An Edge-based Human Activity Recognition in Wearable Systems
Ethan Neal Capuchino, Kathleen Issandra Tuso, Ron Louis Nierva, Carl Lester Fabian, Nathaniel S. Orillaza, Marc Driz Rosales, John Richard E. Hizon · 2025
Human activity recognition (HAR) has seen wearable sensor applications in the healthcare industry. Accelerometer data can be attributed to different activities, including sitting, standing, going upstairs and downstairs, jogging, and walking. Although data is typically gathered from multiple wearable sensors on the body, this study aims to minimize sensor requirements by focusing on one wearable placement while retaining excellent performance. Machine learning (ML) allows for the training and deployment of neural network (NN) models that learn from raw data. This study, then, details a novel approach that uses the resource-constrained XIAO nRF52840 Sense, instead of larger edge devices, GPUs, or manufactured wristwatches, for the classification of 6 human activities in an edge-based HAR with TinyML. Following similar model architectures, the results show that NNs, instead of classical ML, can be used effectively, with quantized multilayer perceptron (MLP), 1D convolutional neural network (CNN), and temporal convolutional network (TCN) models achieving test accuracies of 89.39%, 90.56%, and 90.34%, and F1 Scores of 92%, 93%, and 92% respectively, with 1D CNN achieving the highest scores. MLP had the fastest inference time and lowest memory requirements, while showing comparable performance. The quantized NN models are deployed on the nRF52840 Sense, showing its ability as a small wearable device to not only record data, but also to perform live predictions efficiently. Furthermore, the study implements axis reduction to compare between all possible combinations of x,y,z accelerometer data. This research shows the feasibility of only two axes, instead of the full three, and still reaching up to 80.54% accuracy and 87% F1 score. Hence, this study provides a fair and comprehensive comparison between the models and different axes combinations in terms of accuracy, F1 score, inference time, and memory usage, and a demonstration of the nRF52840 Sense’s capability as a wearable edge device.