A low-power HAR method for fall and high-intensity ADLs identification using wrist-worn accelerometer devices.

The article focuses on a low-power classification algorithm named EKMeans, which combines K-Means and K-Nearest Neighbors (KNN) to identify falls and high-intensity Activities of Daily Living (ADLs) using triaxial accelerometer data from wrist-worn devices. Validated on three public Fall&ADL datasets (UMAFall, UCIFall, and FallAllD), EKMeans demonstrated superior specificity in classifying ADLs compared to baseline KNN and feed-forward Neural Network (NN) models, particularly when datasets were balanced using SMOTE oversampling. Battery consumption tests on a WearOS smartwatch (TICWATCH® E2) showed EKMeans extended battery life to 27.45 hours, outperforming KNN and NN by 5% and 21%, respectively. The study highlights EKMeans as a computationally efficient and battery-friendly solution for wearable human activity recognition, while suggesting further research into alternative low-power classifiers and device platforms for improved fall detection.