Deep Residual Networks for Human Activity Recognition based on Biosignals from Wearable Devices
Sakorn Mekruksavanich, Ponnipa Jantawong, Anuchit Jitpattanakul · 2022
Wearable inertial sensors are frequently employed in the area of human activity recognition (HAR), considering they provide the most meaningful time-series data available for non-visual datasets. Nevertheless, when just motion detectors are supplied, the recognition ability of individual complex activity detection is restricted. This study examines bio-signal data in order to improve the outcomes of complex HAR. Electrocardiogram (ECG) and photoplethysmogram (PPG) data are employed separately to identify complex HAR. We employed a benchmark bio-signal dataset (PPG-DaLiA) comprising ECG, PPG, and tri-axial accelerometer (3D-ACC) signals obtained from 15 people as they performed a series of uncomplicated and complex everyday activities. After incorporating early fusion conditions, we trained and evaluated deep learning models using 5-fold cross-validation. Our findings suggest that integrating biosignal characteristics with the 3D-ACC improves the classifier's effectiveness F1-scores by 2.17% and 2.27%, respectively, for ECG and PPG.