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

Human Activity Recognition (HAR) has attracted significant research attention, leading to impressive recognition accuracy through advanced algorithms and thorough data validation. Despite the advantages of multisensor systems in capturing detailed behavioral data, significant challenges persist in applying these advancements to real-world scenarios, including increased data processing demands, user behavior variability, and operational issues like sensor dropout or positional changes. This study proposes a Generative Hierarchical Light Transformer (GAN-HLT) framework tailored explicitly for multi-IMU sensing networks to tackle these issues. The framework utilizes generative models to expand the quantity and types of collected data, addressing the issue of performance differences among different users in multisensor networks and the impact of unknown sensor changes. In addition, it also integrates a transformer-based hierarchical classifier, which improves the accuracy of behavior recognition while being lightweight, and ensures scalability in various dynamic environments. Extensive experimental evaluations support the effectiveness of the system, demonstrating its ability to overcome the limitations inherent in multisensor HAR systems. The findings highlight the promising potential of the GAN-HLT framework for real-world applications, offering a significant step forward to improve the practical deployment of HAR technologies.Clinical relevance - The GAN-HLT framework is clinically relevant, enabling accurate, scalable activity recognition, continuous monitoring of patients with mobility impairments, and actionable insights for personalized care beyond clinical settings.