Real-Time Implementation of Tiny Machine Learning Models for Hand Motion Classification

Abstract

This paper investigates the design, implementation, and assessment of different embedded neural network models for hand motion classification. Using an Inertial Measurement Unit (IMU) sensor, a dataset of five distinct hand motion classes has been collected from the Arduino nano BLE 33 targeting hand motion analysis. Two different machine learning models namely Convolutional Neural Network (CNN) and Multilayer Perceptron (MLP) are implemented and compared in terms of classification accuracy and hardware complexity. Experimental results show identical classification accuracy of 100% for the two models. After deployment, results show that the MLP network exhibited the lowest processing time and RAM usage, taking 4 ms and 4.0 K, respectively. In contrast, the CNN model needs 313ms and utilizes 15.9 K of RAM. Moreover, the MLP model also had the lowest flash usage with 25.3 K, while the CNN model used 67.7 K.