w-HAR: An Activity Recognition Dataset and Framework Using Low-Power Wearable Devices.

Area of Science:

• Biomedical Engineering
• Machine Learning
• Wearable Technology

Background:

• Human Activity Recognition (HAR) is crucial for applications like patient rehabilitation and movement disorder analysis.
• Existing HAR methods often rely on noisy inertial sensor data and lack publicly available datasets, hindering algorithm comparison and development.
• Current approaches face challenges in data segmentation and activity classification due to sensor noise and data accessibility issues.

Purpose of the Study:

• To introduce the wearable Human Activity Recognition (w-HAR) dataset, a novel resource for HAR research.
• To address limitations of existing HAR datasets by integrating data from inertial and wearable stretch sensors.
• To develop and evaluate a HAR framework utilizing the w-HAR dataset for accurate activity classification.

Main Methods:

• Collected labeled data of seven activities from 22 users using both inertial and wearable stretch sensors.
• Developed a HAR framework involving neural network architecture design space exploration for activity classification.
• Implemented two online learning algorithms to adapt the HAR classifier to new users.

Main Results:

• The w-HAR dataset provides dual-modality sensor data, enabling variable-length segment creation and single-activity segmentation.
• The proposed HAR framework achieved 95% classification accuracy on the w-HAR dataset.
• Online learning algorithms significantly improved accuracy, with gains up to 40% for unseen users.

Conclusions:

• The w-HAR dataset offers a valuable, multi-modal resource for advancing HAR research.
• The developed HAR framework demonstrates high accuracy and adaptability, particularly with the integration of online learning.
• This work facilitates more robust and personalized human activity recognition systems for diverse applications.