Improving Accuracy of Respiratory Rate Estimation by Restoring High Resolution Features With Transformers and Recursive Convolutional Models

Alicja Kwasniewska, Maciej Szankin, Jacek Ruminski, Anthony Sarah, David Gamba

Abstract: Non-contact evaluation of vital signs has been becoming increasingly important, especially in light of the COVID-19 pandemic, which is causing the whole world to examine people's interactions in public places at a scale never seen before. However, evaluating one's vital signs can be a relatively complex procedure, which requires both time and physical contact between examiner and examinee. These requirements limit the number of people who can be efficiently checked, either due to the medical station throughput, patients' remote locations or the need for social distancing. This study is a first step to increasing the accuracy of computer vision-based respiratory rate estimation by transferring texture information from images acquired in different domains. Experiments conducted with two deep neural network topologies, a recursive convolutional model and transformers, proved their robustness in the analyzed scenario by reducing estimation error by 50% compared to low resolution sequences. All resources used in this research, including links to the dataset and code, have been made publicly available.