Abstract
Remote Photoplethysmography (rPPG) enables non-contact heart rate monitoring by detecting subtle skin color changes in video. As the technology transitions from controlled laboratory environments to real-world smartphone applications, the impact of hardware-induced artifacts and evaluation methodologies becomes paramount. This paper investigates three critical but often overlooked factors: rolling shutter effects, frame rate jitter, and temporal window sizes. Our findings indicate that rolling shutter can introduce phase shifts comparable to physiological signals, while inconsistent temporal windowing significantly complicates algorithmic benchmarking.
Introduction
While significant research has addressed major rPPG challenges such as motion and lighting variability, the field lacks standardization regarding hardware artifacts and evaluation metrics. This study systematically evaluates how CMOS sensor characteristics and data processing choices influence the accuracy and reliability of heart rate estimation.
Hardware Artifacts: Rolling Shutter and Jitter
Rolling Shutter Effects
Most modern smartphone cameras utilize CMOS sensors with a rolling shutter, which scans the image line-by-line. This progressive scanning introduces a time delay across a single frame. By utilizing a custom LED modulator to replay precise PPG signals, we measured the resulting phase shifts across various smartphone models.
The results show phase shifts of approximately 0.02 seconds along the horizontal axis. While seemingly negligible, this shift corresponds to roughly 10 degrees at a standard heart rate—a magnitude similar to real physiological phase shifts used in medical diagnostics (e.g., studying pulse transit time). Without proper compensation, these hardware artifacts could be misinterpreted as biological data.
The Jitter Myth
Smartphones frequently record video with irregular frame rates, a phenomenon known as "jitter." It has been hypothesized that this jitter could corrupt the frequency analysis required for heart rate estimation. However, our analysis of "timestamp-aware" interpolation versus constant frame rate assumptions suggests that for current standard algorithms, the impact is minimal.
As shown in the spectral analysis, the distortion caused by jitter is orders of magnitude smaller than the primary pulse signal. Consequently, frame rate irregularity is largely an inessential factor for heart rate estimation in typical smartphone use cases.
Evaluation Inconsistency: The Windowing Problem
A significant source of variance in rPPG research stems from the choice of temporal window size used to calculate "ground truth" heart rates. Heart rate is an average over time, and its value fluctuates depending on the duration of the observation window.
Using the Fantasia Database, we simulated heart rate calculations across windows ranging from 5 to 60 seconds. We found that simply varying the window size can alter the calculated heart rate by 4 to 10 beats per minute (bpm), or up to 10% of the total value. This variability means that head-to-head comparisons between different rPPG methods are invalid unless the researchers utilize identical window sizes.
Conclusion
This investigation highlights the need for greater technical transparency in the rPPG community. Hardware artifacts like rolling shutter must be accounted for when studying fine-grained physiological signals. Furthermore, the standardization of evaluation parameters—specifically temporal window sizes—is essential for fair benchmarking and the continued advancement of the field. Researchers are encouraged to explicitly report these factors to ensure that claims of algorithmic superiority are robust and reproducible.
Summary updated