Continuous Gait Phase Recognition and Prediction Using LSTM Network Across Walking Speeds and Inclinations

Abstract

Real-time gait phase estimation is essential for the safe and accurate control of lower-limb exoskeletons. In this paper, we evaluate the performance of a Long Short-Term Memory (LSTM) recurrent neural network as a method for prediction of gait phases using two inertial measurement units (IMUs) placed on the subject’s thigh and shank. We evaluated the prediction of gait phases across different walking speeds (0. 27m/s to 1.35 m/s), inclinations (0° to 2°), and up to 1.56 s ahead of time, on two able-bodied participants. LSTM-based classifier was trained with the labeled gait data to recognize the current gait phase and to predict the continuous gait phase in the next gait cycle. We present preliminary results measured from two unimpaired participants. Our results suggest the proposed method has the potential to achieve a stable gait phase prediction using 40 sample points from time-history data. The algorithm achieved 85.6% and 73.3% average accuracy during walking experiences in structural and over-ground environments. It achieved the lowest RMSE (Root Mean Square Error) of 0.143 for the current gait phase recognition out of the 40 continuous estimated gait phases.