Available at: https://digitalcommons.calpoly.edu/theses/3213
Date of Award
12-2025
Degree Name
MS in Electrical Engineering
Department/Program
Electrical Engineering
College
College of Engineering
Advisor
Xiao-Hua (Helen) Yu
Advisor Department
Electrical Engineering
Advisor College
College of Engineering
Abstract
Gesture recognition using wrist-based Electromyography (EMG) presents the opportunity to push the boundaries of hand tracking technology while eliminating many limitations of current solutions. However, EMG signals are diverse and difficult to decode. Previous works have opted to reduce the difficulty of the problem by only classifying a small number of gestures, training participant and session specific models, using many electrodes, and/or using large time windows for non-real-time operation.
With the goal of advancing more widely applicable solutions, this thesis develops real-time, participant and session generalized, gesture recognition deep learning models by leveraging the GRABMyo dataset. This dataset has many participants, multiple sessions, and many gestures compared to other datasets, making it ideal for studying the robustness of gesture recognition models.
Unlike most previous works, the proposed models are trained using data of only six electrodes located on the wrist and evaluated with unseen participants and/or unseen sessions. Computer simulation results show that the proposed CNN achieves an average of 82.90% accuracy on 17 gestures while having 10 times fewer trainable parameters than similar existing models, making it more suitable for real-time implementation. The proposed TCN model performs more robustly than CNN on unseen participants, demonstrating a better generalization property and reducing the need to retrain the model for new users.
Included in
Bioelectrical and Neuroengineering Commons, Biomedical Commons, Biomedical Devices and Instrumentation Commons, Other Computer Engineering Commons, Other Electrical and Computer Engineering Commons