Abstract
Hand gesture recognition systems have become increasingly important due to
their use in fields such as prosthetics, rehabilitation, human-computer interaction,
and gaming. This research aims to create an advanced recognition system based on
High-Density surface Electromyography (HD-sEMG) signals. The system achieves
enhanced spatial and temporal resolution by capturing and processing biosignals
with high-density electrode arrays, allowing for more precise gesture identification.
Signal processing techniques are applied to extract meaningful features from
the raw surface Electromyography (sEMG) signals, including temporal, and time
frequency domains. A robust machine learning framework, combining traditional
classifiers, is employed to train the system on a diverse dataset of hand gestures.
Special attention is given to the fusion of sEMG and Force Myography (FMG)
signals to improve recognition accuracy and robustness against noise and artifacts.
The research emphasizes the challenges of handling HD-sEMG data and
optimizing machine learning models for real-time performance. Experimental
results demonstrate the system’s ability to accurately classify complex gestures
across various hand positions and muscle activation levels. The outcomes of this
study highlight the advances of integrating HD-sEMG and FMG signals for
advancing gesture recognition technologies, paving the way for innovative
applications in wearable technology and assistive devices. In particular, the study is
divided into two key parts: the first is the investigation of designing a suitable gesture
recognition system with simple algorithms to understand the effect of the chosen
features and window size on the efficiency of the proposed system. The second is to
improve the suggested system by merging the HD-sEMG signals with the FMG
ones, following the same procedure. Both parts were done considering a time domain
IVfeature extracted from the recording of HD-sEMG signals, and lastly the FMG
signals. The models were evaluated through the integral of several matrices to
investigate the performance of these systems.
Even if the algorithms chosen are characterized by their simplicity, a
significant statistical difference was found only with using the force signals. In
particular, the accuracy improved from (89.5 %) to (99.1 %). The comparison with
previous researchers' work confirms the robustness and reliability of our proposed
approach.
IV
