Automatic Detection Of Heart Rate And Discomfort Based On Facial Video

Traditional contact health monitoring methods have many limitations,such as skin contact will bring discomfort,cable connection will lose the flexibility of detection,interventional detection is easy to cause changes in physiological indicators,and so on.Even for burn patients with fragile skin,premature infants and other special situations,it is not suitable to use contact health detection method.The non-contact health monitoring method based on camera provides a new way of self-care,which not only solves some difficulties of contact detection,but also is easy to set up,convenient and comfortable,and can realize all-weather unmanned monitoring.This kind of noncontact automatic detection of vital signs and physical signs can change passive medical treatment into active examination,change passive treatment into active prevention,and help people get the best health benefits.In this paper,a conventional camera is connected to the back-end intelligent algorithm system,and a series of computer vision related algorithms such as r PPG heart rate detection technology,convolutional neural network and some common signal processing methods are used to process the face video frame to extract useful information,detect the heart rate and discomfort of human body in the daytime and at night,It promotes the development of a new non-contact health monitoring method.Firstly,the algorithm of heart rate detection at night based on infrared camera is studied.For the infrared single channel image and signal with weak pulsation and unable to use the channel combination method,we extract the skin and non skin region pixels to create multiple channels and fuse them,and use the illumination change reflected by the non skin region pixels to eliminate the illumination interference on the skin region pixels,so as to improve the quality of pulsation signal.The channel expansion algorithm proposed in this paper can correct the wrong heart rate value in the original signal,significantly improve the signal-to-noise ratio and reduce the measurement error.On average,the measurement coverage is increased from 50% to83%,and the average signal-to-noise ratio is increased from-8.40 d B to-4.62 d B.This paper provides an effective solution for non-contact measurement of sleep reference heart rate.Secondly,this paper explores a series of low-cost hardware and software settings for daytime heart rate detection.At present,the advanced r PPG technology is mainly based on expensive cameras.In order to promote the large-scale use of r PPG technology,low-cost cameras must be used.In this paper,we choose a network camera and a cheap industrial camera,and design a suitable software solution for them.For the cheaper network camera,the coverage rate of 92% is achieved by using the singular spectrum analysis algorithm in the green channel,which can be applied to family daily health care.For low-cost industrial cameras which are widely used in specific fields of professional scenes,the plane orthogonal skin color algorithm can improve the signalto-noise ratio from-6.70 d B to-2.21 d B,and obtain high-quality heart rate signal,which is convenient to accurately obtain instantaneous heart rate or other heart characteristics.Therefore,it can be used in the use cases which require high system performance,such as hospital clinical heart rate monitoring.We can not only get physiological information such as heart rate from a facial video sequence,but also know the comfort or discomfort of human body.Then,an automatic and continuous video processing system based on 3D convolutional neural network is proposed to detect human discomfort end-to-end.Due to the limitation of data set,this paper only discriminates the comfort status of newborns.In this paper,a threedimensional convolutional neural network based on five channel input of pixel intensity and optical flow is proposed,and the video features of pixel intensity and motion for neonatal discomfort detection and the structure of two-dimensional or threedimensional convolutional neural network are deeply studied.Experiments show that by combining the information from the optical flow channel,the 3D convolutional neural network focuses more on the facial region.In this paper,three-dimensional convolutional neural network is improved to detect discomfort.The AUC value is 0.99 and the accuracy is 0.98,which verifies the robustness of the algorithm.This method can achieve real-time monitoring,does not need the conventional front-end face detection and face tracking steps,and is suitable for dealing with challenging situations such as face occlusion.The system alarm will be triggered by the detected discomfort state to inform the clinical staff to deal with the situation of the newborn timely and properly.Therefore,the system can prevent the occurrence of fatal events and improve the early development of infants.