Abstract

Hand Gesture Recognition (HGR) using surface Electromyogram (sEMG) signals can be considered as one of the most important technologies in making efficient Human Machine Interface (HMI) systems. In particular, sEMG-based hand gesture has been a topic of growing interest for development of assistive systems to improve the quality of life in individuals suffering from amputated limbs. Generally speaking, myoelectric prosthetic devices work by classifying existing patterns of the collected sEMG signals and synthesizing intended gestures. While conventional myoelectric control systems, e.g., on/off control or direct-proportional, have potential advantages, challenges such as limited Degree of Freedom (DoF) due to crosstalk have resulted in the emergence of data-driven solutions. More specifically, to improve efficiency, intuitiveness, and the control performance of hand prosthetic systems, several Artificial Intelligence (AI) algorithms ranging from conventional Machine Learning (ML) models to highly complicated Deep Neural Network (DNN) architectures have been designed for sEMG-based hand gesture recognition in myoelectric prosthetic devices. In this thesis, we, first, perform a literature review on hand gesture recognition methods and elaborate on the recently proposed Deep Learning/Machine Learning (DL/ML) models in the literature. Then, our utilized High-Density sEMG (HD-sEMG) dataset is introduced and the rationales behind our main focus on this particular type of sEMG dataset are explained. We, then, develop a Vision Transformer (ViT)-based model for gesture recognition with HD-sEMG signals and evaluate its performance under different conditions such as variable window sizes, number of electrode channels, and model's complexity. We compare its performance with that of two conventional ML and one DL algorithm that are typically adopted in this domain. Furthermore, we introduce another capability of our proposed framework for instantaneous training, which is its ability to classify hand gestures based on a single frame of HD-sEMG dataset. Following that, we introduce the idea of integrating the macroscopic and microscopic neural drive information obtained from HD-sEMG data into a hybrid ViT-based framework for gesture recognition, which outperforms a standalone ViT architecture in terms of classification accuracy. Here, microscopic neural drive information (also called Motor Unit Spike Trains) refers to the neural commands sent by the brain and spinal cord to individual muscle fibers and are extracted from HD-sEMG signals using Blind Source Separation (BSP) algorithms. Finally, we design an alternative and novel hand gesture recognition model based on the less-explored topic of Spiking Neural Networks (SNN), which performs spatio-temporal gesture recognition in an event-based fashion. As opposed to the classical DNN architectures, SNNs are of the capacity to imitate human brain's cognitive function by using biologically inspired models of neurons and synapses. Therefore, they are more biologically explainable and computationally efficient.