Fully automated real-time approach for human temperature prediction and COVID-19 detection-based thermal skin face extraction using deep semantic segmentation

The COVID-19 pandemic caused by the coronavirus has affected every country. The World Health Organization has declared it a global health emergency by 2020. Early intervention provides a more effective cure and relieves patients of unnecessary intensive care from hospitals and other healthcare facilities. However, that requires an accurate method for early disease detection. COVID-19 illness may eventually infect people, making it a severe condition. Research shows that COVID-19 pneumonia shares many clinical features with other forms of pneumonia, suggesting that COVID-19 can be a serious condition for those affected. The main symptom of the disease is high body temperature. Noncontact thermal imaging is one of the most important methods for measuring body temperature, although it has some performance limitations. Many of these thermal cameras are privately owned, not widely available, and are often owned by for-profit organizations. In this paper, we propose a real-time method for human temperature prediction and tracking based on facial skin temperature extraction for accurate COVID-19 case detection. The main idea of this paper is to use an automated deep-learning model for thermal skin temperature extraction. The optimization is based on using the deep semantic segmentation approach for fully automated thermal skin binary mask prediction using the thermal video only. The predicted mask is projected onto the original thermal video to extract the thermal skin mask, which is used to calculate the average thermal skin face temperature. The proposed deep semantic segmentation model is trained on the thermal skin binary mask dataset that is generated by using our first model a fully automated unsupervised learning approach for camera calibration and thermal skin binary mask extraction. The experimental results show that the proposed approach can automatically predict better binary skin-mak than our first model which leads to a very high-efficiency human temperature tracking using thermal videos only compared with the Ground Truth using the Speaking Faces thermal datasets.