Abstract

Remote photoplethysmography (rPPG) enables non-contact heart rate monitoring from facial videos but it is highly susceptible to motion artefacts, illumination changes, and sensor noise. A framework combining the Plane Orthogonal-to-Skin (POS) method and Singular Spectrum Analysis (SSA) was proposed in this work to address these challenges by first projecting normalised RGB signals onto a skin-tone–orthogonal subspace to suppress illumination and motion distortions and then decomposing the resulting signal into components that isolate physiologically meaningful oscillations. Evaluation on the PFF and UBFC-Phys dataset demonstrates that this approach consistently outperforms conventional single-channel, statistical, and chrominance based methods by achieving a mean absolute error of 4.99 bpm and correlation of 0.76 on PFF, and a mean absolute error of 4.11 bpm with correlation of 0.86 on UBFC-Phys. Furthermore, the comparison with results reported in the existing literature indicates that the proposed framework achieves competitive accuracy relative to popular learning based rPPG approaches. These findings indicate that integrating chrominance projection with adaptive temporal decomposition significantly improves robustness and accuracy for contact-free heart rate estimation.