Research On Non-contact Heart Rate Detection System Based On Deep Learning

Traditional heart rate detection methods generally require close contact between the human body and the detection equipment,which can easily cause health problems,and are not suitable for long-term measurement.Therefore,some scholars have begun to pay attention to non-contact heart rate detection methods,among which the face video heart rate detection method based on i PPG has been favored by researchers.This method only requires a camera to measure the heart rate,which is simple and convenient.However,the current method of heart rate detection in face video based on i PPG still has problems such as high noise interference and low accuracy.The main reason is that factors such as lighting changes,facial expressions,and head movements during face video shooting can cause a large amount of noise in the extracted i PPG signal,which leads to low accuracy of heart rate detection results in face video.In response to the above issues,this article focuses on the following research work based on deep learning technology:1.Aiming at the problem of abnormal data and insufficient data volume in the data set,this article mainly deals with it from two aspects.One is to use statistical methods to analyze the video data in the data set and remove abnormal video data from the data set that exceeds the normal heart rate range;The second is to expand the dataset using horizontal flipping and overlapping sampling methods to obtain more training data samples,making the generalization ability of the network model obtained from the training set stronger.2.Research on facial ROI selection and feature map construction was conducted.Through research and analysis of relevant literature,this article uses the SSD based face detector in the Open CV library for face detection,uses the 68 point labeling method in the Dlib library to locate key points in the face detection results,and selects all face regions except the forehead as ROI to fully utilize the heart rate information of each region of the face;In addition,the method of characterizing the i PPG signal as a spatiotemporal map is used to construct a feature map,which increases the network input characteristics of the sample data and preserves the specific information of the interest rate characteristics to the greatest extent.3.Aiming at the problem that i PPG signals in face video heart rate detection are susceptible to noise interference,an improved DRSN-CW network detection method is proposed.Based on DRSN-CW,this paper introduces a spatial soft threshold attention mechanism to set more reasonable thresholds for different regions in the feature map to better remove the interference of unrelated noise and improve the anti-interference ability of the network.It also replaces the Re Lu activation function in the network with Leaky Re Lu to solve the problem that neurons no longer continue to learn after the Re Lu activation function enters a negative range.Based on this,network performance evaluation experiments and ablation experiments are designed.The experimental results show that the improved DRSN-CW network performance has been greatly improved,and the detection results have higher accuracy and robustness.4.Establish a non-contact heart rate detection system based on face video.Establish a face video heart rate detection system based on improved DRSN-CW network as the core technology.The system includes four functions: user registration and login,heart rate detection,user information management,and detection result management.It allows users to independently complete non-contact heart rate detection,ensuring their heart rate health.