On the use of unsupervised learning techniques in ultrasound-based upper limb movement recognition
Loading...
Date
2026-02-25
Authors
Κασουρίδου, Βασιλική
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Ultrasound has emerged as alternative modality for operating Human-Machine Interfaces, due to its ability to capture morphological muscle activity with high spatial resolution. Existing methods rely on supervised learning techniques but require careful design of handcrafted features for achieving strong intra-session performance. This thesis investigates whether unsupervised learning can replace or be comparable to handcrafted features. A one-dimensional convolutional autoencoder (1D CAE) is used to learn compact latent representations directly from A-mode ultrasound signals in an unsupervised manner. The learned features are constrained to the same dimensionality as handcrafted feature sets, enabling a fair comparison. Experiments are conducted with the only publicly available dataset in the field, namely the Ultrasound-based Adaptive Prosthetic Control (UltraPro) dataset. According to our findings, the feature representations of the CAE provide on par results with handcrafted feature methods while demonstrating reduced computational complexity. Overall, this work establishes a user-agnostic unsupervised convolutional autoencoder as an efficient feature extraction framework for gesture recognition tasks, reducing significantly computational complexity compared to classical handcrafted approaches while maintaining robust classification performance.
Description
Keywords
Hand gesture recognition, A-mode ultrasound, Unsupervised learning, 1D convolutional autoencoder, Feature extraction, Biosignal classification