HARTH: A Human Activity Recognition Dataset for Machine Learning.

Area of Science:

• Biomedical Engineering
• Computer Science
• Machine Learning

Background:

• Existing human activity recognition (HAR) datasets lack consistent sensor placement and reliable annotations, hindering real-world application development.
• Free-living conditions present unique challenges for accurate activity monitoring due to variability in sensor positioning and user behavior.

Purpose of the Study:

• Introduce the Human Activity Recognition Trondheim (HARTH) dataset, a novel benchmark for HAR research.
• Provide a high-quality, reliably annotated dataset to facilitate the development of advanced machine learning models for HAR.
• Evaluate the performance of various machine learning models on the HARTH dataset.

Main Methods:

• Collected data from 22 participants over 90-120 minutes during regular working hours.
• Utilized two three-axial accelerometers (thigh, lower back) and a chest-mounted camera for data acquisition.
• Employed expert annotators with high inter-rater agreement (Fleiss' Kappa = 0.96) to label twelve distinct activities based on video recordings.

Main Results:

• Trained and evaluated seven machine learning models, including SVM, KNN, Random Forest, XGBoost, CNN, and BiLSTM.
• The Support Vector Machine (SVM) model achieved the highest performance with an F1-score of 0.81 (±0.18), recall of 0.85 (±0.13), and precision of 0.79 (±0.22).
• Leave-one-subject-out cross-validation demonstrated the generalizability of the models across different participants.

Conclusions:

• The HARTH dataset offers a robust benchmark for advancing HAR research in free-living environments.
• The findings highlight the potential of machine learning approaches for precise human activity recognition.
• This dataset will empower researchers to develop more accurate and reliable HAR systems.