Facial physiological detection method and system based on signal quality driving ROI selection

Abstract

The invention relates to the technical field of image processing and biological signal detection, discloses a facial physiological detection method and system based on signal quality driven ROI selection, and aims to solve the problem of signal degradation of a traditional fixed geometric ROI in a complex scene. The method comprises the following steps: collecting a user face video stream through a camera and preprocessing the user face video stream; detecting a face bounding box and dividing the face bounding box into a plurality of sub-regions; calculating the signal-to-noise ratio, the periodic intensity and the motion artifact interference degree of each sub-region; screening an optimal sub-region according to a weighted fusion formula to generate a dynamic ROI mask; extracting a pure rPPG signal from the dynamic ROI mask coverage area; detrending and band-pass filtering are carried out on the rPPG signals, and physiological parameters such as the heart rate and the blood oxygen saturation degree are extracted. The system comprises a face video acquisition module, a face region positioning and segmentation module, a signal quality evaluation module, a dynamic ROI selection module, an rPPG signal extraction module and a physiological parameter estimation module. According to the technical scheme, signal degradation caused by local shielding, illumination abrupt change or attitude offset can be effectively avoided, the signal-to-noise ratio and the stability of the rPPG signal are remarkably improved, and the universality and the robustness of the method are enhanced.

Description

Technical Field

[0001] This invention belongs to the field of image processing and biological signal detection technology, specifically relating to a facial physiological detection method and system based on signal quality-driven ROI selection. Background Technology

[0002] With the rapid development of non-contact health monitoring technologies, camera-based remote photoplethysmography (rPPG) has become an important tool for physiological signal detection and health screening due to its advantages of being non-invasive, convenient, and allowing for continuous observation. This technology estimates key physiological parameters such as heart rate and blood oxygen saturation by analyzing subtle light intensity fluctuations caused by blood flow changes in facial videos, showing broad application prospects in smart wearables, telemedicine, and public health screening. However, rPPG signals are highly susceptible to interference from factors such as changes in ambient lighting, facial posture shifts, partial occlusion, and individual physiological differences, leading to a significant decrease in the signal-to-noise ratio and severely limiting its reliability and robustness in real-world, complex scenarios.

[0003] Among these factors, the selection strategy for regions of interest (ROIs) is a core element affecting rPPG performance. Traditional methods generally use fixed geometric rules (such as the subnasal triangle or forehead region) as ROIs, assuming that these regions have stable blood flow signals and are less affected by motion interference. However, this assumption often fails in practical applications: lens reflections caused by wearing glasses, local occlusion caused by beards or makeup, and skin texture distortion caused by drastic facial expressions can all drastically degrade the signal quality of the preset ROI regions, and the system cannot detect or avoid such poor-quality regions, thus introducing significant noise or even complete failure.

[0004] Existing technologies lack real-time perception and dynamic response capabilities for local signal quality in ROI selection. They fail to establish pixel-level or region-level signal reliability assessment mechanisms or adaptively adjust sampling areas based on video frame content. This makes the system vulnerable to individual differences, environmental disturbances, or sudden occlusions, making it difficult to maintain stable and high-precision physiological parameter output. These deficiencies are particularly pronounced in uncontrolled scenarios such as home health monitoring and rapid screening in public places, severely hindering the large-scale deployment of rPPG technology from the laboratory. Therefore, there is an urgent need for a facial physiological detection method that can dynamically drive ROI selection based on real-time signal quality to overcome the limitations of fixed ROI paradigms and improve the system's adaptability and accuracy under complex real-world conditions. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a facial physiological detection method and system based on signal quality-driven ROI selection, which can effectively solve the problems in the background technology. To achieve the above objective, this invention provides the following technical solution: In one aspect, a facial physiological detection system based on signal quality-driven ROI selection, the system comprising the following components:

[0006] The facial video capture module is used to continuously capture raw video streams containing the user's face through a camera and perform preprocessing operations on the video streams, including resolution standardization, frame rate unification, and color space conversion.

[0007] A facial region localization and segmentation module, connected to the facial video acquisition module, is used to detect facial bounding boxes from preprocessed video frames and divide the facial region into multiple independent sub-regions based on the facial key point localization results. The sub-regions include the forehead, left cheek, right cheek, nose, and chin regions.

[0008] The signal quality assessment module is connected to the facial region localization and segmentation module. It is used to perform time-domain and frequency-domain analysis on the pixel sequence in each sub-region and calculate its corresponding signal quality index. The signal quality index includes signal-to-noise ratio, periodicity intensity, and motion artifact interference level.

[0009] The dynamic ROI selection module is connected to the signal quality evaluation module. It is used to sort and filter each sub-region according to the real-time calculated signal quality index, select sub-regions with quality index higher than a preset threshold to form the final ROI of the current frame, and generate the corresponding mask matrix.

[0010] The rPPG signal extraction module is connected to the dynamic ROI selection module and is used to extract the original light intensity signal from the pixel area covered by the selected ROI mask, and to independently separate and normalize the multi-channel color signal.

[0011] The physiological parameter estimation module is connected to the rPPG signal extraction module. It is used to perform detrending and bandpass filtering on the normalized light intensity signal, and then extract physiological parameters such as heart rate and blood oxygen saturation through spectrum analysis or time-frequency transformation.

[0012] Preferably, in the signal quality assessment module, the signal-to-noise ratio (SNR) is calculated based on the ratio of the power spectral density of the pixel sequence within each sub-region to the background noise power in the rPPG frequency band. The power spectral density is calculated using a Hanning window plus a windowed Fast Fourier Transform, with a window length of 512 points and an overlap rate of 50%. The baseline fluctuation amplitude is defined as the standard deviation of the autocorrelation function in the 50-100 sampling point delay interval, and its specific formula is as follows:

[0013]

[0014] Among them, P signal P represents the signal power within the 0.8Hz to 3Hz frequency band. noise The noise power is the average value in the frequency bands of 0.1Hz to 0.7Hz and 3.5Hz to 5Hz.

[0015] Furthermore, the periodic intensity is obtained by calculating the ratio of the peak amplitude of the autocorrelation function of the pixel sequence at the heart rate-corresponding delay to the baseline fluctuation amplitude, with a threshold set to 1.5; the degree of motion artifact interference is estimated by optical flow method to estimate the average pixel displacement in the sub-region and compared with a preset displacement tolerance of 2 pixels.

[0016] Furthermore, the dynamic ROI selection module employs a weighted fusion strategy, with weight allocation determined based on regression analysis of 100 sets of laboratory samples. A comprehensive quality score threshold of 0.7 corresponds to an optimization point with an area under the ROC curve ≥ 0.85. The three indicators—signal-to-noise ratio, periodicity intensity, and motion artifact interference—are assigned weights of 0.5, 0.3, and 0.2, respectively, to calculate the comprehensive quality score of each sub-region. Only when the comprehensive quality score is higher than 0.7 and the signal-to-noise ratio is greater than 8dB is the sub-region included in the final ROI combination.

[0017] Preferably, after separating the RGB three-channel signals, the rPPG signal extraction module uses a blind source separation algorithm to perform independent component analysis on the signals, selects the component with the highest correlation to the blood flow pulse waveform as the effective rPPG signal, and normalizes the amplitude of the component to the [-1,1] interval.

[0018] On the other hand, a facial physiological detection method based on signal quality-driven ROI selection includes the following specific steps:

[0019] Step S110: Capture the user's facial video stream using a camera, and perform resolution normalization and color space conversion on the video stream;

[0020] Step S120: Detect the face bounding box in each video frame, locate the facial key points, and divide the facial region into 5 sub-regions: forehead, left cheek, right cheek, nose, and chin.

[0021] Step S130: For each sub-region, extract its pixel sequence and calculate three signal quality indicators: signal-to-noise ratio, periodicity intensity, and motion artifact interference level.

[0022] Step S140: Calculate the comprehensive quality score of each sub-region according to the weighted fusion formula, filter out sub-regions with a score higher than 0.7 and a signal-to-noise ratio greater than 8dB, and generate a dynamic ROI mask;

[0023] Step S150: Extract the RGB channel light intensity signal from the area covered by the dynamic ROI mask, and obtain the pure rPPG signal after blind source separation and normalization.

[0024] Step S160: The rPPG signal is detrended and bandpass filtered from 0.8Hz to 3Hz. Then, the frequency corresponding to the spectral peak is extracted using a Fast Fourier Transform and converted into a heart rate value. Based on the AC / DC ratio of the red (R) and green (G) channels, the calibration curve is calibrated using an in vitro blood oxygenation simulation device. The fitting formula is as follows: Calculate blood oxygen saturation.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] By evaluating the signal quality of each facial sub-region in real time and dynamically selecting the optimal ROI, signal degradation caused by local occlusion, sudden changes in illumination or pose shift is effectively avoided, significantly improving the signal-to-noise ratio and stability of rPPG signals.

[0027] The ROI screening mechanism, which adopts a multi-index weighted fusion, takes into account both the periodicity of the signal and the anti-interference capability of motion, enabling the system to maintain high-precision physiological parameter output even in complex environments.

[0028] By using blind source separation and normalization, the influence of ambient light and individual skin color differences on rPPG signals was eliminated, enhancing the universality and robustness of the method.

[0029] The overall solution does not rely on a preset fixed ROI, achieving fully adaptive facial physiological detection and greatly expanding the practical value of rPPG technology in uncontrolled scenarios such as home monitoring and public screening. Attached Figure Description

[0030] Figure 1 This is a schematic diagram of the overall technical architecture of the facial physiological detection method based on signal quality-driven ROI selection proposed in this invention;

[0031] Figure 2 This is a schematic diagram of the core principle framework of the dynamic ROI selection mechanism in this invention;

[0032] Figure 3 This is a logical flowchart of facial region localization and segmentation in this invention;

[0033] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between signal quality assessment and rPPG signal extraction in this invention;

[0034] Figure 5 This is a flowchart illustrating the logical process of physiological parameter estimation in this invention. Detailed Implementation

[0035] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the specific embodiments according to the present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments.

[0036] Example 1

[0037] In a home health monitoring scenario, users capture facial video using their smartphone's front-facing camera. Once activated, the facial video capture module continuously captures a raw video stream containing the user's face. This module first performs resolution standardization, adjusting the input video to a uniform 1920×1080 pixel resolution to ensure subsequent processing units receive video frames of consistent size. Next, it performs frame rate unification, stabilizing the video stream at a capture rate of 30 frames per second using frame sampling technology to eliminate frame rate fluctuations caused by differences in device performance. Finally, it performs color space conversion, converting the original BGR color space to the standard RGB color space, and uses gamma correction technology to normalize image brightness, eliminating the impact of ambient lighting differences on color reproduction. Throughout the preprocessing process, the module monitors the data integrity of the video stream in real time, initiating a retransmission mechanism for frame loss due to packet loss to ensure the spatiotemporal continuity of the video sequence.

[0038] See Figure 1 The facial region localization and segmentation module extracts facial region information from preprocessed video frames. This module employs a face detection algorithm based on a multi-task convolutional neural network to generate bounding box coordinates containing the facial region in each frame. The confidence threshold for the bounding box is set to 0.95; only when the detection confidence exceeds this threshold is the face considered validly localized. Subsequently, the module activates the facial landmark localization engine to accurately identify the pixel coordinates of 68 facial feature points, including key anatomical landmarks such as the brow bone contour, eyelid edge, nasal alar boundary, and lip line contour. Based on the spatial distribution of these landmarks, the module divides the facial region into five independent anatomical sub-regions: the forehead region is defined as a rectangular area from the upper edge of the brow to the hairline, and from both sides to the temples; the left cheek region is defined as a triangular area from the outer canthus of the left eye to the front of the left ear, and from the left side of the nasal alar to the angle of the mandible; the right cheek region is divided symmetrically; the nose region is defined as a trapezoidal area from the root of the nose to the tip of the nose, and from both sides to the nasal alar boundary; and the chin region is defined as a rectangular area from the lower lip to the edge of the mandible. Each sub-region generates a corresponding binary mask matrix for subsequent pixel sequence extraction.

[0039] The signal quality assessment module performs comprehensive time-domain and frequency-domain analysis on the pixel sequences within each sub-region. First, the module extracts the mean pixel brightness sequence within the mask coverage area of ​​each sub-region, with a sampling frequency of 30 Hz and a sequence length of 512 sampling points. In the time-domain analysis, the module calculates the mean, variance, and autocorrelation function of the pixel sequence, with the calculation delay range for the autocorrelation function set to 0 to 255 sampling points. In the frequency-domain analysis, the module uses a Fast Fourier Transform to convert the time-domain signal into a frequency-domain representation, achieving a spectral resolution of 0.0586 Hz. The signal-to-noise ratio (SNR) is calculated based on the power spectral density characteristics of the pixel sequence within each sub-region in the rPPG frequency band. The specific execution process is as follows: First, the signal power value in the 0.8 Hz to 3 Hz frequency band is calculated, which corresponds to the normal adult heart rate range of 48 to 180 beats per minute. Then, the average noise power in the 0.1 Hz to 0.7 Hz and 3.5 Hz to 5 Hz frequency bands is calculated, which correspond to respiratory interference and high-frequency noise, respectively. Finally, the SNR value is calculated according to the formula, where the ratio of signal power to noise power is taken as the common logarithm and multiplied by 10, with the unit being standard decibels.

[0040] The periodicity intensity index is obtained by calculating the characteristic value of the autocorrelation function of the pixel sequence at the corresponding delay in heart rate. This module first identifies the main peak frequency from the power spectral density, calculates the expected heart rate cycle based on the frequency value, and then finds the peak amplitude at the corresponding delay position in the autocorrelation function. Simultaneously, it calculates the baseline fluctuation amplitude of the autocorrelation function in the non-periodic delay interval, and takes the ratio of the two as the quantified value of periodicity intensity. This module sets a qualified threshold of 1.5 for periodicity intensity; when the ratio is lower than this threshold, the signal periodicity characteristic of the sub-region is deemed insufficient. The motion artifact interference level index is estimated using the optical flow method. The module uses the Lucas-Kanade optical flow algorithm to calculate the pixel displacement vector in each sub-region between consecutive video frames, and calculates the average displacement of all pixels as the quantified index of motion artifact interference level. This index is compared with a preset 2-pixel displacement tolerance; when the average displacement exceeds this tolerance, the sub-region is deemed severely affected by motion interference.

[0041] The dynamic ROI selection module employs a multi-index weighted fusion strategy for sub-region filtering. First, the module normalizes three quality indices, mapping the original values ​​to a range of 0 to 1. The signal-to-noise ratio (SNR) index is normalized using a piecewise linear mapping: 8 dB corresponds to 0.7, 20 dB to 1.0, and values ​​below 8 dB are directly mapped to 0. The periodicity intensity index uses a linear mapping: 1.5 corresponds to 0.7, and 3.0 to 1.0. The motion artifact interference level index uses an inverse proportional mapping: 0 pixel displacement corresponds to 1.0, 2 pixel displacement to 0.7, and displacements exceeding 2 pixels are directly mapped to 0. After normalization, the module calculates the comprehensive quality score for each sub-region according to a weighted allocation scheme, where the SNR index has a weight of 0.5, the periodicity intensity index has a weight of 0.3, and the motion artifact interference level index has a weight of 0.2. The comprehensive quality score is calculated as the sum of the products of the normalized values ​​of each index and their corresponding weights. The selection criteria are set as an overall quality score higher than 0.7 and a signal-to-noise ratio greater than 8 dB. Only sub-regions that simultaneously meet both conditions are included in the final ROI combination. This module generates a dynamic ROI mask matrix in real time, where a mask value of 1 indicates that the corresponding pixel is selected, and a mask value of 0 indicates that the corresponding pixel is excluded.

[0042] The rPPG signal extraction module extracts the raw light intensity signal from the pixel area covered by the dynamic ROI mask. This module first separates the RGB three-channel color signals, calculating the mean pixel intensity of each channel within the selected area, forming three independent time-series signals. Each signal has a sampling length of 512 points, and the sampling frequency remains constant at 30 Hz. Next, the module uses a blind source separation algorithm to perform independent component analysis on the three-channel signals. Specifically, the data matrix composed of the three-channel signals is first centered by subtracting the mean of each channel; then, the covariance matrix of the data is calculated, and a whitening matrix is ​​obtained through eigenvalue decomposition; finally, the FastICA algorithm is used to achieve blind source separation, with a convergence tolerance of 1e-6 and a maximum of 100 iterations, resulting in three statistically independent source signals. The module calculates the correlation coefficient between each source signal and the ideal blood flow pulse waveform, selecting the component with the highest correlation coefficient as the effective rPPG signal. The signal is then normalized in amplitude, and a linear transformation maps the signal value to a closed interval from -1 to +1. The transformation formula is dynamically adjusted based on the minimum and maximum values ​​of the signal.

[0043] The physiological parameter estimation module further processes the normalized rPPG signal to extract physiological parameters. This module first performs detrending, using a locally weighted regression algorithm to eliminate slowly varying trend components in the signal, with a sliding window width set to 5 seconds. Next, bandpass filtering is performed using a 4th-order Butterworth filter with a passband frequency range of 0.8 Hz to 3 Hz and a stopband attenuation of 40 dB. The filtered signal is then converted to a frequency domain representation using a Fast Fourier Transform. Spectral analysis employs a Hanning window function to reduce spectral leakage, achieving a frequency resolution of 0.0586 Hz. The module identifies the peak with the largest amplitude in the spectrum, takes its corresponding frequency as the fundamental frequency of the heart rate signal, and converts the frequency value to 60 to calculate the heart rate per minute. The calculation of blood oxygen saturation is based on the characteristic of the amplitude ratio of the red light channel and the infrared light channel. The module first extracts the DC and AC components of the red and green light channels from the RGB signal before blind source separation, calculates the AC / DC ratio of each channel, and then maps the ratio to a blood oxygen saturation percentage value through a pre-calibrated curve. Throughout the processing, the module monitors the signal quality in real time, and automatically triggers a re-acquisition mechanism when the coefficient of variation of the heart rate value exceeds 15% for three consecutive calculation cycles.

[0044] See Figure 2 The core principle framework of the dynamic ROI selection mechanism demonstrates the complete process of quality assessment and region selection. Signal quality indicators for each sub-region are processed simultaneously through parallel computing channels. The weighted fusion unit comprehensively evaluates the applicability of each region based on preset weight coefficients. The decision logic unit strictly executes dual threshold selection conditions, ultimately outputting an optimized ROI combination scheme. This design ensures that even when parts of the face are occluded or there is uneven lighting, the system can still automatically select the facial region with the best signal quality for physiological parameter detection.

[0045] See Figure 3 The logical flow of facial region localization and segmentation details the sequential operations from face detection to sub-region division. After the face detector generates bounding boxes, the key point localization engine accurately identifies the location of facial features. Based on prior anatomical knowledge, the region segmentation algorithm divides the face into five functional sub-regions, each corresponding to a specific vascular distribution pattern, providing a structural basis for subsequent signal quality assessment.

[0046] See Figure 4 The multi-level interaction between signal quality assessment and rPPG signal extraction demonstrates the dependencies and information transmission paths in the data processing flow. The indicator data generated by the quality assessment layer directly guides the decision-making process of the ROI selection layer, the selected region mask controls the sampling range of the signal extraction layer, and the extracted raw signal is fed back to the quality assessment layer for closed-loop verification, forming a complete quality-driven adaptive loop.

[0047] See Figure 5 The logical framework for physiological parameter estimation describes the algorithmic chain from preprocessing rPPG signals to the final physiological parameter output. The detrending unit eliminates baseline drift, the bandpass filtering unit preserves the effective frequency band, the spectrum analysis unit extracts periodic features, and the parameter conversion unit maps signal features into clinically readable physiological indicators. The close integration of each step ensures the accuracy and reliability of the measurement results.

[0048] Example 2

[0049] In public health screening scenarios, the system is deployed on self-service clearance equipment at airport immigration checkpoints to conduct non-contact initial screening of physiological parameters for large groups of people. The facial video acquisition module uses a high-definition infrared camera to simultaneously acquire facial video streams in both visible and infrared light modes. Resolution standardization processing uniformly adjusts the video to 1280×720 pixels and sets the frame rate to 25 frames per second to adapt to the limitations of device processing capabilities and network transmission bandwidth. The color space conversion process has been particularly enhanced to adapt to different skin tones and races by establishing a skin tone adaptive mapping table to eliminate the impact of racial differences on color reproduction.

[0050] The facial region localization and segmentation module quickly and accurately locates facial regions in complex backgrounds, employing a lightweight face detection network to ensure a processing speed of over 15 frames per second. Considering that subjects may wear glasses, masks, or other items, the module enhances its robustness against occlusion. When partial occlusion is detected, the keypoint localization strategy is automatically adjusted, dividing the sub-regions only based on the visible area. The signal quality assessment module, when calculating the signal-to-noise ratio (SNR), subdivides the heart rate frequency band into three sub-bands: 0.8 Hz to 1.5 Hz corresponds to bradycardia, 1.5 Hz to 2.5 Hz corresponds to normal heart rate, and 2.5 Hz to 3 Hz corresponds to tachycardia. The SNR of each sub-band is calculated separately, and the weighted average is taken as the final metric.

[0051] The dynamic ROI selection module incorporates an environment-adaptive mechanism into its weighted fusion strategy, dynamically adjusting the weight allocation of each quality indicator based on ambient light intensity. In strong light conditions, the weight of motion artifact interference is increased from 0.2 to 0.3; in low light conditions, the weight of the signal-to-noise ratio (SNR) indicator is increased from 0.5 to 0.6. Simultaneously, the screening threshold is fine-tuned based on population statistical characteristics, with the overall quality score threshold adjusted from 0.7 to 0.65 to meet the efficiency requirements of large-scale screening scenarios. The rPPG signal extraction module employs an incremental blind source separation algorithm, updating the separation matrix only for newly added data segments when processing continuous video streams, significantly improving computational efficiency and ensuring the system meets the real-time requirements of high-throughput screening.

[0052] The physiological parameter estimation module incorporates an anomaly detection mechanism. When the calculated heart rate value is below 40 beats per minute or above 180 beats per minute, it is automatically marked as a suspicious result and a review process is initiated. A temperature compensation factor is introduced into the blood oxygen saturation calculation to eliminate the influence of ambient temperature on the accuracy of optical measurements. The system compares the final measurement results with preset normal ranges, automatically generating health warning reports for cases exceeding the thresholds, providing data support for public health management. The entire processing flow ensures measurement accuracy while keeping the single detection time within 15 seconds, meeting the practical needs of high-throughput screening in public places.

[0053] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A facial physiological detection system based on signal quality-driven ROI selection, characterized in that, The system includes the following components: The facial video capture module is used to continuously capture raw video streams containing the user's face through a camera and perform preprocessing operations on the video streams, including resolution standardization, frame rate unification, and color space conversion. A facial region localization and segmentation module, connected to the facial video acquisition module, is used to detect facial bounding boxes from preprocessed video frames and divide the facial region into multiple independent sub-regions based on the facial key point localization results. The sub-regions include the forehead, left cheek, right cheek, nose, and chin regions. The signal quality assessment module is connected to the facial region localization and segmentation module. It is used to perform time-domain and frequency-domain analysis on the pixel sequence in each sub-region and calculate its corresponding signal quality index. The signal quality index includes signal-to-noise ratio, periodicity intensity, and motion artifact interference level. The dynamic ROI selection module is connected to the signal quality evaluation module. It is used to sort and filter each sub-region according to the real-time calculated signal quality index, select sub-regions with quality index higher than a preset threshold to form the final ROI of the current frame, and generate the corresponding mask matrix. The rPPG signal extraction module is connected to the dynamic ROI selection module and is used to extract the original light intensity signal from the pixel area covered by the selected ROI mask, and to independently separate and normalize the multi-channel color signal. The physiological parameter estimation module is connected to the rPPG signal extraction module. It is used to perform detrending and bandpass filtering on the normalized light intensity signal, and then extract physiological parameters such as heart rate and blood oxygen saturation through spectrum analysis or time-frequency transformation.

2. The facial physiological detection system based on signal quality-driven ROI selection according to claim 1, characterized in that, In the signal quality assessment module, the signal-to-noise ratio (SNR) is calculated based on the ratio of the power spectral density of the pixel sequence within each sub-region to the background noise power in the rPPG frequency band. The power spectral density is calculated using a Hanning window plus a windowed Fast Fourier Transform, with a window length of 512 points and an overlap rate of 50%. The baseline fluctuation amplitude is defined as the standard deviation of the autocorrelation function in the 50-100 sampling point delay interval, and its specific formula is as follows: Among them, P signal P represents the signal power within the 0.8Hz to 3Hz frequency band. noise The noise power is the average value in the frequency bands of 0.1Hz to 0.7Hz and 3.5Hz to 5Hz.

3. The facial physiological detection system based on signal quality-driven ROI selection according to claim 1, characterized in that, In the signal quality assessment module, the periodic intensity is obtained by calculating the ratio of the peak amplitude of the autocorrelation function of the pixel sequence at the heart rate corresponding delay to the baseline fluctuation amplitude, and its threshold is set to 1.5; the degree of motion artifact interference is estimated by optical flow method to estimate the average pixel displacement in the sub-region and compared with the preset displacement tolerance of 2 pixels.

4. The facial physiological detection system based on signal quality-driven ROI selection according to claim 1, characterized in that, The dynamic ROI selection module adopts a weighted fusion strategy, with weight allocation determined based on regression analysis of 100 sets of laboratory samples. The comprehensive quality score threshold of 0.7 corresponds to the optimization point with an area under the ROC curve ≥ 0.

85. The three indicators, signal-to-noise ratio, periodicity intensity, and motion artifact interference, are assigned weights of 0.5, 0.3, and 0.2, respectively, and the comprehensive quality score of each sub-region is calculated. Only when the comprehensive quality score is higher than 0.7 and the signal-to-noise ratio is greater than 8dB, the sub-region is included in the final ROI combination.

5. The facial physiological detection system based on signal quality-driven ROI selection according to claim 1, characterized in that, After separating the RGB three-channel signals, the rPPG signal extraction module uses a blind source separation algorithm to perform independent component analysis on the signals, selects the component with the highest correlation to the blood flow pulse waveform as the effective rPPG signal, and normalizes the amplitude of the component to the [-1,1] interval.

6. The facial physiological detection system based on signal quality-driven ROI selection according to claim 1, characterized in that, The physiological parameter estimation module detrends and applies a 0.8Hz to 3Hz bandpass filter to the normalized rPPG signal, then extracts the frequency corresponding to the spectral peak through Fast Fourier Transform, converts it into a heart rate value, and calibrates the curve using an in vitro blood oxygenation simulation device based on the AC / DC ratio of the red (R) and green (G) channels. The fitting formula is as follows: Calculate blood oxygen saturation.

7. A facial physiological detection method based on signal quality-driven ROI selection, characterized in that, The method includes the following steps: The S110 captures a video stream of the user's face through a camera and performs resolution standardization and color space conversion on the video stream. S120 detects face bounding boxes in each video frame, locates facial key points, and divides the facial region into 5 sub-regions: forehead, left cheek, right cheek, nose, and chin. S130: For each sub-region, extract its pixel sequence and calculate three signal quality indicators: signal-to-noise ratio, periodicity intensity, and motion artifact interference level. S140: Calculate the overall quality score of each sub-region according to the weighted fusion formula, select sub-regions with a score higher than 0.7 and a signal-to-noise ratio greater than 8dB, and generate a dynamic ROI mask; S150 extracts RGB channel light intensity signals from the area covered by dynamic ROI mask, and obtains pure rPPG signals after blind source separation and normalization. S160 performs detrending and bandpass filtering (0.8Hz to 3Hz) on the rPPG signal, then extracts the frequency corresponding to the spectral peak through Fast Fourier Transform, converts it into a heart rate value, and calibrates the curve using an in vitro blood oxygenation simulator based on the AC / DC ratio of the red (R) and green (G) channels. The fitting formula is as follows: Calculate blood oxygen saturation.

8. The facial physiological detection method based on signal quality-driven ROI selection according to claim 7, characterized in that, In step S130, the signal-to-noise ratio is calculated based on the ratio of the power spectral density of the pixel sequence within each sub-region to the background noise power in the rPPG frequency band. The power spectral density is calculated using a Hanning window plus a windowed Fast Fourier Transform with a window length of 512 points and an overlap rate of 50%. The baseline fluctuation amplitude is defined as the standard deviation of the autocorrelation function in the delay interval of 50-100 sampling points, and its specific formula is as follows: Among them, P signal P represents the signal power within the 0.8Hz to 3Hz frequency band. noise The noise power is the average value in the frequency bands of 0.1Hz to 0.7Hz and 3.5Hz to 5Hz.

9. The facial physiological detection method based on signal quality-driven ROI selection according to claim 7, characterized in that, In step S130, the periodic intensity is obtained by calculating the ratio of the peak amplitude of the autocorrelation function of the pixel sequence at the heart rate corresponding delay to the baseline fluctuation amplitude, and its threshold is set to 1.5; the degree of motion artifact interference is estimated by optical flow method to estimate the average pixel displacement in the sub-region and compared with the preset displacement tolerance of 2 pixels.

10. The facial physiological detection method based on signal quality-driven ROI selection according to claim 7, characterized in that, In step S140, a weighted fusion strategy is adopted, assigning weights of 0.5, 0.3, and 0.2 to the three indicators of signal-to-noise ratio, periodicity intensity, and motion artifact interference, respectively, and calculating the comprehensive quality score of each sub-region; only when the comprehensive quality score is higher than 0.7 and the signal-to-noise ratio is greater than 8dB, the sub-region is included in the final ROI combination.

Images