Integration of a Lightweight Customized 2D CNN Model to an Edge Computing System for Real-Time Multiple Gesture Recognition

Abstract

The human-machine interface (HMI) collects electrophysiology signals incoming from the patient and utilizes them to operate the device. However, most applications are currently in the testing phase and are typically unavailable to everyone. Developing wearable HMI devices that are intelligent and more comfortable has been a focus of study in recent times. This work developed a portable, eight-channel electromyography (EMG) signal-based device that can distinguish 21 different types of motion. To identify the EMG signals, an analog front-end (AFE) integrated chip (IC) was created, and an integrated EMG signal acquisition device combining a stretchy wristband was made. Using the EMG movement signals of 10 volunteers, a SIAT database of 21 gestures was created. Using the SIAT dataset, a lightweight 2D CNN-LSTM model was developed and specialized training was given. The signal recognition accuracy is 96.4%, and the training process took a median of 14 min 13 s. The model may be used on lower-performance edge computing devices because of its compact size, and it is anticipated that it will eventually be applied to smartphone terminals.