A benchmark for domain adaptation and generalization in smartphone-based human activity recognition.

Area of Science:

• Computer Science
• Machine Learning
• Human-Computer Interaction

Background:

• Human activity recognition (HAR) relies on smartphone inertial sensors, but data variability (sensor, user, device placement) hinders model generalization.
• Existing HAR datasets lack standardization, leading to poor cross-dataset performance due to incompatible data formats (units, sampling rates, labels).

Purpose of the Study:

• To introduce the DAGHAR benchmark, a standardized collection of datasets for evaluating domain adaptation and generalization in smartphone-based HAR.
• To enable controlled cross-dataset generalization studies by removing trivial biases while preserving intrinsic data differences.

Main Methods:

• Standardized six diverse HAR datasets by harmonizing accelerometer units, sampling rates, gravity component, activity labels, user partitioning, and time window size.
• Developed the DAGHAR benchmark to facilitate reproducible domain adaptation and generalization research.
• Provided baseline performance metrics using state-of-the-art machine learning models.

Main Results:

• The DAGHAR benchmark enables controlled evaluation of HAR model generalization capabilities across different datasets.
• Standardization successfully removed trivial biases, allowing for a clearer assessment of intrinsic dataset differences.
• Baseline metrics offer a crucial reference point for future HAR generalization studies.

Conclusions:

• The DAGHAR benchmark is essential for advancing domain adaptation and generalization research in smartphone-based HAR.
• Standardized datasets are critical for reliable evaluation and development of robust HAR models.
• Future research can leverage DAGHAR to develop more adaptable and generalizable HAR systems.