Related Experiment Video
Updated: Aug 16, 2026
06:36
Three-Dimensional Mapping of the Rotation of Interactive Virtual Objects with Eye-Tracking Data
Published on: October 18, 2024
Siamese masked reconstruction with temporal rotation consistency for inertial-based human activity recognition
Francisco M Calatrava-Nicolás1, Todor Stoyanov2, Oscar Martinez Mozos3,4
1Department of Science and Technology, Örebro University, Örebro, 70182, Sweden. francisco.calatrava-nicolas@oru.se.
Scientific Reports
|August 14, 2026
Summary
This study introduces a self-supervised method for Human Activity Recognition (HAR) using wearable sensors. The approach improves generalization to new users by leveraging signal rotation, reducing the need for extensive labeled data.
Area of Science:
- Computer Science
- Biomedical Engineering
- Signal Processing
Background:
- Human Activity Recognition (HAR) using inertial sensors faces challenges in generalizing to unseen users due to data variability.
- Collecting large labeled datasets for HAR is expensive and time-consuming.
- Existing methods struggle with cross-subject generalization in wearable sensor data.
Purpose of the Study:
- To develop a self-supervised framework for HAR that improves generalization across users without requiring extensive annotated data.
- To leverage the property that signal rotation does not alter the underlying motion content, only its projection onto sensor axes.
- To reduce the gap in HAR performance caused by data heterogeneity.
Main Methods:
- A self-supervised framework using a Siamese masked convolutional autoencoder is proposed.
- The model learns by reconstructing masked inputs and aligning representations of original and rotated signals.
- Temporal consistency loss is applied to enforce agreement between original and rotated signal representations.
Main Results:
- The proposed method shows improved cross-subject generalization in HAR benchmarks, outperforming baselines by +1.4% (linear probe) and +1.2% (MLP probe).
- Achieved competitive performance against fully supervised methods, especially in low-data scenarios.
- Evaluated on diverse HAR datasets including activities of daily living and sports activities.
Conclusions:
- The rotation-based self-supervised approach effectively captures motion content, enhancing HAR model generalization.
- This framework offers a cost-effective solution for HAR by minimizing reliance on labeled data.
- The method demonstrates strong performance and potential for real-world applications of wearable HAR.
