On the Vector Space in Photoplethysmography Imaging

Christian S. PilzWorkshops @ ICCV 2019 · DOI · arXiv

Introduction

Measuring heart rate from a webcam video—a process known as Photoplethysmography Imaging (PPGI)—is a significant challenge in non-contact health monitoring. The primary obstacle is that the subtle pulse signal, caused by minute blood volume changes in the skin, is often overwhelmed by "noise" from head movements, talking, or fluctuating ambient light. Traditional methods attempt to filter this noise using linear projections or blind source separation (e.g., ICA), but these often require specific priors about skin tone or environmental conditions.

This paper proposes a fundamental shift in how we represent PPGI data. By moving away from standard Euclidean linear algebra and adopting a geometric approach based on Group Theory and Differential Geometry, the authors introduce a parameter-free feature operator that maps RGB pixel data onto a Riemannian manifold (the unit sphere). This change in the underlying geometry makes the system inherently robust to motion without needing complex manual tuning.

Methodology: The Spherical Mapping (SPH) Operator

The core of the proposed method is the SPH operator, which treats motion as a group of actions affecting pixel intensities. To achieve motion invariance, the authors map the distribution of skin pixel intensities onto a Riemannian submanifold—specifically, the unit sphere ($S^2$).

The SPH pipeline: mapping raw RGB distributions to a spherical manifold and extracting coordinates.

In this spherical space, the "distance" between measurements is defined by geodesics (curves on the sphere's surface) rather than straight lines. This topological transformation naturally unifies several previous projection-based approaches into a single, implicit, closed-form operator. Furthermore, the authors employ a stochastic model—a Diffusion process—to treat the heart rate as a quasi-periodic signal evolving over time, allowing the system to maintain a lock on the pulse even in noisy data.

Key Insights: Geometry as a Solution

The central thesis of this work is that motion robustness can be achieved through topology rather than just filtering. By changing the "shape" of the feature space from a flat Euclidean cube to a sphere, the mathematical framework can ignore certain types of noise by default.

Conceptual representation of the vector space, illustrating the separation of pulse, color, and motion signals.

Mapping to the unit sphere effectively normalizes intensity variations. Since the pulse signal manifests as a rhythmic change in the direction of the color vector rather than just its magnitude, the spherical representation is a natural fit for capturing these dynamics while remaining invariant to translational motion noise.

Experimental Evaluation

The researchers evaluated the SPH operator on two prominent public databases: UBFC-RPPG (150 sequences with moderate head motion) and LGI Multi-Session (challenging scenarios including rotation, gym exercise, and outdoor conversation).

The SPH operator demonstrated high competitiveness against state-of-the-art methods like SSR (Spatial Subspace Rotation) and POS (Plane-to-Orthogonal-to-Skin). In the "LGI Resting" session, SPH achieved a Pearson correlation of 0.71 and a Root-Mean-Square Error (RMSE) of 1.49 BPM, outperforming all other tested methods. While it performed similarly to the more complex LGI (Local Group Invariance) method, SPH is significantly simpler to implement and requires no parameter tuning.

Limitations and Future Directions

Despite its robustness, the SPH operator faces two main hurdles. First, it is primarily designed to be invariant to translations in the feature space and does not yet account for more complex transformations. Second, extreme uncontrolled lighting remains a significant challenge; in the "LGI Talk" outdoor scenario, all tested algorithms, including SPH, saw a drop in performance.

Future work will focus on expanding the group of transformations the operator can handle. To support further research, the authors have released their code as an open-source MATLAB toolbox.

Summary updated

BibTeX
@inproceedings{pilz2019vector,
  author = {Christian S. Pilz},
  title = {{On the Vector Space in Photoplethysmography Imaging}},
  booktitle = {Workshops @ ICCV},
  year = {2019},
  doi = {10.1109/iccvw.2019.00197},
}
Figure