STCNet: Spatio-Temporal Cross Network with subject-aware contrastive learning for hand gesture recognition in surface...

Area of Science:

• Biomedical Engineering
• Machine Learning
• Human-Computer Interaction

Background:

• Inter-subject variability and environmental factors challenge surface electromyography (sEMG) based hand gesture recognition.
• Existing methods often lack robustness due to electrode shifts and muscle fatigue.
• Accurate sEMG gesture recognition is crucial for advanced prosthetics and human-computer interfaces.

Purpose of the Study:

• To introduce a novel deep learning architecture, Spatio-Temporal Cross Network (STCNet), for robust multi-subject sEMG hand gesture recognition.
• To address and mitigate the impact of inter-subject variability and environmental factors on recognition accuracy.
• To enhance the spatial and temporal feature extraction from sEMG signals.

Main Methods:

• Developed STCNet, a convolutional-recurrent architecture with a spatio-temporal block for feature extraction.
• Incorporated a rolling convolution technique to effectively capture spatial relationships from the sEMG measurement device's circular band structure.
• Proposed a subject-aware contrastive learning framework using subject and gesture labels to align vector space representations.

Main Results:

• STCNet demonstrated superior performance under aggregated conditions, outperforming existing methods.
• Achieved state-of-the-art results on benchmark sEMG datasets.
• Effectively managed and compensated for variability across different subjects.

Conclusions:

• STCNet offers a robust and effective solution for multi-subject sEMG hand gesture recognition.
• The proposed architecture and learning framework significantly improve recognition accuracy and reliability.
• This work advances the field of wearable sensor-based human-computer interaction.