Design, Implementation and Evaluation of a Real-Time Remote Photoplethysmography (rPPG) Acquisition System for Non-Invasive Vital Sign Monitoring

arXiv:2508.18787 · cs.CV · Submitted 2025-08-26 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Design, Implementation and Evaluation of a Real-Time Remote Photoplethysmography (rPPG) Acquisition System for Non-Invasive Vital Sign Monitoring".

Jane: Here's the gist of the paper. The paper presents a real-time remote photoplethysmography (rPPG) system optimized for low-power devices to extract vital signs like heart rate (HR), respiratory rate (RR), and oxygen saturation (SpO2) from facial video streams, addressing the challenges of real-time deployment on resource-constrained platforms System Architecture and Design The system is built on a multithreaded architecture to manage video capture, real-time processing, network communication, and graphical user interface (GUI) updates concurrently The architecture follows a modular pipeline design where the core rPPG system operates sequentially, with each component performing a specific task in sequence The system is organized into four primary functional blocks: the collaborative UI, video capture, real-time processing and analysis, and the network interface The main thread coordinates system operations by using a QTimer class from the Qt framework to trigger the main loop every 33 milliseconds to maintain a consistent frame rate of 30 frames per second Signal Extraction Pipeline The extraction process utilizes the unsupervised, non-learning-based Face2PPG pipeline which integrates computer vision and signal processing techniques This pipeline begins with face detection and alignment, followed by skin segmentation to isolate regions of interest (ROIs) in the facial region for signal extraction Dynamic ROI selection is used to identify areas with optimal signal quality while discarding regions with occlusion, low contrast, or potentially poor signals The RGB signals are transformed into physiological signals (PPG signals) using the CIE-Lab color space for efficient computation and isolation of the chrominance channel Real-Time Optimization Techniques Several optimizations were implemented to ensure robust real-time performance on low-power devices These optimizations include: 1.

Tom: Using YuNet by Wu et al.

42: for face detection, which achieves detection rates of 909 fps for 224x224 (1.1 ms) on an Intel i7-12700K CPU 2.

Tom: Employing Ensemble of Regression Trees (ERT) and Local Binary Features (LBF) algorithms for fast, non-deep-learning-based face alignment 3.

Paper summary: Tom: So to wrap up the design and implementation evaluation of this paper, the core claim is that they built a real-time remote photoplethysmography system specifically optimized for low-power devices to extract vital signs like heart rate and oxygen saturation from facial video streams.

Jane: The paper argues that deploying such systems in real-time on resource constrained platforms is difficult because of issues related to scalability, interoperability, and performance, especially when dealing with mass-market sensor technologies <ref:2508.18787#pg2>.

Lu: They present a prototype architecture that uses a pipeline approach for continuous monitoring while focusing heavily on power consumption and memory usage to ensure robust real-time performance <ref:2508.18787#pg3>.

Meng: What’s the specific promise here? It's not just *a* system, it's one specifically designed to overcome those hard engineering hurdles in a practical setting <ref:2508.18787#pg3>.

Tom: The paper outlines the methodology for extracting physiological signals, which involves face detection, skin segmentation to isolate regions of interest for signal extraction, and then dynamic ROI selection based on signal quality <ref:2508.18787#pg3>.

Jane: They use the CIE Lab color space transformation to efficiently convert RGB signals into the physiological signals needed for heart rate and oxygen saturation measurements <ref:2508.18787#pg3>.

Lu: The optimization techniques they used were key: they employed YuNet for face detection, Ensemble of Regression Trees and Local Binary Features for alignment, a sixty-one tap Finite Impulse Response filter with linear phase design for filtering, and buffering with a C plus deque to hold up to three hundred sixty samples <ref:2508.18787#pg1>.

Meng: I’m looking at the efficiency gains here. They used YuNet for detection, which achieved detection rates of nine hundred nine frames per second on a two two four by two two four input size, which is about one point one millisecond <ref:2508.18787#pg1>. That’s pretty aggressive for low power.

Tom: And the results showed that the RT configuration two performed better than the RT configuration one, achieving one hundred three frames per second compared to seventy one frames per second <ref:2508.18787#pg3>.

Jane: So what this means for the listener is that they’ve developed a practical system that balances speed and accuracy for non-invasive vital sign monitoring in real-time, which is a big step forward <ref:2508.18787#pg3>.

Lalam: They conducted comprehensive speed and accuracy assessments using four benchmark datasets to demonstrate the system’s effectiveness and reliability in real-world scenarios for telemedicine and remote health monitoring <ref:2508.18787#pg3>.

Conclusion: Tom: So looking at the full paper, titled Design, Implementation and Evaluation of a Real-Time Remote Photoplethysmography Acquisition System for Non-Invasive Vital Sign Monitoring, the authors are showing how they built a system that directly tackles the practical challenges of deploying rPPG on low-power hardware.

Jane: The implication is that there’s now a more optimized blueprint for getting continuous, non-invasive health monitoring running on devices where battery and processing power are limited <ref:2508.18787#pg3>.

Lu: It moves the field by providing a concrete design, showing how to structure the pipeline to minimize resource usage while still extracting reliable heart rate and respiratory rate signals <ref:2508.18787#pg3>.

Meng: For me, what’s important is that they didn't just talk about theory; they showed the actual performance trade-offs between a fast real-time version and

Center for Machine Vision and Signal Analysis, University of Oulu · VTT Technical Research Centre of Finland

cs.CV

Submitted: 2025-08-26

Updated: 2026-11-11

Comments: 33 pages, 4 figures, 11 numbered equations, 6 tables, and 67 references. Evaluated on 4 public datasets and in 5 live 10-minute sessions on 2 computers. Code available at https://github.com/Multimodal-Sensing-Lab/rt-rppg

Code: https://github.com/Corvusoft/restbed

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 77/100

The gist: The paper presents a real-time remote photoplethysmography (rPPG) system optimized for low-power devices to extract vital signs like heart rate (HR), respiratory rate (RR)…

Key concepts

rPPG
Real-time remote photoplethysmography is a technique that measures subtle color changes in facial video caused by blood flow. This allows the system to estimate vital signs like heart rate and oxygen levels without needing physical sensors, making it suitable for non-invasive monitoring.
CIE-Lab Color Space
This is a specific color model used to transform standard RGB video colors into a more physiologically relevant space. By using this space, the system can efficiently isolate the chrominance channel, which holds the most useful information for extracting vital signs from skin color changes.
YuNet
YuNet is a specific computer vision model used in this system for face detection. It is chosen because it offers high speed and accuracy, allowing the system to quickly locate and track faces within video streams, which is crucial for real-time processing on low-power hardware.

Terminology

Summary

The paper presents a real-time remote photoplethysmography (rPPG) system optimized for low-power devices to extract vital signs like heart rate (HR), respiratory rate (RR), and oxygen saturation (SpO2) from facial video streams, addressing the challenges of real-time deployment on resource-constrained platforms

System Architecture and Design

The system is built on a multithreaded architecture to manage video capture, real-time processing, network communication, and graphical user interface (GUI) updates concurrently The architecture follows a modular pipeline design where the core rPPG system operates sequentially, with each component performing a specific task in sequence The system is organized into four primary functional blocks: the collaborative UI, video capture, real-time processing and analysis, and the network interface The main thread coordinates system operations by using a QTimer class from the Qt framework to trigger the main loop every 33 milliseconds to maintain a consistent frame rate of 30 frames per second

Signal Extraction Pipeline

The extraction process utilizes the unsupervised, non-learning-based Face2PPG pipeline which integrates computer vision and signal processing techniques This pipeline begins with face detection and alignment, followed by skin segmentation to isolate regions of interest (ROIs) in the facial region for signal extraction Dynamic ROI selection is used to identify areas with optimal signal quality while discarding regions with occlusion, low contrast, or potentially poor signals The RGB signals are transformed into physiological signals (PPG signals) using the CIE-Lab color space for efficient computation and isolation of the chrominance channel

Real-Time Optimization Techniques

Several optimizations were implemented to ensure robust real-time performance on low-power devices These optimizations include:

  1. Using YuNet by Wu et al. p. 5 [42] for face detection, which achieves detection rates of 909 fps for 224x224 (1. p. 51 ms) on an Intel i7-12700K CPU

  2. Employing Ensemble of Regression Trees (ERT) and Local Binary Features (LBF) algorithms for fast, non-deep-learning-based face alignment

  3. Using a 61-tap Finite Impulse Response (FIR) filter with a linear phase design for signal filtering

  4. Implementing buffering with a C++ deque to maintain up to 360 samples of the achannel values from the CIE-Lab color space for rolling computations

Performance Evaluation and Results

The system was evaluated under two configurations: Face2PPG-RT, designed for real-time performance on low-power systems, and Face2PPG-Server, designed for high accuracy in cloud environments Performance metrics included Mean Absolute Error (MAE), Root-Mean-Square Error (RMSE), and Pearson Correlation Coefficient (PCC) against four benchmark datasets The results show that RT Config. 2 performs better than RT Config. p. 2 1, achieving a commendable 103 FPS compared to 71 FPS for RT Config. p. 5

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Fast-track face detection by using optimized models like YuNet to achieve high frame rates, as YuNet has 85,000 parameters, achieving detection rates of 909 fps for 224x224 (1.1 ms). This speeds up the initial stage significantly compared to traditional methods.

  2. Enhance face alignment performance by employing fast non-deep-learning models like Ensemble of Regression Trees (ERT) and Local Binary Features (LBF) instead of computationally demanding Deep Alignment Network (DAN) models, as Our optimized models significantly outperform existing implementations of these algorithms [46]. This keeps the system responsive on low-power devices.

  3. Improve signal extraction accuracy by using a novel geometric segmentation method based on facial landmark points to create a spatiotemporal matrix of normalized faces, which ensures consistent signal measurement across frames, regardless of pose or movement. This stabilizes the input for subsequent processing steps.

  4. Optimize RGB to BVP transformation by computing CIE-Lab values only for the average RGB values within the ROI, as this reduces the computational cost by performing a single conversion per ROI rather than processing every pixel. This minimizes overhead while targeting blood-related signals.

  5. Implement a predictive stability flag within the buffering mechanism, where Once 8 seconds (240 samples) of uninterrupted data is collected, the system marks the signal as stable, to ensure that only uninterrupted and reliable data is used for further processing and transmission. This prevents spurious vital sign readings during brief occlusions.

  6. Utilize Welch’s method in spectral analysis when signal instability occurs, as it improves the reliability of frequency analysis when the signal is less stable or contains transient disturbances. This provides more robust HR and RR estimates in challenging scenarios like motion artifacts.

  7. Deploy a modular pipeline architecture where each block is optimized at both algorithm and data levels, including selecting algorithms with reduced computational complexity and applying pruning strategies to limit unnecessary computations. This ensures the system operates efficiently under low-power constraints.

  8. Implement a reactive programming model combining Functional Reactive Programming (FRP) and the Actor Model to create an event-driven design, which allows operations to be triggered dynamically by events like face detection, ensuring efficient use of computational resources. This maintains continuous operation without blocking threads.

  9. Integrate real-time quality metrics, specifically estimating face movement, into the processing pipeline so that only high-quality data is processed, thereby improving efficiency and speed by avoiding unnecessary computations on low-quality inputs. This dynamically adjusts processing based on input signal integrity.

Sources

Related papers