A Fully Embedded Adaptive Real-Time Hand Gesture Classifier Leveraging HD-sEMG and Deep Learning

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Rehabilitation Technology

Background:

• Advanced prosthetic limbs require intuitive and precise control systems.
• Surface electromyography (sEMG) offers a non-invasive method for detecting muscle activity.
• Current systems often lack the fine gesture recognition needed for complex tasks.

Purpose of the Study:

• To develop a real-time fine gesture recognition system for multi-articulating hand prostheses.
• To enhance the accuracy and responsiveness of prosthetic hand control.
• To integrate an embedded artificial intelligence (AI) platform for on-device processing.

Main Methods:

• Utilized a custom 32-channel high-density surface electromyography (HDsEMG) flexible electrode array.
• Developed a frequency-time-space cross-domain preprocessing technique for muscle activation maps.
• Implemented an embedded convolutional neural network (CNN) for gesture classification on an AI platform.

Main Results:

• Achieved 98.15% accuracy in fine gesture recognition using a majority vote over 5 CNN inferences.
• Real-time recognition achieved within 100-200 ms.
• Demonstrated reliable, secure, and low-latency operation via edge computing.

Conclusions:

• The developed system shows significant promise for improving state-of-the-art commercial hand prostheses.
• Co-design of hardware, signal processing, and AI ensures a power-efficient and reliable embedded solution.
• Edge AI enables on-demand training, fine-tuning, and calibration for personalized control.