Privacy-preserving remote heart rate estimation using Block-wise Frequency Domain Transformation

Abstract

Remote heart rate estimation has attracted much attention in recent years in fields such as medical monitoring, mental health assessment, and exercise monitoring due to its non-contact characteristic. Most remote heart rate estimation methods usually adopt remote photoplethysmography (rPPG) to obtain the blood volume pulse (BVP) signal from the facial video. However, facial information is highly sensitive and poses a significant risk of privacy leakage. Currently, most face privacy preserving methods suitable for remote heart rate estimation are based on facial perturbations, if these methods are used to protect facial information, the accuracy of heart rate estimation might decrease. To address the above problems, we proposed a Privacy-Preserving Remote Heart Rate Estimation (PP-RHRE) architecture, with Block-wise Frequency Domain Transformation (BFDT) privacy-preserving method. This BFDT method can balance face privacy preserving effect and original remote heart rate estimation model accuracy. Specifically, we crop each raw image of the facial videos according to the face area for removing environmental visual information without rPPG signals. Subsequently, we divide the images into blocks, apply discrete cosine transform (DCT) on the each block, and then we discard some high-frequency parts to blur the images and enhance the representation of the BVP signal. Extensive experiments demonstrate that our method can improve original model accuracy while protecting face privacy in most BFDT scheme.