TinyML for Real-Time Embedded HD-EMG Hand Gesture Recognition with On-Device Fine-Tuning

Area of Science:

• Biomedical Engineering
• Artificial Intelligence
• Wearable Technology

Background:

• Myoelectric prosthetics often rely on complex, external hardware for control.
• Real-time processing of high-density surface electromyography (HD-sEMG) signals is crucial for intuitive prosthetic control.
• Integrating artificial intelligence (AI) directly into prosthetic devices can enhance functionality and reduce costs.

Purpose of the Study:

• To develop a fully embedded wireless platform for real-time hand gesture recognition using TinyML and HD-sEMG.
• To evaluate the performance of the Coral Tensor Processing Unit (TPU) accelerator for on-device inference.
• To explore the effectiveness of on-device versus cloud-based model fine-tuning and data quantization techniques.

Main Methods:

• Developed an embedded wireless platform using off-the-shelf components and the Coral TPU accelerator.
• Implemented TinyML for real-time hand gesture recognition utilizing 64-channel HD-sEMG data.
• Investigated on-device and cloud-based model fine-tuning strategies.
• Explored 8-bit data quantization techniques to optimize hardware compatibility and performance.

Main Results:

• Achieved a general inference time of 2.96 ms, demonstrating the TPU's suitability for real-time tasks.
• On-device fine-tuning improved gesture recognition accuracy by up to 36.15% in intersession tests, comparable to cloud-based methods.
• 8-bit quantization maintained or slightly improved performance, with a best-case improvement of 0.96% over unquantized data.
• The platform offers a cost-effective, self-sufficient solution for AI integration in prosthetics.

Conclusions:

• The developed embedded platform provides a robust foundation for on-device HD-sEMG based hand gesture recognition.
• This approach offers a more accessible and practical solution for myoelectric prosthetic control.
• The findings support the integration of TinyML and efficient hardware accelerators for advanced wearable AI applications.