Non-Contact Measurement of Temperature Using RGB Facial Images

With the arrival of COVID-19, non-contact temper-ature measurement methods quickly gained popularity and have since been used ubiquitously, from malls to government buildings. But research claims that the existing non-contact techniques are less accurate than the contact methods. To obviate this issue, this study proposes a novel algorithm for estimating body temperature without skin contact, using a smartphone camera that captures RGB facial videos. Involving 83 volun-teers, temperature measurements were taken from various body locations using infrared, digital, and analog thermometers. A 30-second facial video was recorded for each participant. Machine learning models were trained to estimate core body temperature from these images. The algorithm is based on a combination of image preprocessing techniques and hyperspectral bands extraction that quantify skin reddening with increasing temperature. The performance of different machine learning algorithms was discussed. Performance metrics revealed that the Extra Trees Regression model achieved the highest accuracy and the lowest Root Mean Square Error. Support Vector Machine required the least amount of computation time. Also, K-Nearest Neighbors algorithm has a good mix of accuracy, Root Mean Square Error, and computation time. Clinical validation against thermal camera measurements confirmed the algorithm's reliability.