Accurate Non-Contact Body Temperature Measurement with Thermal Camera under Varying Environment Conditions

Abstract

Non-contact measurement of body temperature is preferred due not only to the convenience it provides but also to the necessity for preventing medical staffs and patients from infection and safety risk. For non-contact body temperature measurement, thermal cameras have been used to measure the temperature of facial skins. However, the problem is that temperature of facial skins varies according to varying environmental conditions such as outside temperature, subject activities prior to measurement, etc. Efforts to compensate the temperature of facial skin locations that are least affected by environmental conditions have shown only a limited success, leaving further improvement in accuracy as necessary. This paper presents a deep learning approach to body temperature prediction based on thermal camera facial skin images that provides highly accurate body temperature under varying environmental conditions. We achieve high accuracy by measuring temperature distributions of several Region-of-Interests (ROIs) on facial skins and learning the relationship between the ground truth body temperatures and the temperature distributions on ROIs. The results indicate that we can obtain around 0.2°C average error in body temperature estimation despite that subjects are exposed to hot and cold temperature, engaged in different physical activities.