A Review of the Current Methods and Challenges of Facial EMG Based Gesture Recognition

Abstract

Advances in technology related to the Internet-of-Things and wearable health technology has lead new research in the field of surface EMG based gesture recognition in different fields of study such as medical rehabilitation, EMG controlled prosthetic limbs, and human-computer interaction for people with disabilities. Current research pertaining to EMG based gesture recognition focuses on a variety of different muscles across the human body and as such, different gestures as well. This has made it difficult to evaluate studies against each other because of the different study set ups used across different fields of research and the inability to combine datasets or compare them due to these differences. This thesis aims to study a publicly available Facial EMG dataset to accomplish the following goals. Compare gesture recognition performance between different feature sets and classifiers by comparing amplitude time-domain based features that are used in literature to discrete wavelet transform based time-frequency features and compare the performance of these two feature sets. Highlight the challenges involved in creating a facial gesture recognition model that is able to identify gestures with high accuracy. Compare the results presented in this thesis to related works and highlight how current research may present overly optimistic results concerning classification accuracy. Present a new alternative lightweight classification model, which requires no feature engineering, based on a convolution neural network, which is supplied discrete wavelet transform coefficients derived from Facial EMG signals as an input.