The role of face regions in remote photoplethysmography for contactless heart rate monitoring

Figures

• 

  Fig. 1. Flow diagram of rPPG estimation steps. a Face Detection algorithms with Color Channel Enhancement and Photometric Normalization (Illumination Equalization, Color Normalization)—the light-blue area with a dashed border; followed by ROI Selection methods and Pre-Processing (Bandpass Filter, Baseline Correction, Signal Filtering Gaussian/Median) for (b) rPPG Signal detection and filtering—the light-purple area with a dashed border. Next, (c) HR detection—the light-yellow area with a dashed border. Either using ML-based methods(Option 1: the light-red area with a dashed border) with labeled data to estimate HR, or using traditional methods (Option 2: the light-brown area with dashed border) without training data for HR prediction, e.g., via PSD or FFT.

• 

  Fig. 2. Example of ROI Generation algorithms. a Using whole (holistic) face, separate facial patches (regions), and superpixel techniques (segmentation of the image into smaller, meaningful regions based on pixel similarity). Later, (b) single, multiple or set of combinations of sub-ROIs can be used.

• 

  Fig. 3. Search workflow with the identification, screening, eligibility, and inclusion phases in the review according to PRISMA guidelines. Out of 133 initially identified papers, 70 were finally included in the systematic literature review.

• 

  Fig. 4. MAE and RMSE per ROIs with LOESS function on datasets COHFACE, MAHNOB-HCI and PURE. The best results with error < 1.00 are concentrated around 3 and 63 ROIs (also marked with vertical lines). A total of 33 values for MAE and 29 values for RMSE were used.

• 

  Fig. 5. Most commonly analyzed regions of interest (ROIs) in reviewed studies and their pixel intensity characteristics. The top row illustrates the percentage of studies utilizing specific ROIs, including the whole face (36.8%), forehead (24.5%), cheeks (21.7%), nose (6.6%), chin (4.7%), and other regions (5.7%). Gaussian probability density plots show the normalized intensity distributions within each ROI across studies. The bottom row depicts the spatial intensity maps (viridis colormap) of pixels within each ROI, highlighting areas of high and low intensity. Asymmetric ROIs, such as those from the nose and eyes, are shaped by the Mediapipe algorithm, which uses triangular regions that may reduce accuracy. Environmental factors, such as lighting, reflections, and head rotation, also impact ROI selection algorithms.

• 

  Fig. 6. Accumulated averaged number of ROIs per Paper per Year. There is a notable trend of increasing amount of papers starting from the year 2019.