Edge AI-Enabled EMG Finger Gesture Recognition on Low-Cost Microcontroller

Abstract Each hand and finger movement causes Electromyogram (EMG) signal in relevant muscle groups. This EMG signal is used in a variety of applications like prosthetic arm operation which involves classification of hand and finger movements. The algorithms used for classification are usually executed in high performance computer environments. The goal of this study is implementing the classification model in a microcontroller platform, which is characterized by its cost effectiveness. The first step was to collect data and create a dataset of five different finger movements. Collected data have undergone preprocessing such as filtering and detection of onset and offset points of signal. Time, frequency, and wavelet features were extracted from the final data set. After feature selection process, an artificial neural network model was trained. The resulting model was executed independently on computer, single-board computer, and microcontroller platforms. The model’s accuracy and classification times on these platforms were then compared. As a result, an artificial intelligence model that utilizes EMG signals for the classification of finger movements has been obtained. The resulting model achieved 98% test accuracy on all platforms. The model completed classification in 0.0044 ms in computer environment and 0.16 ms in microcontroller environment. This model has been successfully implemented on a low-cost microcontroller with significantly lower specifications compared to a computer while achieving high classification performance with a notably low latency. Furthermore, a shorter classification time has been achieved compared to other microcontroller-based studies in the literature. This work represents a remarkable advancement in future real-time robotic/prosthetic hand control.