PaperlayerView exact revision on arXiv

Source-linked AI summary

DistancePPG: Robust non-contact vital signs monitoring using a camera

Mayank Kumar, Ashok Veeraraghavan, Ashutosh Sabharval

arXiv:1502.08040v2cs.CV

TL;DR

Camera-based vital-sign monitoring offers a non-contact alternative but is challenged by darker skin tones, low lighting, and motion. DistancePPG combines tracked facial-region color signals using automatically determined weights based on light intensity and blood perfusion, improving PPG SNR and reducing pulse-rate estimation error. The paper evaluates this approach across diverse skin tones, lighting conditions, and motion scenarios while identifying frame-rate and large-motion limitations.

  • Problem

    Camera-based vital-sign monitoring is desirable but challenged by weak signals under darker skin tones or low lighting and by motion artifacts.

  • Method

    DistancePPG tracks facial regions and combines their color-change signals with automatically determined weights reflecting incident light intensity and blood perfusion.

  • Results

    DistancePPG improved estimated-PPG SNR by 4.1 dB on average across skin tones and reduced pulse-rate error across stationary and motion scenarios.

  • Takeaways & Limitations

    The evaluated method improves camera-based pulse-rate and pulse-rate-variability estimation across diverse skin tones, lighting conditions, and natural motion scenarios.

  • Takeaways & Limitations

    PRV RMSE is constrained by the 30 fps camera’s ±16.6 ms peak-detection ambiguity, while large talking motions can make PPG estimates inaccurate.

Abstract

from arXiv · show

Vital signs such as pulse rate and breathing rate are currently measured using contact probes. But, non-contact methods for measuring vital signs are desirable both in hospital settings (e.g. in NICU) and for ubiquitous in-situ health tracking (e.g. on mobile phone and computers with webcams). Recently, camera-based non-contact vital sign monitoring have been shown to be feasible. However, camera-based vital sign monitoring is challenging for people with darker skin tone, under low lighting conditions, and/or during movement of an individual in front of the camera. In this paper, we propose distancePPG, a new camera-based vital sign estimation algorithm which addresses these challenges. DistancePPG proposes a new method of combining skin-color change signals from different tracked regions of the face using a weighted average, where the weights depend on the blood perfusion and incident light intensity in the region, to improve the signal-to-noise ratio (SNR) of camera-based estimate. One of our key contributions is a new automatic method for determining the weights based only on the video recording of the subject. The gains in SNR of camera-based PPG estimated using distancePPG translate into reduction of the error in vital sign estimation, and thus expand the scope of camera-based vital sign monitoring to potentially challenging scenarios. Further, a dataset will be released, comprising of synchronized video recordings of face and pulse oximeter based ground truth recordings from the earlobe for people with different skin tones, under different lighting conditions and for various motion scenarios.

1. Introduction

Camera-based vital-sign monitoring could avoid contact sensors but remains difficult under darker skin tones, low lighting, and motion. DistancePPG addresses these challenges by combining tracked facial-region signals with automatically determined weights, improving SNR and vital-sign estimation accuracy.

  • Motivation: Contact sensors are inconvenient in some settings, including NICU care, motivating non-contact vital-sign monitoring.Contact probes can cause skin damage in premature babies.
  • Challenges: Camera-based monitoring is feasible but performs poorly for darker skin tones, low lighting, and subjects who are not nearly at rest.
  • Method: DistancePPG combines color-change signals from different facial regions using a weighted average to improve camera-based PPG SNR.
  • Method: The algorithm automatically determines weights from the subject video to capture incident-light intensity and underlying blood perfusion.
  • Method: DistancePPG tracks facial regions separately during movement using a deformable face tracker and KLT feature tracker.
  • Results: 4.1 dB average SNR improvement across pale-white-to-brown skin tones increased to 6–7 dB for non-white skin and 4.5 dB across three natural motion scenarios.
  • Results: Pulse-rate estimation achieved mean bias −0.02 bpm with 95% limits of agreement −0.75 to 0.72 bpm across 12 subjects, compared with −0.40 bpm and −4.5 to 3.7 for prior methods.

2. Background and problem definition

The camera records reflected facial intensity over time, which combines illumination with skin reflectance. Because illumination is treated as constant during the estimation window, temporal intensity changes are used to estimate skin-reflectance changes related to PPG.

  • Video representation: A camera-based PPG system records video frames as intensity signals V(x,y,t) over a two-dimensional pixel grid.
  • Signal model: Measured reflected-light intensity is decomposed into illumination intensity and skin-surface reflectance.
  • Assumption: PPG estimation generally assumes illumination remains constant during the 5–60 second estimation window.
  • Signal model: Skin reflectance contains surface reflection and subsurface reflection, or backscattering.
  • PPG mechanism: Subsurface light is absorbed by blood chromophores, whose cardiac-cycle volume changes modulate the reflected signal.
  • Signal interpretation: With constant illumination, temporal recorded-intensity changes are proportional to skin-reflectance changes, though surface reflection generally dominates.

3. Key challenges and insights for camera-based PPG estimation

Camera PPG signals are weak and vulnerable to motion artifacts, while signal quality varies substantially across facial regions. DistancePPG therefore selects and weights regions according to their signal quality, while tracking them during movement.

  • Challenge 1: Very low signal strength: Camera PPG has very low signal strength because only a small blood volume undergoes a small cardiac-synchronous change.The paper reports that the skin vascular bed contains 2–5% of total blood and blood volume changes by about 5%.
  • Challenge 1: Very low signal strength: PPG quality differs substantially across facial patches, so whole-face averaging can combine stronger and weaker estimates indiscriminately.
  • Regional variation: A region’s useful signal depends on incident light intensity and local blood perfusion, while poorly perfused patches contribute more noise than signal.
  • Regional selection: An automated method is needed to reject noisy facial regions and combine retained regions for maximum SNR without ground-truth PPG.
  • Challenge 2: Motion artifact: Because the face is non-rigid, different regions should be tracked independently during movements such as tilting, smiling, shifting, and talking.
  • Challenge 2: Motion artifact: Motion through spatially varying illumination violates the constant-illumination assumption and corrupts PPG estimates even when tracking is accurate.
  • Challenge 2: Motion artifact: Small motion-induced changes in light direction can produce large surface-reflectance changes that overwhelm the cardiac subsurface signal.
  • Motion handling: Regions showing large intensity changes can be rejected because those changes are mostly attributed to incident-light variation rather than PPG.

4. DistancePPG: camera based PPG estimation algorithm

DistancePPG combines filtered PPG signals from tracked facial regions using video-derived goodness weights, while rejecting motion- or illumination-corrupted regions.

  • MRC algorithm: DistancePPG combines region-level camera signals with maximum-ratio weighting to improve the overall PPG signal-to-noise ratio.The method treats facial regions as channels containing different strengths of the same PPG signal and different noise levels.
  • MRC algorithm: Each facial ROI is spatially averaged, temporally bandpass filtered from 0.5Hz to 5Hz, and combined into a final estimate over T seconds.Filtering suppresses out-of-band surface-reflection components and other noise before combination.
  • Region based motion tracking algorithm: The algorithm tracks facial regions separately during movement and rejects regions with unusually large amplitudes that indicate artifacts.The tracking design combines a deformable face tracker with KLT feature tracking, while amplitude thresholding removes corrupted regions.
  • MRC algorithm: The method assumes PPG components from nearby facial regions are locally coherent because measured inter-region delays are less than 10 ms.This delay is within one sample period for cameras operating at 30–60 Hz.
  • Estimating goodness metric: The goodness metric estimates regional signal quality from power near the pulse rate relative to power across the filter passband.A coarse pulse-rate estimate is obtained from the spectral peak before computing the regional metric.
  • Estimating goodness metric: Goodness weights adapt to regional illumination and blood perfusion using only the recorded face video.Forehead and cheek regions produced stronger estimates than mouth and eye regions, with goodness values of 5.52, 2.30, 0.89, and 0.00 respectively.

5. Dataset and Performance Evaluation

The evaluation compares distancePPG with prior camera-based methods using synchronized camera and earlobe pulse-oximeter recordings across skin tones, motion, and lighting conditions.

  • Data acquisition: The study records synchronized monochrome green-channel camera video and earlobe pulse-oximeter PPG for comparison.The camera operates at 30 frames per second, while the pulse oximeter samples at 500 Hz from a 0.5 m distance.
  • Evaluation design: The experiments characterize MRC and region-based motion tracking against face averaging while varying skin tone, motion, and ambient light intensity.The same raw video feed is used for evaluating the compared methods.
  • Dataset: The dataset includes static recordings from 12 subjects spanning pale-white to dark brown/black skin tones.Subjects faced the camera for 40 seconds, with involuntary motions unrestricted.
  • Evaluation scenarios: Additional recordings cover reading, watching video, talking, illumination from 50 to 650 lux, and harsh color-video comparisons with ICA-based methods.These scenarios represent natural user motion and varying lighting conditions relevant to camera-facing device use.
  • Dataset: The resulting public dataset provides synchronized facial video and earlobe pulse-oximeter ground truth for varied skin tones, lighting, and motion scenarios.The authors state that other researchers can access the released dataset.
  • Metrics: Performance is quantified using camera-to-earlobe PPG SNR and pulse-rate accuracy, with an amplitude-independent SNR definition.The SNR measure is intended to capture waveform similarity despite independent sensor architectures and analog gains.
  • Metrics: Pulse rate is estimated from the highest-power FFT frequency in overlapping 10-second windows and compared with synchronously acquired contact PPG.Each window uses a Hamming window with 5-second overlap, and PR is converted as PR = 60·fPR bpm.

6. Results

DistancePPG improves camera-based PPG quality and vital-sign estimation across skin tones, lighting conditions, and small-to-medium motion. Performance degrades under large non-rigid motion, especially talking, where motion artifacts remain substantial.

  • Performance for different skin tones: 4.1 dB average SNR improvement was observed across skin tones, with approximately 6–7 dB gains for darker skin tones.Figure 4 compares light/fair, medium/olive, and brown/dark skin-tone categories.
  • Performance for different skin tones: −0.02 bpm mean bias with 95% limits of agreement from −0.75 to 0.72 bpm was achieved for pulse-rate estimation across 12 subjects.Face averaging produced −0.40 bpm bias with limits from −4.5 to 3.7 bpm, excluding four extreme outliers.
  • Performance for different skin tones: PRV estimation improved substantially, with light- and medium-skin-tone RMSE below 16 ms using distancePPG.At 30 fps, half the sampling interval is 16.6 ms, limiting how much lower the RMSE can reasonably become.
  • Performance under various motion scenario: 4.5 dB average SNR improvement was obtained across stationary, reading, watching-video, and talking scenarios, with clearer PPG during motion up to 4 px per frame.Under motion of at least 5 px per frame, uncompensated motion artifacts degrade distancePPG performance.
  • Performance under various motion scenario: Pulse-rate estimation under motion improved from 7.17 bpm mean bias with limits of −18.70 to 33.04 bpm to 0.48 bpm with limits of −5.73 to 6.70 bpm.For non-talking motion, the mean bias was 0.17 bpm with limits of −4.52 to 4.86 bpm; talking remained difficult because of large non-rigid motion.
  • Lighting and comparison with ICA-based method: Lighting-dependent SNR gains averaged 1.9 dB for pale white skin and 6.5 dB for brown skin, while talking produced low SNR and large pulse-rate errors.In the talking scenario, distancePPG had −4.62 dB mean SNR and 14.52±12.82 bpm mean pulse-rate error.

7. Discussion

The discussion explains how distancePPG uses signal-quality information and face-region weighting to preserve PPG waveform quality, while identifying sampling and motion limits.

  • Goodness metric as signal quality index: Goodness metric is a good substitute for SNR above −3 dB, but overestimates SNR at lower values.Regions with Gi < −6 dB or Gi < 0.25 are rejected during PPG estimation.
  • Goodness metric as signal quality index: Rejecting regions with Gi < 0.25 improves overall camera-based PPG SNR by 0.2 dB.
  • Goodness metric as signal quality index: Goodness metric estimates regional PPG signal quality from spectral structure centered on the pulse rate.The metric exploits PPG periodicity and its dominant pulse-rate frequency.
  • PRV estimation using camera PPG: Figure 9 compares PRV peak-timing RMSE and missing-peak percentages across running, watching, talking, and stationary conditions.The plots cover five subjects with varying skin tones and compare camera PPG against pulse-oximeter ground truth.
  • Goodness metric as signal quality index: Narrow filtering around pulse rate cannot preserve the wider 0.5Hz–5Hz PPG waveform needed for inter-beat and peak information.
  • Goodness metric as signal quality index: DistancePPG re-estimates regional goodness because illumination changes over time and affects the face-wide quality profile.The profile reflects both spatial illumination and blood-perfusion-dependent modulation strength.
  • PRV estimation using camera PPG: At 30 fps, stationary PRV RMSE for fairer skin tones remains around 16–20 ms because frame spacing creates ±16.6 ms peak-detection ambiguity.Interpolation cannot fully recover details lost through the low sampling rate.
  • PRV estimation using camera PPG: Higher frame rates could reduce PRV timing error but shorten exposure and lower camera-PPG SNR, making pulse-minimum detection harder.DistancePPG’s algorithmic SNR gain may compensate for some SNR loss from higher frame rates.

8. Conclusion

The conclusion reports that distancePPG improves camera-based vital-sign estimation across skin tones, lighting conditions, and natural motion, while acknowledging substantial limitations during talking.

  • Conclusion: DistancePPG improves vital-sign estimation across pale-to-dark skin tones, 50–650 lux lighting, and natural reading, watching, and talking scenarios.The method accounts for regional blood perfusion and incident light and separately tracks face regions during motion.
  • Limitations: During talking, large in-band changes in skin-surface reflectance can obscure the smaller subsurface changes encoding the PPG signal.The paper reports inaccurate PPG estimates after rejecting much of the face region during large motion.
  • Applications: DistancePPG is intended to extend camera-based vital-sign monitoring toward mobile-phone and computer applications beyond users resting in well-lit conditions.The authors report developing a real-time PC application and plan mobile-platform implementations.
Loading 1502.08040v2…