Remote Estimation of Heart Rate Based on Multi-Scale Facial ROIs

While most rPPG approaches extract the pulse signals based on single facial region of interest (ROI), this research proposes a new method to extract pulse signals from ROIs with multiple scales. The idea is that rich pulse features can be extracted by varying ROI scales and combining these features would contribute to the accuracy improvement. The proposed framework consists of three main steps: 1) constructing facial ROI pyramid with multiple scale levels, 2) blood volume pulse (BVP) signals extraction, and 3) signal fusion using convex combination with Gaussian and uniform priors, respectively. This paper also investigates how the commonly used algorithms perform under multi-scale ROIs. Experiments were conducted using one publicly available dataset and one self-collected dataset. The results show that the ROI with a size slightly smaller than the face boundary achieves on average higher measurement accuracy. The high-quality pulse signal appears not consistently in one scale level but rather in multiple levels according to measurement environments and motion statuses. Therefore, the fusion of multiple pulse signals is beneficial to the measurement accuracy improvement.