ViT-HGR: Vision Transformer-based Hand Gesture Recognition from High Density Surface EMG Signals

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Deep Learning (DL) models for surface Electromyogram (sEMG) based hand gesture recognition face challenges with sparse signals, large training times, and data requirements.
• Existing DL models often require data augmentation or transfer learning due to their complexity and memory constraints.

Purpose of the Study:

• To investigate and design a Vision Transformer (ViT) based architecture for hand gesture recognition using High Density sEMG (HD-sEMG) signals.
• To develop a framework that overcomes the limitations of existing DL models, enabling accurate recognition from scratch without extensive data preparation.

Main Methods:

• A novel Vision Transformer-based Hand Gesture Recognition (ViT-HGR) framework was designed and implemented.
• The architecture leverages the attention mechanism of transformers for improved input parallelization and feature extraction from HD-sEMG signals.
• The framework was evaluated on a dataset of 65 isometric hand gestures using a 64-sample window size.

Main Results:

• The ViT-HGR framework achieved an average test accuracy of 84.62 ± 3.07% on the HD-sEMG dataset.
• The model demonstrated high performance without requiring data augmentation or transfer learning.
• The framework utilizes a compact architecture with only 78,210 learnable parameters.

Conclusions:

• The proposed ViT-HGR framework offers an efficient and accurate solution for hand gesture recognition from HD-sEMG signals.
• The compact nature of the ViT-based model highlights its potential for practical applications, particularly in prosthetic control.
• This research presents a novel approach to sEMG-based gesture recognition, advancing the field of intelligent prosthetics.