignore the correction of positioning errors in non line of sight (NLOS) environments. The methods use simple threshold division as the basis for NLOS environment judgment, which is largely related to the positioning scene and site size. In order to solve the above problems, a fusion positioning method for underground personnel considering NLOS environment is proposed. Firstly, UWB technology is used to calculate the position of underground personnel. After obtaining the preliminary position of personnel through the trilateral positioning algorithm, the least squares method is used to optimize the position. Polynomial fitting is used to achieve the fitting between the actual value and the measured value between the base station and the tag in the NLOS environment, reducing the ranging error in the NLOS environment and improving the positioning precision. Secondly, the PDR algorithm is used for gait recognition and analysis. The PDR algorithm uses gait data collected by inertial navigation sensors to update the target position through gait recognition, step size estimation, and direction estimation. Thirdly, the convolutional neural network (CNN) - long short term memory (LSTM) network is used to analyze the features of channel impulse response (CIR) and achieve line of sight (LOS)/NLOS recognition. It solves the problem of scene limitations in NLOS environment judgment. Finally, the fusion coefficient is determined based on the LOS/NLOS recognition results to achieve the fusion of UWB and PDR positioning results. The experimental results show that after polynomial fitting, the average ranging error of UWB is reduced by 0.59 m. The average accuracy of LOS/NLOS recognition is 95.3%, and the recall rate and F1 score are both above 90%, verifying that CNN-LSTM has good recognition performance. The average error of the fusion positioning method is 0.31 m, which is 1.57 m lower than UWB and 1.41 m lower than PDR.