GAN-HLT: Generative Hierarchical Light-Transformer for Extendable Human Activity Recognition.

Area of Science:

• Human Activity Recognition
• Multisensor Networks
• Wearable Technology

Background:

• Multisensor systems offer rich behavioral data but face real-world deployment challenges.
• Existing Human Activity Recognition (HAR) methods struggle with data processing, user variability, and sensor inconsistencies.
• Advancements in HAR algorithms show promise but require adaptation for practical, dynamic environments.

Purpose of the Study:

• To propose a novel Generative Hierarchical Light Transformer (GAN-HLT) framework for multi-Inertial Measurement Unit (IMU) sensing networks.
• To address performance variations across users and adapt to sensor changes in multisensor HAR systems.
• To enhance the accuracy, scalability, and practical applicability of HAR in real-world settings.

Main Methods:

• Utilized generative models to augment data quantity and diversity, mitigating user performance differences and sensor variability.
• Integrated a lightweight, transformer-based hierarchical classifier for improved behavior recognition accuracy.
• Developed a framework specifically for multi-IMU sensing networks to handle dynamic environmental conditions.

Main Results:

• The GAN-HLT framework demonstrated effectiveness in overcoming limitations of traditional multisensor HAR systems.
• Experimental evaluations confirmed the system's ability to improve recognition accuracy and scalability.
• The proposed framework shows significant potential for real-world HAR deployment and continuous patient monitoring.

Conclusions:

• The GAN-HLT framework offers a robust solution for multisensor HAR, enhancing accuracy and scalability.
• This approach facilitates practical deployment of HAR technologies, particularly for continuous patient monitoring.
• The framework provides actionable insights for personalized care, extending beyond clinical settings.