Real-time Gait Trajectory Prediction Based on Convolutional Neural Network with Multi-head Attention

Abstract

The lower limb exoskeleton can effectively improve the ability of users. Accurate gait trajectory prediction can enhance the effect of lower limb exoskeletons. The current gait trajectory prediction methods have the disadvantages of insufficient prediction accuracy and long calculation time. This paper proposes a convolutional neural network model with multi-head attention to predict the gait trajectory. Compared with the widely used recurrent neural network model, the model we proposed in this paper can predict gait trajectory with higher accuracy and shorter calculation time. The error is reduced by up to 21.3% and the average calculation time is reduced by 62.8%. Based on the proposed model, we design a controller of the lower limb exoskeleton to achieve better effects in mixed gait and eliminate the delay.