Spatially Integrating sEMG and Hand Landmarks for Continuous Hand Pose Inference
Xizhe Huang · 2023
Surface Electromyography (sEMG)-based gesture classification holds significant promise for applications in prosthetic limb control and human rehabilitation. While static gesture classification has received extensive research attention, there remains a limited exploration of continuous gestures based on sEMG. This study introduces a comprehensive approach, employing a complete Recurrent Neural Network Autoencoder (RNN Autoencoder) architecture for training on gesture data. By incorporating positional encoding and a Transformer core with a self-attention mechanism, the model predicts the potential representation of right-hand gestures through sEMG eigenvalue regression from the right forearm. The prediction results are subsequently reconstructed and visualized using the RNN Autoencoder decoder. Our research shows that it is effective to use sEMG for continuous gesture prediction through the Transformer model, and proves the generalization of the model in different actions.