Thermal memory of chair: Physics-guided generalization for sitting posture inference using thermal imaging

Thermal imaging offers a viable approach for contactless posture monitoring due to its privacy-preserving nature and ability to capture residual thermal patterns. Existing methods exhibit limited generalization capabilities across different materials and thermal decay stages, coupled with a lack of reliable physical interpretability. To address these challenges, this study proposes an integrated paradigm combining generative data augmentation, visual transformer classification, and finite element (FE) simulation. The proposed pipeline first enhances data diversity through a generative model, then employs a Transformer-based classifier to achieve accurate recognition of 9 sitting postures. Finally, a heat conduction model is constructed to simulate the real thermal decay temperature field, decoding the influence of material and time on buttock thermal patterns. Through this paradigm, we identify a critical temperature difference threshold of 2.6 ± 0.06 K, beyond which model performance significantly degrades. Systematic analysis demonstrates that maintaining surface temperatures above this threshold during the initial 30 s enables the model to sustain accuracy above 85%. Furthermore, we quantified the direct impact of material thermophysical parameters on the effective detection window, revealing that materials with lower thermal conductivity (e.g., plastics) extend reliable identification duration. Validation on an independent test set featuring two materials and varying decay durations demonstrated a classification accuracy of 0.9162. This study establishes a thermal imaging-based posture analysis paradigm, providing a theoretical foundation and practical solutions for real-world applications in privacy-sensitive scenarios by decoding buttock thermal patterns. The dataset and code supporting this study are publicly available at: https://github.com/AJ-1995/Thermal-Memory-of-Chair.