Robust Human Activity Recognition Using a Transformer-Based Model for Aging Society
Qixin Liang, Anuchit Jitpattanakul, Sakorn Mekruksavanich · 2025
Recognizing human activities is vital for enhancing healthcare and assisted living solutions for older people, particularly in applications such as continuous health monitoring, fall prevention, and smart home support. Despite advancements, current human activity recognition (HAR) systems face significant challenges when applied to older adults due to variations in gait, reduced mobility, and inconsistent activity patterns. This study introduces Transformer-Net, a robust deep learning model tailored for activity recognition in older adults. The model harnesses the self-attention mechanism of transformer architectures to effectively capture long-term temporal relationships and subtle motion patterns commonly observed in older individuals. Unlike traditional sequence models, the transformer framework excels in managing irregular activity sequences and sensor noise, which is prevalent in real-world monitoring scenarios involving older adults. We evaluate the performance of Transformer-Net using the HAR70+ dataset, which comprises movement data from 18 elderly participants aged 70-95, engaging in everyday activities such as walking, sitting, standing, and lying down, with some using walking aids. Experimental results indicate that the proposed model achieves$\text{9 8. 0 2 \%}$accuracy, 98.56% precision, 98.23% recall, and an F 1 -score of 98.39 %, outperforming traditional models such as CNN (96.85%), LSTM (97.17%), BiLSTM ($\text{9 7. 1 9 \%}$), GRU ($\text{9 7. 0 7 \%}$), and BiGRU ($\text{9 7. 3 7 \%}$). The model demonstrates strong robustness to inter-individual variability and maintains high classification performance across diverse activity types and durations. These results suggest that transformer-based architectures are particularly effective for elderly-focused HAR tasks, offering greater dependability for real-world applications in smart home systems and healthcare monitoring platforms. This work advances the development of age-aware assistive technologies and lays the groundwork for future research in personalized HAR for aging societies.