A Hybrid EMG Model for the Estimation of Multijoint Movement in Activities of Daily Living

Accurately identifying human's intent of motion from electromyography (EMG) signals is the key to implement EMG-based HRI (Human-Robot Interface) systems. Human's intent of motion includes motion modes and continuous movement variables. In this paper, a hybrid EMG-to-motion model is constructed by combining a classification model and a regression model. Based on a proper division for joints, the classification model is utilized to recognize the motion modes of 'small' joints; meanwhile, the regression model is utilized to estimate the continuous movement variables of 'big' joints. Furthermore, a Bayesian network (BN) model, which sufficiently employs context information of a task, is also involved into the hybrid model to improve its performances for motion estimation. Experiments have been conducted with three subjects to demonstrate the feasibility of the proposed methods. In these experiments, the motion modes of hand and wrist, and the continuous elbow angles are estimated with sEMG signals considering a 'drinking' task. Finally, an upper limb prosthetic is controlled to simulate human's movement in a 'drinking' task.