Matrix and Tensor Completion on a Human Activity Recognition Framework.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Pervasive Healthcare

Background:

• Sensor-based activity recognition is vital for pervasive healthcare and biomedical research.
• Unobserved measurements hinder machine learning algorithms in extracting context from data streams.

Purpose of the Study:

• To accurately estimate missing multimodal inertial data.
• To propose a classification framework incorporating data reconstruction during the test phase.

Main Methods:

• Forming available data streams into low-rank 2-D and 3-D Hankel structures.
• Utilizing matrix and tensor completion for data imputation.
• Evaluating reconstruction impact on classification performance with state-of-the-art classifiers.

Main Results:

• Robust classification accuracy achieved through data recovery, even with extremely subsampled data.
• Analysis of data structuring, reconstruction volume, and missing data levels on performance.
• Examination of the trade-off between subsampling accuracy and energy conservation in wearable platforms.

Conclusions:

• Reconstruction techniques significantly improve activity recognition from incomplete sensor data.
• The proposed framework effectively handles missing inertial data in pervasive healthcare applications.
• Matrix and tensor completion are viable methods for enhancing sensor-based activity recognition systems.