Abstract
We propose a real-time activity recognition system based on data from multiple wearable inertial sensors. They are worn at the user's right wrist, waist, and right ankle to collect acceleration and angular velocity data, which are then transmitted via Bluetooth to a computer. The data are used to train a convolutional neural network (CNN) model to recognize 13 types of daily activities. The trained model has been ported to Tensorflow Lite running on Raspberry Pi to enable edge processing. The latency for recognizing the first motion takes 2.6 seconds, and subsequent ones are 1.325 seconds. Experimental results show that among the six participants, our model achieves high accuracy of 98.62% and 99.84% for leave-one-out cross-validation and 10-fold cross-validation, respectively.