Thermal-Only Facial Sensing of Stress and Boredom Toward Occupant-Centric Control

Occupant stress and boredom significantly impact comfort, productivity, and safety in indoor environments, underscoring the need for scalable, non-intrusive affect sensing. This study presents a thermal-only facial-sensing framework for differentiating between stress and boredom using infrared thermography, without relying on contact-based sensors or privacy-sensitive RGB imaging. Facial thermal video data were collected from 11 participants using a within-subjects protocol comprising baseline, stress-inducing, boredom-inducing, and recovery conditions. Region-of-interest (ROI)-based facial thermal signals were baseline-normalised to reduce inter-individual variability, and both static distributional features and temporal dynamic features were extracted. Feature redundancy was mitigated using Pearson correlation and Variance Inflation Factor (VIF) analyses, resulting in a compact, physiologically relevant feature set. Multiple machine learning classifiers were evaluated, including Random Forest, XGBoost, LightGBM, SVC, kNN, Logistic Regression, and MLP. Random Forest achieved the highest performance under within-dataset evaluation, with an accuracy of 0.82 and balanced F1-scores exceeding 0.82 using combined static-temporal features. Region-wise importance analysis identified the periorbital, forehead, and temporal regions as dominant contributors. While these findings demonstrate proof-of-concept feasibility under controlled indoor conditions, further validation across larger and environmentally diverse cohorts is required before large-scale deployment in adaptive building systems.