High-Density Surface EMG-Based Gesture Recognition Using a 3D Convolutional Neural Network.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Signal Processing

Background:

• High-density surface electromyography (HD-sEMG) is increasingly utilized for gesture recognition.
• Deep learning models, particularly Convolutional Neural Networks (CNNs), are applied to analyze sEMG data represented as images.
• Existing 2D CNNs struggle to capture temporal dynamics in sequential sEMG images.

Purpose of the Study:

• To introduce and evaluate a 3D CNN for HD-sEMG-based gesture recognition.
• To compare the performance of 3D CNNs against 2D CNNs in capturing spatio-temporal features.
• To assess the effectiveness of 3D CNNs in recognizing muscle electrical activity patterns over time.

Main Methods:

• Development of a 3D CNN model utilizing 3D kernels to process sequential sEMG images.
• Comparative analysis of 3D CNN and 2D CNN architectures on benchmark datasets (CapgMyo DB-a and CSL-HDEMG).
• Evaluation of model performance based on accuracy across different recognition window lengths (40 ms and 150 ms).

Main Results:

• The 3D CNN demonstrated superior performance compared to the 2D CNN across both datasets.
• Accuracy improvements for 3D CNN over 2D CNN ranged from 1% to 1.5% on CapgMyo DB-a.
• Significant accuracy gains of 15.3% and 18.6% were observed for 3D CNN on CSL-HDEMG at 40 ms and 150 ms window lengths, respectively.
• The 3D CNN achieved competitive results against established baseline methods.

Conclusions:

• 3D CNNs are effective in capturing both spatial and temporal information from sequential HD-sEMG images.
• The proposed 3D CNN approach offers enhanced accuracy for gesture recognition compared to 2D CNNs.
• This method shows promise for advanced applications in human-computer interaction and prosthetic control using sEMG signals.