Personalized Human Activity Recognition using Wearables: A Manifold Learning-based Knowledge Transfer.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Human Activity Recognition (HAR) is crucial for healthcare applications like elderly care and patient monitoring.
• Current HAR systems heavily rely on supervised machine learning, with accuracy dependent on training and operating context similarity.
• A need exists for HAR algorithms that can readily adapt to diverse operating contexts.

Purpose of the Study:

• To propose a cross-subject transfer learning algorithm for adaptable human activity recognition.
• To develop a method for linking source and target subjects to personalize HAR models.
• To improve the accuracy of HAR models in new or unseen contexts.

Main Methods:

• Developed a cross-subject transfer learning algorithm utilizing feature-level representations.
• Constructed manifolds from source subject data to link to target subjects.
• Assigned labels to unlabeled target data using the learned manifold for personalized model development.

Main Results:

• The proposed algorithm effectively links source and target subjects for personalized HAR.
• Demonstrated the algorithm's efficacy using a publicly available HAR dataset.
• Achieved significant improvements in activity recognition accuracy, up to 24%.

Conclusions:

• The cross-subject transfer learning framework enhances the adaptability and accuracy of HAR systems.
• This approach offers a viable solution for developing personalized HAR models in healthcare settings.
• The method shows promise for improving context-aware applications in health-care.