Vision-based personal thermal comfort modeling under facial occlusion scenarios

Personal thermal comfort modeling can accurately identify the transient thermal comfort states of individuals, facilitating occupant-centric indoor thermal comfort regulation. Facial temperature is the important data source for developing personal thermal comfort model. However, facial occlusion often occurs in daily life, such as wearing eyeglasses or masks, would hinder the acquisition of facial temperature. Previous studies have ignored the facial occlusion scenarios, which narrowed the application scope of the model. This study proposed a method fusing visible and infrared images to fill this knowledge gap. Firstly, the facial occlusion scenarios and corresponding Regions of Interest (ROIs) were recognized from the visible images based on YOLOv8 and FaceMesh. Secondly, the coordinates of ROIs were mapped from visible images onto the infrared images, and the temperature features of each ROI were calculated. Finally, Random Forest (RF) algorithm-based models were developed to predict the subjective thermal comfort indices. 3029 sets of data were collected in the experiment to verify the prediction models under four facial occlusion scenarios (i.e., without occlusion, wearing eyeglasses, wearing mask, wearing both). The results showed that: (1) the accuracy of the proposed prediction models was improved by 3.30% to 14.17% compared with the baseline model based on environmental parameters, (2) temperature features of hand and median temperature feature type were important for personal thermal comfort modeling, and (3) the addition of air temperature and subjects’ Body Mass Index (BMI) could significantly improve the model performance by 6.34% and 5.39%.