Gesture Recognition Through Mechanomyogram Signals: An Adaptive Framework for Arm Posture Variability

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Signal Processing

Background:

• Hand gesture recognition is crucial for human-computer interaction.
• Classifying gestures across varying arm postures presents significant challenges due to dynamic muscle activity.
• Existing methods often rely on electromyogram (EMG) sensors requiring skin contact.

Purpose of the Study:

• To develop a robust hand gesture recognition system that addresses arm posture variability.
• To utilize a wearable mechanomyogram (MMG) device, eliminating the need for electrical skin contact.
• To evaluate the effectiveness of unsupervised domain adaptation for improving gesture classification accuracy across different postures.

Main Methods:

• Employed a wearable mechanomyogram (MMG) device for muscle activity sensing.
• Utilized Continuous Wavelet Transform (CWT) for feature extraction from MMG signals.
• Implemented Domain-Adversarial Convolutional Neural Networks (DACNN) with unsupervised domain adaptation for gesture classification.
• Compared DACNN performance against supervised classifiers across multiple arm postures.

Main Results:

• The proposed DACNN method demonstrated consistent improvement in classification accuracy compared to supervised methods across various arm postures.
• Achieved an average prediction accuracy of 87.43% for intra-posture and 64.29% for inter-posture classification of 5 hand gestures.
• Expanding the MMG segmentation window to 600 ms increased intra-posture accuracy to 92.32% and inter-posture accuracy to 71.75%.

Conclusions:

• The developed method effectively improves gesture recognition generalization across dynamic arm posture changes.
• Mechanomyogram (MMG) shows potential as a viable alternative sensor for gesture recognition, comparable to electromyogram (EMG).
• The system is suitable for non-laboratory usage, offering a user-friendly setup and high performance.