Multimodel Lightweight Transformer Framework for Human Activity Recognition

Human Activity Recognition (HAR) finds extensive application across diverse domains. Yet, its integration into healthcare remains challenging due to disparities between prevailing HAR systems optimized for rudimentary actions in controlled settings and the nuanced behaviors and dynamic conditions pertinent to medical diagnostics. Furthermore, prevailing sensor technologies and deployment scenarios present formidable hurdles regarding wearability and adaptability to heterogeneous environments. While navigating these constraints, this investigation evaluates the requisite monitoring simplicity and system adaptability crucial for medical contexts. A HAR framework is proposed, leveraging a Lightweight Transformer architecture with a multi-sensor fusion strategy employing five Inertial Measurement Units (IMUs) as sensors. A Real-world HAR dataset is assembled to authenticate the system's suitability, and a comprehensive array of experiments is conducted to showcase its potential utility.