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

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• Electromyography (EMG) signals offer potential for pre-emptive hand gesture recognition.
• Key parameters like sliding window size, overlap, and majority voting significantly impact accuracy.
• Previous research on these parameters is limited, especially with large subject groups.

Purpose of the Study:

• To investigate the influence of varying window and overlapping sizes on machine learning performance using a large EMG dataset.
• To develop and evaluate a novel multi-window scheme for improved gesture recognition accuracy compared to conventional single-window methods.

Main Methods:

• Utilized a large, publicly available EMG dataset from forty healthy subjects.
• Varied window sizes (50-500ms) and overlapping percentages (0-80%).
• Employed six machine learning algorithms (k-NN, LDA, Logistic Regression, Naïve Bayes, SVM, Random Forest) for six hand gesture classifications.

Main Results:

• Increased window size, overlapping size, and majority voting votes significantly improved classification accuracy (p < 0.05).
• Random Forest algorithm achieved the highest accuracy.
• The proposed multi-window scheme demonstrated a substantial improvement in overall accuracy over conventional majority voting.

Conclusions:

• Parameter optimization, particularly window size and voting strategies, is crucial for accurate EMG-based hand gesture recognition.
• The novel multi-window scheme offers a significant advancement for improving gesture recognition.
• This method holds promise for enhancing the control of prosthetic and exoskeleton devices.