CNN-based Human Activity Recognition on Edge Computing Devices

Human Activity Recognition (HAR) is a research area that involves wearable devices integrating inertial and/or physiological sensors to classify human actions and status across various application domains, such as healthcare, sports, industry, and entertainment. However, executing HAR algorithms on remote devices or the cloud can lead to issues such as latency, bandwidth requirements, and energy consumption. Transitioning towards Edge HAR can be a more effective and versatile solution, overcoming the challenges of traditional HAR techniques. We present a novel HAR model for computation on edge devices: we design a Convolutional Neural Network (CNN) Deep Learning approach and compare its performance with cloud-computing HAR models. The paper is accompanied by a self-collected dataset. The experiments on this dataset demonstrate that the proposed edge computing model achieves promising results (\geq 92 %) in terms of Precision, Recall, and Fl-score. Furthermore, the model exhibits significantly reduced latency, with only 117 ms, and utilizes minimal memory, with a peak of 18.8 Kb RAM and 956 Kb Flash memory.