Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

EMG subspace alignment and visualization for cross-subject hand gesture classification

Martin Colot, Cédric Simar, Mathieu Petieau, Ana Maria Cebolla Alvarez, Guy Cheron, Gianluca Bontempi

TL;DR

This work tackles the cross-subject generalization gap in EMG-based hand gesture recognition by leveraging unsupervised domain adaptation on a 14-subject dataset. It shows that pooling sources without adaptation yields limited accuracy, but identifying a common low-dimensional subspace using cosine KPCA and aligning it to a target subject via Subspace Alignment markedly improves cross-subject performance, reaching 79.5% in leave-one-subject-out tests. The combination of dimensionality reduction, visualization, and alignment provides practical insights and a scalable path toward faster calibration for EMG-based BCIs, though a gap to intra-subject performance remains. Limitations include the simplicity of gestures and the modest dataset, with future work aimed at expanding subjects, exploring supervised subspace methods, and adopting more advanced alignment techniques to better exploit source variability.

Abstract

Electromyograms (EMG)-based hand gesture recognition systems are a promising technology for human/machine interfaces. However, one of their main limitations is the long calibration time that is typically required to handle new users. The paper discusses and analyses the challenge of cross-subject generalization thanks to an original dataset containing the EMG signals of 14 human subjects during hand gestures. The experimental results show that, though an accurate generalization based on pooling multiple subjects is hardly achievable, it is possible to improve the cross-subject estimation by identifying a robust low-dimensional subspace for multiple subjects and aligning it to a target subject. A visualization of the subspace enables us to provide insights for the improvement of cross-subject generalization with EMG signals.

EMG subspace alignment and visualization for cross-subject hand gesture classification

TL;DR

This work tackles the cross-subject generalization gap in EMG-based hand gesture recognition by leveraging unsupervised domain adaptation on a 14-subject dataset. It shows that pooling sources without adaptation yields limited accuracy, but identifying a common low-dimensional subspace using cosine KPCA and aligning it to a target subject via Subspace Alignment markedly improves cross-subject performance, reaching 79.5% in leave-one-subject-out tests. The combination of dimensionality reduction, visualization, and alignment provides practical insights and a scalable path toward faster calibration for EMG-based BCIs, though a gap to intra-subject performance remains. Limitations include the simplicity of gestures and the modest dataset, with future work aimed at expanding subjects, exploring supervised subspace methods, and adopting more advanced alignment techniques to better exploit source variability.

Abstract

Electromyograms (EMG)-based hand gesture recognition systems are a promising technology for human/machine interfaces. However, one of their main limitations is the long calibration time that is typically required to handle new users. The paper discusses and analyses the challenge of cross-subject generalization thanks to an original dataset containing the EMG signals of 14 human subjects during hand gestures. The experimental results show that, though an accurate generalization based on pooling multiple subjects is hardly achievable, it is possible to improve the cross-subject estimation by identifying a robust low-dimensional subspace for multiple subjects and aligning it to a target subject. A visualization of the subspace enables us to provide insights for the improvement of cross-subject generalization with EMG signals.
Paper Structure (9 sections, 5 figures)

This paper contains 9 sections, 5 figures.

Table of Contents

  1. Introduction
  2. Related Work
  3. The experimental setting
  4. Baseline classifiers
  5. Dimensionality reduction for cross-subject learning
  6. Visualization of the low-dimensional subspace
  7. Subspace alignment for domain adaptation
  8. Conclusion
  9. Acknowledgements

Figures (5)

  • Figure 1: (a) Intra-subject classification accuracy for different dimensionality reduction algorithms and an increasing number of dimensions of the projection space. (b) Classification accuracy for an increasing number of dimensions of the KPCA subspace.
  • Figure 2: (a) KPCA projections of the samples from all subjects in a common subspace. (b) t-SNE visualization of the samples from all subjects in the common KPCA subspace
  • Figure 3: Comparison of the classification accuracy obtained with KPCA SA and with the other strategies for an increasing number of dimensions of the KPCA subspace.
  • Figure 4: Average rank of the different classifiers (the lower, the better). Classifiers that are not significantly different are connected (at p = 0.05 found by a Nemenyi test demvsar2006statistical).
  • Figure 5: Comparison of the results from the different models