A User-Adaptive Online Learning Approach of Real-Time Gait Phase Detection for Walking Assistance

Abstract

This paper presents a novel real-time gait phase detection method for wearable gait assist robotics such as lower limb exoskeletons. In order to achieve individualization and user-adaptive gait assistance, we use an online learning approach to incrementally learn and update gait parameters for a six states Hidden Markov Model. The advantage of this method is to avoid the manually parameter calibration process. The gait phase detector in this paper can automatically adapt to the wearer’s gait changes, such as in a rehabilitation process. To evaluate the method, the experiments of level walking at different speeds of five healthy subjects were carried out. Results show that the method we proposed can realize stable real-time gait phase detection of normal gait without any offline training. The average time error is 0.064s and the average detection accuracy is 94.6%.