Real-Time Multiple Gesture Recognition: Application of a Lightweight Individualized 1D CNN Model to an Edge Computing...

Area of Science:

• Biomedical Engineering
• Human-Machine Interface (HMI) Technology
• Wearable Technology

Background:

• Current human-machine interfaces (HMIs) primarily detect electrophysiological signals but remain largely in the testing phase.
• Advancements in wearable HMI devices focus on enhancing intelligence and user comfort.
• Electromyography (EMG) signal detection is crucial for developing intuitive control systems.

Purpose of the Study:

• To design and develop an intelligent, wearable eight-channel EMG-based system for recognizing 21 distinct gestures.
• To create a practical HMI system with potential for real-world application and public accessibility.
• To establish a comprehensive EMG gesture signal database for system training and validation.

Main Methods:

• Fabrication of an integrated EMG signal acquisition device with an elastic armband, incorporating a custom analog front end (AFE) integrated chip (IC).
• Development of the SIAT database by collecting EMG gesture signals from 10 volunteers, encompassing 21 different gestures.
• Construction and individualized training of a lightweight 1D Convolutional Neural Network (CNN) model using the established SIAT database.

Main Results:

• The developed system achieved a maximum EMG gesture recognition accuracy of 89.96%.
• The average model training time was recorded at 14 minutes and 13 seconds.
• The lightweight CNN model demonstrates suitability for deployment on lower-performance edge computing devices.

Conclusions:

• The intelligent wearable EMG system offers a high degree of accuracy in gesture recognition.
• The system's compact design and efficiency make it suitable for integration into resource-constrained edge devices.
• Future applications are envisioned for smartphone terminals, enhancing mobile HMI capabilities.