ABSTRACT
We present a method for evaluating COVID-19 contamination risk based on social distancing between individuals and face mask usage. Our method employs images captured by surveillance cameras as input to a system that computes a health risk indicator in real time. This system can handle real-world situations, performing detections in large public spaces, such as squares and streets, as well as other potentially crowded areas like restaurants and shopping centers. Our system uses the number of people with and without masks and their proximity to evaluate the risk of COVID-19 contamination. We employed deep neural networks to detect people with and without masks, and we used computer vision to measure the distance between them. Both cases presented challenges, including distinguishing face masks at wildly different distances and positions concerning the camera, occlusions, shape variance, etc. We have built and made public a face mask detection dataset (44,402 faces) with images that include these challenging scenarios and used them to train our deep neural networks. Our best deep neural network architecture achieved 91.41% precision, 82.88% accuracy, and 89.88% recall on face mask detection. © 2022 IEEE.