Hybrid Learning Models for IMU-Based HAR with Feature Analysis and Data Correction.

Area of Science:

• Biomedical Engineering
• Wearable Technology
• Machine Learning for Health

Background:

• Human Activity Recognition (HAR) is crucial for healthcare and monitoring.
• Existing vision-based HAR systems have limitations in privacy and environmental dependence.
• Wearable Inertial Measurement Unit (IMU) sensors offer a promising alternative for unobtrusive HAR.

Purpose of the Study:

• To develop and evaluate a novel real-time HAR system using wearable IMU sensors.
• To implement a hybrid learning model for accurate human activity classification.
• To demonstrate the system's effectiveness for common daily activities.

Main Methods:

• A real-time HAR system architecture was designed utilizing wearable IMU sensors.
• Data involved four body movement classes: stand-up, sit-down, run, and walk.
• A hybrid learning approach combined an eXtreme Gradient Boosting (XGBoost) classifier and a Convolutional Variational Autoencoder (CVAE) generator.

Main Results:

• The proposed system achieved a high classification accuracy of 96.03%.
• The hybrid model effectively processed sensor data for activity recognition.
• The system demonstrated robust performance across different human activities.

Conclusions:

• The wearable IMU-based HAR system offers a viable and accurate alternative to vision-based methods.
• The hybrid XGBoost and CVAE model is effective for real-time human activity classification.
• This approach has significant potential for applications in health monitoring and assistive technologies.