Existing accelerometer-based human activity recognition (HAR) benchmark datasets that were recorded during free living suffer from non-fixed sensor placement, the usage of only one sensor, and unreliable annotations. We make two contributions in this work. First, we present the publicly available Human Activity Recognition Trondheim dataset (HARTH). Twenty-two participants were recorded for 90 to 120 min during their regular working hours using two three-axial accelerometers, attached to the thigh and lower back, and a chest-mounted camera. Experts annotated the data independently using the camera’s video signal and achieved high inter-rater agreement (Fleiss’ Kappa $=0.96$). They labeled twelve activities. The second contribution of this paper is the training of seven different baseline machine learning models for HAR on our dataset. We used a support vector machine, k-nearest neighbor, random forest, extreme gradient boost, convolutional neural network, bidirectional long short-term memory, and convolutional neural network with multi-resolution blocks. The support vector machine achieved the best results with an F1-score of $0.81$ (standard deviation: $\pm 0.18$), recall of $0.85\pm 0.13$, and precision of $0.79\pm 0.22$ in a leave-one-subject-out cross-validation. Our highly professional recordings and annotations provide a promising benchmark dataset for researchers to develop innovative machine learning approaches for precise HAR in free living.

Activity monitoring combined with machine learning (ML) methods can contribute to detailed knowledge about daily physical behavior in older adults. The current study (1) evaluated the performance of an existing activity type recognition ML model (HARTH), based on data from healthy young adults, for classifying daily physical behavior in fit-to-frail older adults, (2) compared the performance with a ML model (HAR70+) that included training data from older adults, and (3) evaluated the ML models on older adults with and without walking aids. Eighteen older adults aged 70–95 years who ranged widely in physical function, including usage of walking aids, were equipped with a chest-mounted camera and two accelerometers during a semi-structured free-living protocol. Labeled accelerometer data from video analysis was used as ground truth for the classification of walking, standing, sitting, and lying identified by the ML models. Overall accuracy was high for both the HARTH model (91%) and the HAR70+ model (94%). The performance was lower for those using walking aids in both models, however, the overall accuracy improved from 87% to 93% in the HAR70+ model. The validated HAR70+ model contributes to more accurate classification of daily physical behavior in older adults that is essential for future research.

Self-supervised learning (SSL) has emerged as a promising alternative to purely supervised learning, since it can learn from labeled and unlabeled data using a pre-train-then-fine-tune strategy, achieving state-of-the-art performances across many research areas. The field of accelerometer-based human activity recognition (HAR) can benefit from SSL since unlabeled data can be collected cost-efficiently due to the ubiquitous nature of sensors embedded in smart devices, which is in contrast to labeled data, that require a costly annotation process. Motivated by the success of SSL and the lack of surveys on SSL for HAR, this survey comprehensively examines 52 SSL methods applied to HAR, and categorizes them into four SSL paradigms based on pre-training objectives. We discuss SSL strategies, evaluation protocols, and utilized datasets. We highlight limitations in current methodologies, including little large-scale pre-training, the absence of foundation models, as well as the scarcity of systematic domain shift experiments and domain knowledge utilization. Notably, the diversity in evaluation protocols across papers poses a considerable challenge when comparing methods. Future directions outlined in this survey include the development of an SSL framework for HAR to enable standardized benchmarking and large-scale pre-training, along with integrating domain knowledge to enhance model performance.

Annotating accelerometer-based physical activity data remains a challenging task, limiting the creation of robust supervised machine learning models due to the scarcity of large, labeled, free-living human activity recognition (HAR) datasets. Researchers are exploring self-supervised learning (SSL) as an alternative to relying solely on labeled data approaches. However, there has been limited exploration of the impact of large-scale, unlabeled datasets for SSL pre-training on downstream HAR performance, particularly utilizing more than one accelerometer. To address this gap, a transformer encoder network is pre-trained on various amounts of unlabeled, dual-accelerometer data from the HUNT4 dataset: 10, 100, 1k, 10k, and 100k hours. The objective is to reconstruct masked segments of signal spectrograms. This pre-trained model, termed SelfPAB, serves as a feature extractor for downstream supervised HAR training across five datasets (HARTH, HAR70+, PAMAP2, Opportunity, and RealWorld). SelfPAB outperforms purely supervised baselines and other SSL methods, demonstrating notable enhancements, especially for activities with limited training data. Results show that more pre-training data improves downstream HAR performance, with the 100k-hour model exhibiting the highest performance. It surpasses purely supervised baselines by absolute F1-score improvements of 7.1% (HARTH), 14% (HAR70+), and an average of 11.26% across the PAMAP2, Opportunity, and RealWorld datasets. Compared to related SSL methods, SelfPAB displays absolute F1-score enhancements of 10.4% (HARTH), 18.8% (HAR70+), and 16% (average across PAMAP2, Opportunity, RealWorld).

Purpose: Body-worn accelerometers are commonly used to estimate sleep duration in population-based studies. However, since accelerometry-based sleep/wake-scoring relies on detecting body movements, the prediction of sleep duration remains a challenge. The aim was to develop and evaluate the performance of a machine learning (ML) model to predict accelerometry-based sleep duration and to explore if this prediction can be improved by adding skin temperature data, circadian rhythm based on the estimated midpoint of sleep, and cyclic time features to the model. Patients and Methods: Twenty-nine adults (17 females), mean (SD) age 40.2 (15.0) years (range 17– 70) participated in the study. Overnight polysomnography (PSG) was recorded in a sleep laboratory or at home along with body movement by two accelerometers with an embedded skin temperature sensor (AX3, Axivity, UK) positioned at the low back and thigh. The PSG scoring of sleep/wake was used as ground truth for training the ML model. Results: Based on pure accelerometer data input to the ML model, the specificity and sensitivity for predicting sleep/wake was 0.52 (SD 0.24) and 0.95 (SD 0.03), respectively. Adding skin temperature data and contextual information to the ML model improved the specificity to 0.72 (SD 0.20), while sensitivity remained unchanged at 0.95 (SD 0.05). Correspondingly, sleep overestimation was reduced from 54 min (228 min, limits of agreement range [LoAR]) to 19 min (154 min LoAR). Conclusion: An ML model can predict sleep/wake periods with excellent sensitivity and moderate specificity based on a dual-accelerometer set-up when adding skin temperature data and contextual information to the model.