Non-Invasive Thermal Comfort Recognition by Monitoring Human Thermal Adaptive Actions and Facial Temperatures

Precisely grasping human thermal comfort indoors is crucial for controlling HVAC systems and enhancing building energy efficiency. This paper proposes a non-invasive approach for thermal comfort prediction by monitoring human thermal adaptive actions and facial temperatures simultaneously. The Mediapipe tool is applied to detect thermal adaptive actions of seated human body with detection accuracy of 99%. The YOLOv5 (You Only Look Once) algorithm is used to identify key human facial regions with average precision exceeding 96%. In order to predict thermal comfort more comprehensively and accurately, a fusion model is proposed by combining two non-invasive approaches together using the Random Forest algorithm. Field experiment results show that the accuracy of fusion model for thermal comfort prediction reaches 85.71%. For the data containing thermal adaptive actions detected, the thermal comfort prediction accuracy could reach 97.37%.