A Multi-Window Majority Voting Strategy to Improve Hand Gesture Recognition Accuracies Using Electromyography Signal

The electromyography (EMG) signal has great potential to determine the hand gestures automatically before the actual move begins. However, parameters of the sliding window along with the EMG signal, such as window size and overlapping size, as well as the number of votes in post-processing, such as majority voting, can significantly influence the gesture recognition accuracy. These phenomena have been investigated only in a few studies on a small number of subjects. The aim of this study is two-fold. First, to determine the influence of different window and overlapping sizes on the machine-learning performance using a large database consists of forty healthy subjects. Second, to develop a novel multi-window scheme to accumulate a large number of votes compared to the conventional single-window majority voting to improve gesture recognition accuracy. A large publicly available EMG dataset was used in this study. The window and overlapping sizes were varied between 50ms and 500ms, and between 0% and 80%, respectively. Six machine-learning algorithms, including k-Nearest Neighbor, Linear Discriminant Analysis, Logistic Regression, Naive Bayes, Support Vector Machine, and Random Forest were used to classify six different hand gestures. Results show that the overall classification accuracy can be substantially improved by increasing the window size, overlapping size, and the number of votes in the majority voting strategy (p < 0.05). The maximum accuracy was achieved using the Random Forest algorithm. The two-way repeated measure analysis of variance shows that the proposed multi-window scheme substantially improved the overall accuracy of the machine-learning algorithms compared to the conventional majority voting. The proposed method can be instrumental for efficient control of prosthetic or exoskeleton devices.