Enhancing human activity recognition with TB-ConvAtt: A multi-dimensional attention framework
Authors: Yang, H., Wang, Y., Hu, R., Geng, Y., Wang, A., Zhou, X., Yu, H., Zhao, Q.
Journal: Biomedical Signal Processing and Control
Publication Date: 01/12/2025
Volume: 110
eISSN: 1746-8108
ISSN: 1746-8094
DOI: 10.1016/j.bspc.2025.108314
Abstract:The growing prevalence of wearable technology in healthcare highlights the essential need for accurate and efficient human activity recognition (HAR) using wearable sensor data. In clinical settings, HAR plays a pivotal role in patient monitoring, rehabilitation, and personalized healthcare management. This study introduces TB-ConvAtt, a lightweight and multi-dimensional framework that integrates Convolutional Neural Networks (CNNs) with specialized attention mechanisms to effectively balance the extraction of independent temporal, spatial, and spatio-temporal features from wearable multi-sensor data. TB-ConvAtt consists of three distinct branches: the Temporal Attention Dimension (TAD), the Spatial Attention Dimension (SAD), and the Spatio-temporal Attention Dimension (STAD). The performance of TB-ConvAtt is thoroughly evaluated on four public HAR datasets (UNIMIB-SHAR, OPPORTUNITY, PAMAP2, MHEALTH). Comparative studies and detailed ablation experiments demonstrate that TB-ConvAtt achieves state-of-the-art performance while maintaining a lightweight design, enabling efficient deployment in resource-constrained environments.
Source: Scopus