Abstract: The posture detection of underground personnel can provide effective information for disaster warning and accident rescue. The postures of the underground personnel are complex and diverse and they are time series data. The existing human posture detection methods are difficult to process continuous related posture data. And the real-time performance is poor due to the complex algorithm and the need to configure an independent computer. In order to solve the above problems, a human posture detection method in coal mine based on improved long short term memory network(LSTM) is proposed. The pressure sensor and angle sensor are used to obtain the posture data of underground personnel, such as foot pressure, waist and leg angle, etc. The portable edge operation decision unit carried by the personnel can discriminate the posture. It can realize the real-time detection of five postures of underground personnel, such as standing, walking, bending, squatting (sitting) and lying down. In order to reduce the dimension of the original sampled data of human posture and improve the compute efficiency, LSTM is improved. The long short term memory sparse autoencoder(LSTMSA) is designed. The characteristics of the original sampled data is extracted by the sparse autoencoder(SA) to reduce the dimension, and then the human posture is detected by the LSTM. Human posture data are collected in the laboratory environment, and LSTMSA, LSTM and recursive neural network(RNN) are trained and tested respectively. The results show that under the same experimental settings and sampling data, the accuracy of LSTMSA for five kinds of human posture detection reaches more than 90%, which is close to LSTM and greater than RNN. The computing time of LSTMSA is shortened by more than 50% compared with LSTM, which meets the real-time requirements of human posture detection in coal mine.