Abstract
In recent years, Hand Gesture Recognition (HGR) devices have been designed to recognize gestures
in real time using machine-learning classifiers (MLCs). However, the performance of these classifiers
heavily relies on the tuning of their hyperparameters on real-time data. In this regard, this study
provides a Linear Population Size Reduction Success-History Adaptation Differential Evolution
(L-SHADE)-based optimized Extra Tree (ET) MLC framework for HGR. The study includes real-time
sEMG signals from two forearm muscles to capture six distinct hand gesture movements. To recognize
the gesture, this work employed ten MLCs. Among these ET classifier demonstrates the highest
accuracy without optimizing the hyperparameters. To further enhance performance, ten optimization
algorithms, along with the ET classifier, are considered, where the L-SHADE optimized ET framework
outperforms the others. To validate the proposed framework, a consistent system environment
has been used for both acquired and public datasets. On the acquired data, the mean accuracy
improves from 84.14% to 87.89% using ET with the L-SHADE optimization framework while the mean
computational time is reduced from 8.62 to 3.16 milliseconds. Similarly, the publicly available 15-hand
gesture classification dataset demonstrated a mean accuracy improvement of more than 3.0%.