Remote respiratory rate measurement from torso video in challenging scenarios

Remote respiratory rate (RR) measurement methods can continuously and unobtrusively monitor human pulmonary rhythms, holding great potential for applications in both clinical and home-based health monitoring. However, in challenging scenarios such as subject motion or drastic illumination changes, the weak respiratory signals are easily overwhelmed by noise, making existing methods prone to failure. To address these challenges, this paper proposes a robust RR estimation framework based on a reflection model of torso video. First, we establish a reflection model of the torso region and construct a plane orthogonal to the white-light direction to suppress the interference induced by light intensity variations. Then, the torso is divided into multiple regions of interest (ROIs), and principal component analysis (PCA) is applied to reconstruct the common respiratory components across regions while suppressing local random motion noise. Finally, to address the uncertainty in the ordering of PCA components, we construct a respiratory signal selection index that integrates both time- and frequency-domain features; the RR estimation results are then refined based on the continuity of physiological parameter variations, further improving the practicality and stability of the algorithm. Results on our private dataset across multiple challenging scenarios validate the robustness of the proposed algorithm, while results on public datasets demonstrate its generalization ability.