Automated Thermal Screening for COVID-19 using Machine Learning
This addresses the problem of accurate and efficient COVID-19 screening in public places, offering an incremental improvement by leveraging thermal imaging to overcome lighting limitations.
The paper tackles automated COVID-19 screening by using machine learning on thermal video streams for face and mask detection and temperature measurement, showing that thermal imaging performs as effectively as visual imaging in high illumination and maintains performance under low-light conditions where visual methods degrade by over 50%.
In the last two years, millions of lives have been lost due to COVID-19. Despite the vaccination programmes for a year, hospitalization rates and deaths are still high due to the new variants of COVID-19. Stringent guidelines and COVID-19 screening measures such as temperature check and mask check at all public places are helping reduce the spread of COVID-19. Visual inspections to ensure these screening measures can be taxing and erroneous. Automated inspection ensures an effective and accurate screening. Traditional approaches involve identification of faces and masks from visual camera images followed by extraction of temperature values from thermal imaging cameras. Use of visual imaging as a primary modality limits these applications only for good-lighting conditions. The use of thermal imaging alone for these screening measures makes the system invariant to illumination. However, lack of open source datasets is an issue to develop such systems. In this paper, we discuss our work on using machine learning over thermal video streams for face and mask detection and subsequent temperature screening in a passive non-invasive way that enables an effective automated COVID-19 screening method in public places. We open source our NTIC dataset that was used for training our models and was collected at 8 different locations. Our results show that the use of thermal imaging is as effective as visual imaging in the presence of high illumination. This performance stays the same for thermal images even under low-lighting conditions, whereas the performance with visual trained classifiers show more than 50% degradation.