The use of deep learning for smartphone-based human activity recognition

• DOI: 10.3389/fpubh.2023.1086671
• Publication Date: 2023-03-06T16:59:52Z
• Authors: Tristan Stampfler; Mohamed Elgendi; Richard Ribon Fle...; Carlo Menon
• Keywords: Overfitting; Residual neural network; Dropout (neural networks); Smoothing; Activity Recognition; F1 score; Ranging

The emerging field of digital phenotyping leverages the numerous sensors embedded in a smartphone to better understand its user's current psychological state and behavior, enabling improved health support systems for patients. As part this work, common task is use accelerometer automatically recognize or classify behavior user, known as human activity recognition (HAR). In article, we present deep learning method using Resnet architecture implement HAR popular UniMiB-SHAR public dataset, containing 11,771 measurement segments from 30 users ranging age between 18 60 years. We unified approach based on that consistently exceeds state-of-the-art accuracy F1-score across all classification tasks evaluation methods mentioned literature. most notable increase disclose regards leave-one-subject-out evaluation, rigorous method, where push 78.24 80.09% 78.40 79.36%. For such results, resorted techniques, hyper-parameter tuning, label smoothing, dropout, which helped regularize training reduced overfitting. discuss how our could easily be adapted perform real-time future research directions.