Position-aware Human Activity Recognition with Smartphone Sensors based on Deep Learning Approaches
Sakorn Mekruksavanich, Anuchit Jitpattanakul · 2023
Numerous research studies on human activity recognition (HAR) have utilized smartphone sensors, such as accelerometers, gyroscopes, and magnetometers. The accelerometer, in particular, has been considerably explored as the primary sensor in HAR research. Recently, researchers have integrated different smartphone sensors to enhance recognition interpretation. Nevertheless, it is still challenging to determine the optimal usage of these sensors, either individually or in combination, for better identification effectiveness in various motion situations. In this article, we explore how different motion sensors behave during the activity recognition process, using various deep learning (DL) approaches to classify physical actions based on smartphone sensor data. We evaluate the DL models using a publicly benchmarked dataset of data collected from ten participants performing eight movements while carrying smartphones in different situations. Our experimental results reveal that, except for the magnetometer, each sensor can play a leading role in HAR, depending on the recognized activity, body position, data features employed, and classification approach. Additionally, we find that the combination of sensors only improves overall recognition interpretation when their individual performances are low.