A low-cost deep learning framework for thermal comfort prediction using Eulerian Video Magnification in smart buildings

Predicting occupants’ thermal comfort is vital for optimising indoor energy systems, enhancing building energy efficiency, and ensuring healthier, more comfortable spaces. Traditional methods often use expensive tools such as wearables or thermal cameras to monitor skin temperature (e.g., face or hands), limiting their practicality for everyday household use and broader smart home integration. This study introduces a low-cost, non-intrusive thermal comfort prediction model that leverages Eulerian Video Magnification and deep learning. Using standard video cameras, the framework amplifies subtle facial colour changes linked to skin temperature, allowing detection of thermal sensation levels. A feasibility test was conducted in four office scenarios, where occupant videos were processed through Eulerian Video Magnification at varying magnification levels. These enhanced videos were then used to train and test the You Only Look Once (YOLO)v8 deep learning model. After comparing training results, classification accuracy, and generalisability, the full-frame models at 15× and 20× magnification performed best, achieving mAP50 scores of 80.6% and 81.7%, respectively. These findings highlight the potential of using everyday cameras for accurate, non-invasive thermal comfort prediction. The research offers a foundation for developing integrated, multi-parameter approaches to support more energy-efficient, intelligent home environments. • Developed low-cost, non-invasive comfort prediction framework using ordinary cameras. • Integrated Eulerian Video Magnification with YOLOv8 for thermal sensation detection. • Feasibility test conducted in café, office, and residential building scenarios. • Achieved mAP50 of 80.6% and 81.7% for original 15x and 20x magnification models. • Correlated predicted thermal comfort with actual indoor temperatures to adjust set points.