A Deep Learning-Based EIT System for Robust Gesture Recognition Under Confounding Factors

Area of Science:

• Biomedical Engineering
• Human-Computer Interaction
• Signal Processing

Background:

• Electrical impedance tomography (EIT) offers low-cost, high-temporal-resolution gesture recognition for human-machine interaction.
• Consistent EIT performance under real-world interference, such as varying contact impedance and limb position, remains a significant challenge.

Purpose of the Study:

• To develop a novel method for robust EIT-based gesture recognition that mitigates the impact of confounding factors.
• To enhance the stability of EIT measurements by addressing contact impedance variations.
• To improve the accuracy and reliability of gesture classification in dynamic and non-ideal conditions.

Main Methods:

• Designed a novel EIT armband with a two-layer electrode structure to stabilize contact impedance against changes in applied force.
• Utilized equivalent circuit analysis to inform the armband design for improved impedance stability.
• Developed a spatial-temporal fusion network, Fold Atrous Spatial Pyramid Pooling-Gated Recurrent Unit (FASPP-GRU), for gesture classification.

Main Results:

• The proposed two-layer electrode effectively maintained stable contact impedance despite variations in skin contact force.
• The FASPP-GRU network achieved 80% accuracy in gesture recognition under confounding factors like limb position changes and dynamic muscle state variations.
• The developed system demonstrated superior performance compared to conventional classifiers in challenging environments.

Conclusions:

• The novel EIT armband design and FASPP-GRU network provide a robust solution for EIT-based gesture recognition.
• The system effectively alleviates the impact of confounding factors, enabling reliable performance in real-world scenarios.
• With a fast inference time of 87 μs, the proposed method holds significant promise for real-time human-machine interaction applications.