Large-Scale Pre-Training for Dual-Accelerometer Human Activity Recognition

Abstract

The annotation of physical activity data collected with accelerometers for human activity recognition (HAR) remains challenging despite the growing interest in large public health studies. Existing free-living accelerometer-based datasets are limited, hindering the training of effective deep learning models. To address this limitation, some studies have explored self-supervised learning (SSL), i.e., training models on both labeled and unlabeled data. Here, we extend previous work by evaluating whether large-scale pre-training improves downstream HAR performance. We introduce the SelfPAB method, which includes pre-training a transformer encoder network on increasing amounts of accelerometer data (10-100K hours) using a reconstruction objective to predict missing data segments in the spectrogram representations. Experiments demonstrate improved downstream HAR performance using SelfPAB compared to purely supervised baseline methods on two publicly available datasets (HARTH and HAR70+). Furthermore, an increase in the amount of pre-training data yields higher overall downstream performance. SelfPAB achieves an F1-score of 81.3% (HARTH), and 78.5% (HAR70+) compared to the baselines' F1-scores of 74.2% (HARTH) and 63.7% (HAR70+). Additionally, SelfPAB leads to a performance increase for activities with little training data.