Respiration Monitoring Through Depth Camera in Moving Scenarios

Motion-robust, continuous, and contactless respiration monitoring has the potential to enhance clinical assessments by enabling more frequent inclusion of respiratory rate in cardiopulmonary risk evaluation. Depth sensing can be used to extract the torso surface area trend over time, which is the result of the breathing process. Movement artifacts still represent a major problem. In this paper, four subjects are asked to walk and breathe freely in a volume measurement. A commercial wearable chest strap is employed as a ground truth. Each subject pose is tracked by deep learning. Four regions of interest are evaluated for respiratory extraction: the entire torso, chest, abdomen, and chest centroid. To compensate for the effect of motion, the nose, the middle of the shoulder, and the abdomen are considered as possible movement reference points and subtracted from the signal extracted from the previously mentioned region of interest. The best results given by the chest area correspond to a Mean Absolute Error (MAE) of 0.958 bpm, a standard deviation (STD) of 2.132 bpm, and a Root Mean Square Value (RMSE) of 2.338 bpm.