Abstract: The traditional pedestrian dead reckoning (PDR) algorithm has low positioning precision due to the accumulated errors of step size and heading, which can not meet the requirements of precise positioning of underground personnel. In order to solve the problem, a PDR algorithm for precise positioning of underground personnel based on long short-term memory (LSTM) personalized step size estimation is proposed. Firstly, the acceleration and gyroscope inertia information in the movement of underground personnel is collected, and the movement distance of each step is calculated to construct step size data. The LSTM model of personalized step size estimation of the underground personnel is obtained through off-line training. Secondly, in the online prediction stage, the underground personnel movement data such as acceleration, gyroscope and geomagnetism are collected in real-time through the mine intrinsically safe smart phone. The underground personnel movement step and step size of each step are obtained by using the step detection algorithm and personalized step size estimation model respectively. The heading angle is obtained by using the Kalman filtering and heading estimation algorithm. Finally, the current position of underground personnel is predicted according to step size estimation and heading angle. In Inner Mongolia Ordos Gaotouyao Coal Mine, the underground personnel movement data is collected for testing, and the results show as follows. The PDR algorithm for precise positioning of underground personnel based on LSTM personalized step size estimation has a step detection precision of 96.5% and a step size prediction precision of 90%. The algorithm has a relative positioning error of 2.33% in the real underground environment, which improves the personnel positioning precision in coal mine.