Remote Optical Estimation of Respiratory Rate Based on a Deep Learning Human Pose Detector

Abstract

Respiratory rate (RR) is a handy parameter in the clinical field since it allows the timely detection of diverse pathologies. However, RR is currently acquired using expensive devices which are attached to the patients and therefore may be uncomfortable to use. In this paper, we present a method for the estimation of respiratory rate through a non-contact optical method based on a deep learning human pose detector. The proposed method is tested using a database of videos of subjects performing different respiratory maneuvers to obtain the respiratory signal, and the instantaneous respiratory rate automatically. The proposed method obtained a correlation of 0.8 on static breathing maneuvers with respect to the ground truth signal. For the instantaneous respiratory rate, it was observed through a time-frequency analysis, that the obtained signal shares the same frequency bandwidth as that contained in the ground truth, indicating that the information of both signals is on the same frequency band. Our results indicate the feasibility of employing the proposed method for estimation respiratory rate frequency.