Abstract: As the traditional pedestrian dead reckoning (PDR) algorithm being used for underground personnel positioning, the positioning error gradually increases due to the cumulative error of the step frequency detection, step length estimation and heading estimation phases. Moreover, the commonly used error correction methods such as zero speed correction, heading drift elimination and gait signal optimization cannot change the inherent defects of the PDR algorithm, and the positioning accuracy needs to be improved. It is proposed to use an improved peak detection method to achieve step frequency detection in the PDR algorithm, and to achieve step length estimation based on a deep recurrent neural network (RNN). The improved PDR algorithm is used for underground personnel positioning. Firstly, cell phone accelerometer, gyroscope and magnetometer are used to obtain pedestrian movement data. Secondly, the improved peak detection method is used to obtain the average step frequency in a fixed time interval. The average step frequency, the time interval, acceleration and acceleration variance are used as features to be input to the trained deep RNN model for step length estimation. Finally, the estimated heading angle is added to predict the current position of the personnel. The experimental results show that the improved PDR algorithm for underground personnel positioning has a relative error of 5.9% in predicting the test set data and has a relative error of 1.6%-3.9% in positioning the actual test route. The relative error is smaller than the positioning error of the traditional PDR algorithm and the proposed PDR algorithm has improved the accuracy of underground personnel positioning effectively.