Human hand movement recognition using infinite hidden Markov model based sEMG classification

Abstract

Hand movement recognition based on surface electromyography (sEMG) is challenging because sEMG signals are stochastic, noisy, and difficult to model and have limited datasets. This work improves the recognition accuracy using a small number of sEMG sequences and achieves the negative lag recognition for online applications. To handle the sEMG stochasticity, we applied the hidden Markov model (HMM) to decode the action primitives and characterise the transitions between them. On the challenging issue of selecting the number of hidden states, the parametric HMM was generalised to an infinite model by the hierarchical Dirichlet process, and a Gibbs sampling framework was implemented with the No-U-Turn sampler (NUTS) embedded to learn the parameters of the infinite HMM. For the recognition with negative lag, a generative classifier was built by all hand movement models and then was used with the online probability update. In the experiments of classifying 17 different hand movements, the proposed infinite hidden Markov model learned the number of states from data automatically and the generative classifier achieved an accuracy of 98.85%, 1.5% higher than that by parametric HMMs. The proposed recognition method can be applied to develop more reliable and efficient interfaces for prosthetics, rehabilitation, and robot teleoperation.