Intelligent non-invasive vital signs estimation from image analysis

This study aims to build a fully intelligent noninvasive vital-sign signal detection from image analysis in terms of clinical scenarios to extract breathing rate, heart rate, and blood pressure values. The state-of-the-art object detection Yolov3 is used to localize the interesting bounding boxes (chest, face, palm), these ROIs are then tracked by the Mosse algorithm to boost the processing performance. Next, the Pyramidal Lucas-Kanade and remote photoplethysmography techniques are used for extracting the motion signals (breath, pulse) and subtle color change induced by pulse, respectively. Besides, digital signal processing is applied to remove undesired noises for obtaining a clean bio-signal. From experiments conducted, our system can detect breathing rate, heart rate in real-time at a long distance in terms of motion scenarios. Similar to the noninvasive blood pressure estimation system, the proposed deep learning model overcomes the dependence of the high-speed camera in previous works. It satisfies two medical standards (British Hypertension Society and Association for the Advancement of Medical Instrumentations) in estimating Systolic Blood Pressure (SBP) and Diastolic Blood Pressure (DBP) with the root mean squared error and mean absolute error for SBP/DBP are 7.942/7.912 mmHg and 6.556/6.372, respectively. The proposed approach estimates blood pressure reliably by only an ordinary webcam with 30 fps in a non-contact continuous manner. Thus, it can be concluded that our system can be applied to healthcare applications.