The Research And Implementation Of Non-Contact Heart Rate Measurement Algorithm Based On Face Video

As an important vital sign information,heart rate has great relevance for a subject’s affective and physical state.The traditional heart rate measurement method is mainly contact measurement,which brings many inconveniences to patients no matter psychologically or physiologically.Recently,remote heart rate measurement through the camera has raised a lot of attention.However,remote measurement of heart rate can be very challenging due to the noisy variations such as ambient light changes and subjects’movements.This thesis focuses on video-based non-contact heart rate measurement method under complex environmental conditions.The main work is as follows:1)As for the problem of not being accurate enough for the region of interest(ROI)positioning when acquiring pulse signal from facial videos,an improved facial landmarks detection method is proposed,which cascades the SSD face detector with facial landmarks detector to improve the accuracy of landmarks detection.ROI is located based on detected landmarks,and a reliable ROI region is selected to reduce the noise in the pulse signal.2)Four signal preprocessing methods were studied.The effectiveness of each method was verified through experiments,the chrominance-based method and second-order central difference method were combined to improve the expression ability of pulse signals,thereby improving accuracy.3)Aiming at the poor anti-noise ability in traditional heart rate extraction method,this thesis proposed a deep-learning based heart rate extraction algorithm.Convolutional neural network(CNN)is designed as the feature extractor of heart rate information.The time-frequency representation(TFR)which combined pulse signal’s periodic characteristics and spectral features was used as input,and multi-task learning method was used for training.The experimental results show that the heart rate measurement method based on deep learning has improved both accuracy and robustness compared to the method without deep learning.4)The heart rate extraction algorithm of this paper is applied to face anti-spoofing filed.The heart rate feature extracted from the heart rate extraction model is proposed as a marker for distinguishing between living and non-living subjects.Experiments show that this method can effectively distinguish between living and non-living,and has a clear semantic definition.