BioPoint: Single-site, Multi-sensor Compound Gesture Recognition

Authors

  • Félix Chamberland
  • Xavier Isabel
  • Evan Campbell
  • Gabriel Gagné
  • Benoit Gosselin
  • Erik Scheme
  • Gabriel Gagnon-Turcotte
  • Ulysse Côté-Allard

DOI:

https://doi.org/10.57922/mec.2505

Abstract

EMG-based gesture recognition tasks have received a lot of attention in recent years, mostly focused on multi-channel EMG sensors, leading to issues in ease of use and computational requirements of such systems. The present study leveraged the BioPoint, a smartwatch-like device, to proceed to multi-sensor deep-learning based gesture recognition. 3 hand gestures in 3 wrist orientations were targeted by measuring the EMG, PPG and IMU waveforms on able-bodied subjects (n=10). Preprocessing and feature extraction allowed the modalities to be used in a two-head neural network trained for the simultaneous classification of hand gestures and wrist rotation. During evaluation, the model obtained an average classification accuracy for hand gestures of 83.5 ± 12.4% and 94.3 ± 9.7% for wrist position. Overall, this study showed the potential of single-site, multi-sensor approaches for compound gesture recognition.

Downloads

Published

2024-08-15

How to Cite

[1]
F. Chamberland, “BioPoint: Single-site, Multi-sensor Compound Gesture Recognition”, MEC Symposium, Aug. 2024.

Conference Proceedings Volume

Section

Myoelectric Control Algorithms