Non-contact Heart Rate Estimation Method Based On Face Perception

The monitoring of heart rate has important indications for the warning and prevention of human cardiovascular diseases.As a non-inductive,non-invasive,convenient and efficient new detection technology,non-contact heart rate detection has broad application prospects.The development of deep learning technology has made great progress in non-contact heart rate detection technology,but the current method ignores the impact of color space and pulse wave label noise on the model;under complex lighting conditions and head movement conditions,the existing heart rate estimation method appears to be inadequate.In order to address the aforementioned issues,this paper conducted extensive research on the non-contact heart rate estimation method based on face perception.The main research results are as follows:（1）This paper analyzes the characteristics of each color space,and study the influence of spatio-temporal representations in different color spaces on heart rate estimation.Experimental analysis proves that the YUV color space has a more significant heart rate characterization,which is conducive to the network to obtain more accurate heart rate signals.Based on this,aiming at the problem of pulse wave label noise distortion,a loss function based on signal quality analysis is proposed to constrain the network model,so that it can adaptively adjust the loss weight and reduce the interference of noisy labels.The experimental results on the VIPL-HR dataset show that the constraint of the adaptive loss function helps to improve the accuracy of the model.（2）This paper proposes a heart rate estimation method based on multi-level spatial atten-tion upsampling module（Hierarchical Attentive Upsampling Module,HAUM）.This method extracts significant heart rate features through hierarchical spatial attention upsampling,thereby constructing implicit heart rate spatio-temporal information and generating high-resolution spatio-temporal feature maps.Then uses the Res Net-18 to perform feature extraction and heart rate regression.The model uses the L₁loss function to constrain.Experimental results on VIPL-HR and MAHNOB-HCI datasets prove that our method is better than the current main-stream heart rate estimation method.（3）This paper proposes a self-attention signal codec heart rate estimation method（Self-Attention Signal Codec Network,SASC-Net）.Aiming at the problem of feature space selection of heart rate characterization information,this method introduces a self-attention mechanism,characterizes the characteristics of different spaces and channels according to their correla-tion,and enhances the ability of network analysis.This method constrains the pulse wave and heart rate at the same time through the joint loss function,and then guides the neural network to learn more discriminative heart rate characteristics.Experimental results on the VIPL-HR dataset show that the average absolute error of this method reaches 4.47 bpm,which is signifi-cantly higher than the current mainstream heart rate estimation method.At the same time,the experimental results on other datasets also prove the effectiveness of the method.