Thermal sensation inference in an integrated RGB-EVM framework for occupant-centric HVAC control: deep learning model benchmarking and comparison with thermal imaging

Thermal comfort monitoring is essential for creating energy-efficient, healthy, and comfortable indoor environments. Accurate monitoring reduces energy waste by aligning HVAC operation with occupants’ actual needs, preventing over- or under-conditioning. Our previous work introduced a framework combining Eulerian Video Magnification (EVM) with a YOLOv8-based detector to infer thermal sensation from RGB video, offering a low-cost and scalable solution. A key next step is evidence-based model selection, as detection architectures influence robustness to motion and lighting variation, generalisation to unseen occupants, and computational efficiency. Recent advances in both lightweight single-stage and accurate two-stage detectors motivate a systematic benchmark within the RGB-EVM pipeline. Additionally, the approach has not been rigorously compared with thermal imaging methods, which remain a common reference for non-contact assessment. To address these gaps, this study evaluates representative single-stage and two-stage detectors within the RGB-EVM framework under consistent training and resource conditions, and compares the best model with a thermal imaging baseline. Models from YOLOv8n to YOLOv12n, along with Faster R-CNN, were assessed. YOLOv12n achieved the best balance of accuracy, efficiency, and transferability. The RGB-based model reached an mAP50 of 78.9% and mAP50-95 of 73.7%, compared with 81.8% and 63.3% for thermal imaging. While overall accuracy remained comparable, the RGB-based approach showed more consistent classification behaviour and greater transferability. These results suggest that low-cost RGB cameras, enhanced by EVM and appropriate deep learning architectures, provide a competitive and scalable alternative to thermal imaging for thermal sensation inference in smart building and HVAC applications.