基于人脸视频的逻辑多任务非接触心率估计

Logical multitask contactless heart rate estimation based on face video

为提高远程光体积描记术(remote photoplethysmography,rPPG)在自然场景中心率(heart rate,HR)估计的精度,提出一种基于多头自注意力的多任务心率估计模型。设计一种特征表示方法将视频转换为时空表示,抑制非皮肤区域和背景引入的噪声;使用改进的多头自注意力模型学习长距离帧之间的依赖关系,解决rPPG对光照变化和头部运动鲁棒性不足的问题;将学习到的特征用于估计rPPG信号和HR两个任务,通过rPPG信号的频域约束HR,提高心率估计精度。实验结果表明,提出方法在UBFC-RPPG和PURE两个公开数据集上的心率估计准确度优于对比方法,验证了其有效性。

To improve the accuracy of remote heart rate estimation in natural scenes,a multitask physiological prediction model based on multi-head self-attention was proposed.A feature representation algorithm was designed to convert the video into a spatiotemporal representation,suppressing the noise introduced by non-skin area and background.The improved multi-head self-attention model was used to learn the dependencies between long-distance frames to solve the problem of insufficient robustness of rPPG to illumination changes and head motion.The extracted features were used to estimate the rPPG signal and HR,and HR was constrained by the frequency domain of the rPPG signal,thereby improving the accuracy of HR estimation.Experimental results show that the proposed method outperforms the comparative methods in heart rate estimation on two public datasets,UBFC-RPPG and PURE,which verifies the effectiveness of the proposed method.