AUT Human Activity Recognition (HAR)

Abstract

Human Activity Recognition (HAR) datasets are essential for developing and evaluating wearable sensing systems for health monitoring and activity analysis. This dataset comprises synchronized bioimpedance and inertial sensor recordings collected from 20 healthy adults (13 males and 7 females, aged 18 years and above) while performing eight Activities of Daily Living (ADLs): sitting and typing, sitting and writing, sitting and eating, squatting and lifting a load, toe standing, standing and twisting the hand, standing and cleaning, and walking. Data were acquired using two custom AUT Bio-Bands: a forearm band incorporating eight tetrapolar bioimpedance channels and a 9-axis inertial measurement unit (IMU), and a leg band incorporating four tetrapolar bioimpedance channels and a 9-axis IMU. Bioimpedance and IMU signals were sampled at 250 Hz and 100 Hz, respectively, synchronized, and resampled to 50 Hz. Signal preprocessing included median, Savitzky-Golay, and Butterworth filtering to reduce noise and motion artefacts. The processed signals were segmented into overlapping windows of 5 s (250 samples) and 2.56 s (128 samples) with 50% overlap, from which feature matrices were extracted. To support subject-independent evaluation, the dataset was partitioned on a participant basis into 70% training subjects and 30% testing subjects. The dataset includes raw sensor recordings, segmented sensor windows, extracted feature matrices, activity labels, subject identifiers, and supporting documentation. By combining multi-channel bioimpedance measurements with dual-location IMU sensing, this dataset provides a unique resource for investigating multimodal wearable sensing approaches for robust human activity classification research.