A Lightweight Deep Learning Network for Sensor-based Human Activity Recognition using IMU sensors of a Low-Power Wearable Device

Abstract

Human Activity Recognition (HAR) is an intriguing approach to healthcare monitoring that necessitates the ongoing utilization of wearable sensors to capture everyday activities. The most advanced studies using wearable devices have used a cloud computing paradigm that categorized data at remote systems. This method has drawbacks in terms of energy consumption, latency, and privacy. As a result, we adhere to a low-power computing structure in which wearable device solutions achieve an acceptable level of performance while being energy and memory-efficient. This study proposes a lightweight deep learning model termed a gate recurrent unit (GRU) network for energy-efficient HAR appropriate for low-power wearable devices. The introduced GRU network is evaluated against a benchmark dataset named the w-HAR dataset, including essential deep learning networks. The proposed GRU surpasses baseline deep learning networks in terms of overall accuracy, with a score of 95.160%.