Flexible Sensor Network-Based Digital Human Reconstruction Using Spatiotemporal Attention Mechanism and Physical Constraints

In recent years, significant progress in single-walled carbon nanotube (SWNT)-based flexible strain sensors has facilitated the development of low-cost, flexible, and wearable sensor networks for human motion sensing. However, due to the complex and nonlinear mapping between sensor outputs and human motion, accurately achieving digital human reconstruction from such sensor signals remains a challenging task. This paper proposes a spatiotemporal attention-based and physically constrained algorithm for digital human reconstruction using flexible sensor networks. An improved Transformer architecture is designed to model both temporal dynamics and spatial correlations among multiple sensors through a spatiotemporal attention mechanism. Furthermore, a composite loss function incorporating motion continuity and physical plausibility constraints is introduced to improve the smoothness and realism of predicted motion trajectories. The proposed method achieves real-time, end-to-end prediction of 35 human body joints from raw sensor data. Extensive experiments on a custom dataset demonstrate the effectiveness of the approach in terms of both accuracy and temporal stability. The results demonstrate its potential for real-world deployment in applications such as intelligent sports, rehabilitation, immersive digital human systems, and wearable human–computer interaction platforms.