Improving robustness against electrode shift of sEMG based hand gesture recognition using online semi-supervised learning

Abstract

Electrode shift of a prosthetic device is one of most challengeable problems in surface Electromyography (sEMG) based hand gesture recognition. Electrode shift is usually caused by repositioning, donning or doffing of a prosthetic device. Accuracy of gesture recognition may significantly drop since a pattern of collected signals may change after electrode shift. Although re-training a recognition system after every reposition is able to maintain accurate recognition, collecting labeled samples is inconvenient to users. In this paper, we apply an online semi-supervised learning in which a classifier is trained with a small amount of labeled samples and then is updated with unlabeled samples online to hand gesture recognition. A well-known online semi-supervised learning algorithm, online multi-channel semi-supervised growing neural gas (OSSMGNG) algorithm, is used in this preliminary study. OSSMGNG is compared with an intuitive method which learns from the initial label training set only in experiments. The data is collected from able-bodied individuals across three days for experiments. The results indicate OSSMGNG achieves a higher classification accuracy than others. It suggests that the online semi-supervised learning algorithm enhances robustness of hand gesture identification against electrode shift.

Lab Members

Patrick Chan

Associate Professor, Vice Dean

Patrick Chan works on machine learning, deep learning, image processing, adversarial learning, and secure machine learning.