Adaptive Parameter Optimization for Robust Remote Photoplethysmography

AI summary

Overview

Research area: Computer vision / remote photoplethysmography (rPPG) — contactless heart-rate estimation from ordinary RGB video, positioned within time-series-for-health research.

Technical level: Intermediate. The paper assumes familiarity with signal processing concepts (FFT, spectral power, smoothing splines, chrominance projections), but the core idea is described at the level of a general technical reader.

Scope: A single training-free algorithm (PRISM) for extracting heart rate from facial video, evaluated on two standard benchmarks (PURE and UBFC-rPPG) against unsupervised and supervised baselines.

Paper details: arXiv:2511.21903v1 [cs.CV], 26 Nov 2025, CC BY 4.0, workshop track "Learning from Time Series for Health." Authors: Cecilia G. Morales, Fanurs Chi-En Teh, Kai Li, Pushpak Agrawal, Artur Dubrawski — affiliations Carnegie Mellon University (USA), University of Toronto (Canada), and Vellore Institute of Technology. Morales, Teh, and Li are listed as co-first authors; the paper states the first two contributed equally.

What This Paper Is About

Remote photoplethysmography estimates a person's pulse by detecting tiny color changes in the skin across video frames. The step that matters most is how the raw red, green, and blue channel signals are combined into one pulse signal — and most existing unsupervised methods use fixed parameters chosen for particular lighting and camera conditions. This paper introduces PRISM (Projection-based Robust Signal Mixing), which keeps the interpretable structure of the existing POS algorithm but replaces its fixed numbers with parameters selected online, per video window, based on a measure of signal quality.

Key Contributions

1. A training-free adaptive rPPG method. PRISM jointly optimizes two parameters — a color-mixing weight α and a spline smoothness parameter λ — instead of leaving them fixed, requiring no training data or labeled examples.

2. A signal-quality objective for parameter selection. The method minimizes k · TV(λ, α) − C(λ, α), combining spectral concentration C (fraction of signal power in the plausible heart-rate band) with temporal variation TV (how much the windowed heart-rate estimate jumps between consecutive windows), where k = 1/3 is set empirically. The paper states this objective correlates strongly with lower heart-rate error across datasets.

3. State-of-the-art unsupervised benchmark results. MAE of 0.77 bpm on PURE and 0.66 bpm on UBFC-rPPG, with accuracy at a ±5 bpm threshold of 97.3% and 97.5% respectively (the abstract's figures; Table 1 lists 97.2% for PURE while the ablation table lists 97.3%).

4. A quantified set of implementation improvements to the rPPG-Toolbox. FFT bins increased from N=2⁹=512 (≈3.5 bpm resolution) to N=2¹⁴ (≈0.1 bpm), 60-second windows replaced with 10-second non-overlapping windows, YOLOv5 face detection replacing HaarCascade, dynamic rather than static face tracking, and largebox = False. Appendix D separates these gains from PRISM's own algorithmic contribution.

Main Findings

• Best among unsupervised methods on both benchmarks. PRISM reports 0.77 bpm MAE and 1.69 bpm RMSE on PURE, versus the previous best unsupervised method POS at 2.23 bpm MAE and 8.00 RMSE — described as a 65% MAE reduction and 79% RMSE reduction. On UBFC-rPPG, PRISM reports 0.66 bpm MAE and 1.54 bpm RMSE against POS's 1.08 bpm MAE, with a stated 65% RMSE improvement.

• Accuracy at the ±5 bpm threshold. PRISM reports 97.3% on PURE and 97.5% on UBFC-rPPG in the abstract; Table 1 reports 97.2% for PURE and 97.5% for UBFC-rPPG. The ablation table (Table 2) lists 97.3% for the full model on PURE. The paper does not reconcile these PURE figures.

• Statistically comparable to supervised methods on UBFC-rPPG. Paired t-tests give p=0.208 versus TS-CAN, p=0.880 versus PhysFormer, and p=0.595 versus DeepPhys. PRISM matches DeepPhys's 97.5% accuracy and approaches TS-CAN's 0.51 bpm MAE on that dataset. On PURE, the reported supervised MAEs are considerably worse (DeepPhys 9.42, PhysNet 9.33, TS-CAN 4.54, PhysFormer 11.58, EfficientPhys-C 6.55).

• Optimal fixed parameters differ by dataset — evidence for adaptation. The best fixed α is 0.6 for PURE but 0.8 for UBFC-rPPG, and the best fixed λ is 0.5 for PURE but 0.05 for UBFC-rPPG, which the authors present as verification that a single fixed setting cannot generalize.

• Both objective terms are necessary. Removing the temporal-variation penalty (k=0, "Concentration-only") raises PURE MAE to 4.02 bpm and UBFC-rPPG MAE to 1.83 bpm. Removing the concentration term (C=0, "TV-only") is far worse: 14.30 bpm MAE on PURE and 4.74 bpm on UBFC-rPPG. The authors hypothesize the concentration term acts as a frequency-domain constraint separating genuine pulse signals from smooth but non-physiological drift such as illumination change.

• Fixing one parameter can lower RMSE but not MAE. Best fixed α gives 0.82 MAE / 1.26 RMSE / 97.5% accuracy on PURE, and 0.70 / 1.15 / 97.6% on UBFC-rPPG; best fixed λ gives 0.99 / 1.74 / 96.4% and 1.04 / 1.94 / 96.6% respectively. The authors attribute the lower RMSE to avoidance of occasional large errors from one subject, while the higher MAE indicates reduced overall accuracy.

• Robust across motion conditions on PURE. Per-condition MAE ranges from 0.40 bpm (small head rotation) to 1.48 bpm (talking), with 100% accuracy at ±5 bpm for steady sitting, slow translation, small head rotation, and medium head rotation; talking scores 88.89% and fast translation 93.75%.

• Large gains from toolbox modifications, and PRISM still helps. Under the standard rPPG-Toolbox configuration PRISM achieves 2.66 bpm MAE on PURE and 3.10 bpm on UBFC-rPPG; the enhanced configuration brings 0.77 bpm (71% improvement) and 0.66 bpm (79% improvement). POS improves 39% on PURE and 75% on UBFC-rPPG under the same changes. TS-CAN degrades 23% on PURE (3.69 to 4.54 bpm) but improves 60% on UBFC-rPPG (1.29 to 0.51 bpm), which the authors suggest indicates supervised models may be tuned to standard toolbox characteristics.

• Harmonic handling. Because harmonic frequencies can dominate rPPG spectra and yield predictions at twice the true rate, PRISM evaluates two bands — high [0.75 Hz, 4.0 Hz] and low [0.5 Hz, 3.0 Hz] — and selects the low-band signal when the high-band mean estimate is approximately twice the low-band mean.

Methodology in Plain English

The pipeline starts by detecting and tracking a face rectangle in each video frame with YOLOv5, averaging pixel intensities inside it to get one red, one green, and one blue time series. Each channel is divided by a smooth spline baseline fit to it, which strips out slow brightness drift from lighting and camera gain while keeping the fast pulse ripples. The smoothing spline's stiffness is governed by λ.

For combining channels, PRISM keeps POS's green-minus-red/blue shape but writes it with a single adjustable weight: s(t) = Ĝ(t) − (α·B̂(t) + (1−α)·R̂(t)). When α is near 1 the formula behaves like POS's primary projection (Ĝ − B̂); when α is small it moves toward the secondary projection (Ĝ + B̂ − 2R̂).

The video is then cut into 10-second non-overlapping windows. For every candidate (λ, α) pair on a small grid — α ∈ {0.5, 0.6, 0.7, 0.8, 0.9, 1.0} and λ ∈ {0.01, 0.05, 0.1, 0.5, 1.0} — the method builds the pulse signal, takes an FFT, reads off a heart-rate estimate per window, and scores the pair with the combined objective. The winning pair is fixed and used for the final estimates. Because α values below 0.5 were seldom optimal, they are excluded from the search. Because the decision depends only on recent windows, the method runs online: after roughly one minute of initialization it settles on stable parameters and can either keep updating them or hold them if conditions stay constant. The paper notes inference can start immediately with predefined α and λ values.

Why This Matters

Research impact. The paper argues that adaptive time-series optimization — not just bigger trained models — can close much of the gap between unsupervised and supervised rPPG. It also shows that benchmark numbers for all methods depend heavily on preprocessing choices, since the toolbox modifications lift every unsupervised method and affect supervised methods inconsistently across datasets. That is a methodological caution for anyone comparing rPPG results across papers.

Real-world applications (the first two are the ones named in the paper; the others follow from its framing of autonomous and human-in-the-loop care):

• Telemedicine consultations, where a patient's pulse could be read from a webcam without any contact device.
• Detecting driver drowsiness from in-cabin cameras.
• Deployment on robot-mounted or outdoor platforms, which the authors list as ongoing work.
• Any setting with uncontrolled or shifting illumination, where the paper argues fixed projection axes fail.

Industry relevance. PRISM requires no training dataset and no GPU, runs in real time on CPU, and is designed as a drop-in replacement for projection-based pipelines such as POS. The authors also commit to integrating it into the rPPG-Toolbox repository for reproducibility and easier access. Lower data and compute requirements reduce the cost of deploying contactless vital-sign monitoring at scale.

Funding. The work was partially supported by the Defense Advanced Research Projects Agency (award HR00112420329) and the National Science Foundation (awards 2427948 and 2406231).

Future Directions

• Evaluating across diverse skin tones. The conclusions state that the current evaluation focuses primarily on lighter skin tones present in the tested datasets, and that more comprehensive testing is needed before generalizability can be claimed.

• Testing in uncontrolled environments. The authors report ongoing work on outdoor and robot-mounted scenarios where environmental conditions are completely uncontrolled.

• Toolbox integration. PRISM is planned for integration into the rPPG-Toolbox repository to improve access and reproducibility.

• Open questions the paper leaves implicit. Since the best fixed parameters differ so much between PURE and UBFC-rPPG, it is not yet established how the online search behaves on datasets or populations outside these two benchmarks, how often parameters should be re-updated when conditions change, or why supervised methods like TS-CAN respond so unevenly to the same preprocessing improvements.

Target Audience

Researchers and engineers working on camera-based vital-sign monitoring, rPPG, or physiological time-series analysis, particularly those who need a method that runs without training data or GPU resources. It is also useful for practitioners comparing supervised and unsupervised approaches on the PURE and UBFC-rPPG benchmarks, and for anyone evaluating how sensitive rPPG benchmark results are to preprocessing and toolbox configuration.

Authors’ abstract

Remote photoplethysmography (rPPG) enables contactless vital sign monitoring using standard RGB cameras. However, existing methods rely on fixed parameters optimized for particular lighting conditions and camera setups, limiting adaptability to diverse deployment environments. This paper introduces the Projection-based Robust Signal Mixing (PRISM) algorithm, a training-free method that jointly optimizes photometric detrending and color mixing through online parameter adaptation based on signal quality assessment. PRISM achieves state-of-the-art performance among unsupervised methods, with MAE of 0.77 bpm on PURE and 0.66 bpm on UBFC-rPPG, and accuracy of 97.3\% and 97.5\% respectively at a 5 bpm threshold. Statistical analysis confirms PRISM performs equivalently to leading supervised methods ($p > 0.2$), while maintaining real-time CPU performance without training. This validates that adaptive time series optimization significantly improves rPPG across diverse conditions.