A non-contact physiological state rapid screening method, system and medium

Abstract

The present application relates to the technical field of work personnel safety management, and discloses a non-contact physiological state rapid screening method, system and medium, which comprises collecting an RGB three-channel space-time image sequence of a face region of interest; an adaptive gamma correction algorithm is used to correct the illumination of the RGB three-channel space-time image sequence, a one-dimensional convolutional neural network self-encoder denoising algorithm is used to perform end-to-end denoising processing on the illumination-corrected image sequence, and a remote photoplethysmography pulsatile signal is separated out; physiological indexes of heart rate, heart rate variability and blood oxygen saturation are calculated based on the remote photoplethysmography pulsatile signal; the physiological indexes are graded and determined based on a grading standard of the physiological indexes, a final job suitability determination grade is determined in combination with a bucket effect model, and a grading early warning signal is generated. The present application can realize accurate and rapid evaluation of the physiological state of work personnel in a complex environment, comprehensively and objectively assess the health degree of an individual, and provide decision support for the reasonable allocation of posts and the efficient configuration of personnel.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of work personnel safety management, and particularly relates to a non-contact physiological state rapid screening method, a system and a medium. BACKGROUND

[0002] In high-risk industries such as electric power operation, mining and chemical industry, the health status of work personnel is directly related to production safety and life security. Traditional health detection methods mainly rely on wearable devices or manual physical examination. Wearable devices (such as heart rate bands, blood oxygen meters and smart bracelets) need to be in direct contact with the human body through electrode patches or sensors, and the operation process is complex and easy to cause discomfort. For example, electrocardiogram electrodes need to be attached to the skin for a long time, which may cause allergic reactions (such as contact dermatitis) or skin damage, especially in high-temperature and high-humidity environments, the chemical reaction of sweat and electrode material will further aggravate skin irritation. In addition, there is a risk of cross-infection when multiple people share the device, such as blood residue or microbial contamination. Studies have shown that the average installation time of wearable devices is about 5-10 minutes per person, which significantly reduces the efficiency of operation preparation, and the cost of device cleaning and maintenance is high. Traditional manual physical examination needs to complete detection of heart rate, blood oxygen saturation and the like item by item, and each check takes an average of 2-3 minutes, and a single person takes about 15-30 minutes to complete the physical examination process. In large-scale operation scenes such as electric power repair and mine blasting, such methods are difficult to meet the rapid screening demand. For example, a substation maintenance project needs to detect 50 work personnel at the same time, and the traditional method needs to take 2.5-7.5 hours, while the operation window period usually only allows 1-2 hours. In addition, manual operation depends on the professional level of the physical examination personnel, and the process of device carrying and calibration further slows down the progress, resulting in operation delay and even safety hazards. In summary, wearable devices have the problems of contact limitation, complicated operation, easy to cause skin allergy, cross-infection risk and low efficiency, and manual physical examination has the problems of strong subjectivity and low efficiency, which are difficult to meet the demand of rapid pre-job screening of large-scale operation team.

[0003] In recent years, non-contact detection technology based on remote photoplethysmography has been developed, which extracts physiological parameters such as heart rate by analyzing the weak skin color changes caused by blood volume changes in facial video. However, the traditional remote photoplethysmography algorithm has the following problems in actual industrial scene application:

[0004] 1) Illumination interference problem: visible light cameras are easy to produce image distortion, noise or overexposure under strong light outdoors, indoor shadows or different light sources, which leads to increased error in remote photoplethysmography pulsatile signal extraction;

[0005] 2) Insufficient dynamic environment adaptability: traditional remote photoplethysmography algorithm is sensitive to motion artifacts and illumination changes, and is difficult to operate stably in complex work scenarios, which cannot meet the demand of rapid and accurate pre-service screening of power work. SUMMARY

[0006] In order to overcome the above-mentioned shortcomings of the prior art, the present application provides a non-contact physiological state rapid screening method, system and medium, aiming at solving the problems of illumination interference and insufficient dynamic environment adaptability of the existing non-contact detection technology, and realizing rapid, accurate and large-scale pre-service physiological state screening under complex illumination.

[0007] The technical scheme adopted by the present application to solve its technical problems is: a non-contact physiological state rapid screening method, comprising:

[0008] S1: acquiring a face video of a subject, and collecting an RGB three-channel spatiotemporal image sequence of a face region of interest from the face video;

[0009] S2: performing frame-by-frame illumination correction on the RGB three-channel spatiotemporal image sequence by using an adaptive gamma correction algorithm to obtain an illumination correction image sequence; the adaptive gamma correction dynamically calculates an optimal gamma value by K-means clustering analysis of image pixel intensity distribution;

[0010] S3: performing end-to-end noise reduction processing on the illumination correction image sequence by using a one-dimensional convolutional neural network autoencoder denoising algorithm to separate remote photoplethysmography pulsatile signals from mirror reflection and motion noise;

[0011] S4: calculating physiological indexes of heart rate, heart rate variability and blood oxygen saturation based on the separated remote photoplethysmography pulsatile signals.

[0012] As a further improvement of the present application: further comprising the following step: S5: grading the calculated physiological indexes based on the physiological index grading standard, and determining the final job suitability grade in combination with the bucket effect model, and generating a grading warning signal.

[0013] As a further improvement of the present application: the calculation formula of the adaptive gamma correction algorithm is:

[0014] ;

[0015] wherein, is the intensity value of the illumination correction image sequence; is the intensity value of the overexposed or underexposed RGB three-channel spatiotemporal image sequence; is the optimal gamma value dynamically calculated by K-means clustering analysis of image pixel intensity distribution. ​

[0016] As a further improvement of the present application: the optimal gamma value is dynamically calculated by K-means clustering analysis of image pixel intensity distribution, specifically comprising:

[0017] Normalizing the image pixel intensity to the range of [0, 1];

[0018] Applying K-means clustering algorithm to divide the pixels into K clusters, and calculating the mean value of each cluster;

[0019] According to the distribution characteristics of the cluster, the optimal gamma value is dynamically adjusted , wherein the cluster with a mean value greater than 0.9 is an overexposure cluster, and the cluster with a mean value less than 0.1 is an underexposure cluster. The optimal value is dynamically calculated for the overexposure cluster and the underexposure cluster.

[0020] As a further improvement of the present application: the calculation formula of heart rate is:

[0021] HR=60 / MeanRR;

[0022] ;

[0023] Wherein, HR is the heart rate, MeanRR is the average value of R-R interval, Number of R-R intervals represents the total number of heartbeats.

[0024] As a further improvement of the present application: the calculation formula of blood oxygen saturation is:

[0025] ;

[0026] ;

[0027] ;

[0028] Wherein, is the blood oxygen saturation, ΔA R is the signal AC / DC ratio of R channel, ΔA G is the signal AC / DC ratio of G channel, AC R is the AC component of R channel, DC R is the DC component of R channel, AC G is the AC component of G channel, DC G is the DC component of G channel, a and b are experimental calibration coefficients.

[0029] As a further improvement of the present invention: the barrel effect model selects the worst-performing physiological indicator among all physiological indicators as the final job suitability assessment level; the graded early warning signals include green (normal), yellow (warning), and red (prohibited from working).

[0030] This invention also provides a non-contact rapid physiological state screening system for implementing the aforementioned non-contact rapid physiological state screening method, comprising:

[0031] The perception layer includes a data acquisition module, which is used to acquire RGB three-channel spatiotemporal image sequences of facial regions of interest.

[0032] The processing layer includes an edge computing module, which runs an adaptive gamma correction algorithm and a one-dimensional convolutional neural network autoencoder denoising algorithm to calculate physiological indicators.

[0033] The decision-making level is used to run physiological indicator grading standards and the barrel effect model to generate graded early warning signals.

[0034] As a further improvement of the present invention: the sensing layer also includes a 4K high-definition industrial-grade visible light camera, which integrates an infrared fill light unit and an automatic white balance adjustment unit.

[0035] As a further improvement of the present invention: the edge computing module includes an illumination correction unit, a one-dimensional convolutional neural network noise reduction unit, and a physiological index calculation unit;

[0036] The illumination correction unit dynamically calculates the optimal γ value through K-means clustering and uses an adaptive gamma correction algorithm to perform frame-by-frame illumination correction on the RGB three-channel spatiotemporal image sequence to obtain an illumination-corrected image sequence.

[0037] The one-dimensional convolutional neural network noise reduction unit is used to perform end-to-end noise reduction processing on the illumination-corrected image sequence using a one-dimensional convolutional neural network autoencoder noise reduction algorithm, separating the remote photoplethysmography pulsation signal from specular reflection and motion noise.

[0038] The physiological index calculation unit is used to calculate heart rate, heart rate variability and blood oxygen saturation based on the separated remote photoplethysmography pulse signal.

[0039] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned non-contact rapid screening method for physiological states.

[0040] Compared with the prior art, the beneficial effects of the present invention are:

[0041] 1. This invention can reduce the detection error of physiological signals in environments with strong light (direct sunlight), shadow (building obstruction) or mixed light sources (LED + incandescent lamp), and improve the robustness of light illumination; it can complete the screening of a single person within 10 seconds, support the rapid detection of operators, significantly shorten the pre-job preparation time, and has real-time and high efficiency.

[0042] 2. This invention enables accurate and rapid assessment of the physiological state of workers in complex environments, comprehensively and objectively considering individual health levels, providing solid and reliable decision support for the rational allocation of positions and efficient personnel deployment, while ensuring individual health and high-quality work progress.

[0043] 3. This invention optimizes the global brightness distribution of the acquired facial video frame by frame using an adaptive gamma correction algorithm, and dynamically calculates the optimal value based on the pixel intensity distribution. The values ​​effectively correct overexposed / underexposed areas and optimize overall brightness uniformity. Subsequently, a one-dimensional convolutional neural network autoencoder denoising algorithm is used to perform end-to-end denoising on the corrected illumination-corrected image sequence, achieving separation of remote photoplethysmography pulsation signals from specular reflection and motion noise.

[0044] 4. This invention utilizes a 4K high-definition industrial-grade visible light camera integrating an infrared fill light unit and an automatic white balance adjustment unit to ensure image quality in multi-light source scenarios; it dynamically eliminates interference from strong light, shadows, and mixed light sources by employing an adaptive gamma correction algorithm and a one-dimensional convolutional neural network autoencoder noise reduction algorithm; through an edge computing module, it achieves single-person screening within 10 seconds, supports multi-person parallel processing and data caching, and adapts to the needs of high-concurrency operation scenarios; it grades the calculated physiological indicators using a physiological indicator grading standard, and quantifies individual health shortcomings using the barrel effect model to determine the final job suitability assessment level, ensuring the scientific and objective nature of the assessment results; and it enables real-time risk intervention and job suitability decision-making through early warning responses via LED screen colors (green / yellow / red) and audible and visual alarms. Attached Figure Description

[0045] Figure 1 This is a flowchart of a non-contact rapid screening method for physiological states.

[0046] Figure 2 This is a structural block diagram of a non-contact rapid physiological state screening system.

[0047] Figure 3 This is a block diagram of the edge computing module. Detailed Implementation

[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or devices. The present invention will now be further described in conjunction with the accompanying drawings and embodiments:

[0049] Please see Figure 1 An embodiment of the present invention provides a non-contact rapid screening method for physiological states, comprising:

[0050] S1: Acquire RGB three-channel spatiotemporal image sequence data: Acquire facial video of the examinee and extract RGB three-channel spatiotemporal image sequence of the region of interest of the face from the facial video;

[0051] S2: Illumination correction of the RGB three-channel spatiotemporal image sequence: An adaptive gamma correction algorithm is used to perform frame-by-frame illumination correction on the RGB three-channel spatiotemporal image sequence to eliminate overexposed and underexposed areas, resulting in an illumination-corrected image sequence; the adaptive gamma correction uses K-means clustering analysis to analyze the pixel intensity distribution of the image to dynamically calculate the optimal... value;

[0052] S3: Denoising the illumination-corrected image sequence: End-to-end denoising of the illumination-corrected image sequence is performed using a one-dimensional convolutional neural network autoencoder denoising algorithm to separate the remote photoplethysmography pulsation signal from specular reflection and motion noise.

[0053] S4: Physiological index calculation: Based on the separated remote photoplethysmography pulse signal, calculate the physiological indexes of heart rate (HR), heart rate variability (HRV), and blood oxygen saturation (SpO2).

[0054] S5: Generation of graded early warning signals: Construct a grading standard for physiological indicators, grade the calculated physiological indicators based on the grading standard, determine the final suitability level by combining the barrel effect model, and generate graded early warning signals.

[0055] Data is acquired by using RGB three-channel spatiotemporal image sequences of the facial region of interest. No wearable devices are required, which avoids the risks of skin allergies and cross-infection caused by contact, while also saving on equipment cleaning and maintenance costs and improving the convenience of screening.

[0056] Secondly, adaptive gamma correction uses K-means clustering to dynamically analyze pixel intensity distribution, accurately calculate the optimal γ value and process it frame by frame. It can specifically eliminate overexposed and underexposed areas, effectively solve the problem of image distortion under complex lighting conditions, reduce the detection error of physiological state signals in strong light (direct sunlight), shadow (building obstruction) or mixed light source (LED + incandescent lamp) environments, and improve lighting robustness.

[0057] One-dimensional convolutional neural network autoencoder end-to-end processing can efficiently separate remote photoplethysmography pulse signals from specular reflection and motion noise, significantly improving the purity of physiological signals, ensuring excellent noise reduction, and guaranteeing the accuracy of index calculation.

[0058] By simultaneously calculating multiple physiological indicators such as heart rate, heart rate variability, and blood oxygen saturation, and combining physiological indicator grading standards with the barrel effect model, the final suitability assessment level is determined and a graded early warning signal is generated. The assessment is comprehensive and the judgment is scientific, thus improving the reliability of screening.

[0059] In some implementations, the calculation formula for the adaptive gamma correction algorithm is as follows:

[0060] ;

[0061] in, Intensity values ​​of the image sequence are corrected for illumination. Intensity values ​​of an overexposed or underexposed RGB three-channel spatiotemporal image sequence; The optimal value obtained by dynamically calculating the pixel intensity distribution of an image through K-means clustering analysis. value.

[0062] Optimum values ​​are dynamically calculated using K-means clustering. The value can accurately adapt to overexposure and underexposure under different lighting conditions; compared to a fixed value... Value correction, dynamic This makes illumination correction more intelligent and flexible, effectively improving image quality under complex lighting conditions and enhancing the illumination robustness of non-contact physiological detection.

[0063] In some implementations, the step of dynamically calculating the optimal γ value by analyzing the image pixel intensity distribution through K-means clustering specifically includes:

[0064] Normalize the image pixel intensity to the range [0,1];

[0065] The K-means clustering algorithm is applied to divide the pixels into K clusters, and the mean of each cluster is calculated.

[0066] Dynamically adjust according to the distribution characteristics of the cluster Values ​​were calculated, with those having a mean greater than 0.9 being overexposed clusters and those having a mean less than 0.1 being underexposed clusters. Optimal values ​​were dynamically calculated for both overexposed and underexposed clusters. value.

[0067] By normalizing image pixel intensity to the [0,1] range, the analysis scale is unified, providing a consistent benchmark for clustering; K-means clustering accurately divides pixel clusters, objectively identifying overexposed (>0.9) and underexposed (<0.1) areas through cluster means, avoiding subjective judgment bias; dynamic calculation is performed for overexposed and underexposed clusters. Values ​​are used to achieve differentiated correction, which is relatively fixed. It is better suited to complex lighting scenarios, can specifically optimize image quality in extreme lighting areas, improve the accuracy and adaptability of lighting correction, and lay a more reliable image foundation for subsequent remote photoplethysmography pulsation signal extraction.

[0068] In some implementations, the heart rate is calculated using the following formula:

[0069] HR = 60 / MeanRR;

[0070] ;

[0071] Wherein, HR is the heart rate, representing the average heart rate; MeanRR is the average RR interval, representing the average time interval between adjacent heartbeats; Total duration represents the total duration of the electrocardiogram signal; and Number of RR intervals represents the total number of heartbeats.

[0072] MeanRR, based on the average of the RR intervals, directly correlates with the actual time interval between adjacent heartbeats in the electrocardiogram signal, accurately reflecting the heartbeat cycle and making the heart rate results more physiologically accurate. Calculating the average by combining the total duration and the total number of heartbeats smooths out instantaneous fluctuations, reduces the interference of accidental abnormalities on the results, and improves the stability and accuracy of heart rate calculation.

[0073] In some embodiments, the formula for calculating the blood oxygen saturation is:

[0074] ;

[0075] ;

[0076] ;

[0077] in, Blood oxygen saturation, ΔA R The AC / DC ratio of the signal in channel R, ΔA G The AC / DC ratio of the signal in channel G, AC R For the AC component of the R channel, DC R For the DC component of channel R, AC G For the AC component of the G channel, DC G Here, denoted as DC component of channel G, and a and b are experimental calibration coefficients.

[0078] The AC / DC ratio ΔA of the signal through channel R R The AC / DC ratio ΔA of the G channel signal G By combining calculations and utilizing the different sensitivities of the two channels to changes in blood oxygen, the detection dimensionality and accuracy are improved. The introduction of alternating current (AC) and direct current (DC) components can effectively separate blood oxygen-related pulsating signals from the baseline light intensity, reducing interference from non-physiological factors. The coefficients a and b are experimentally calibrated and can be adapted to specific detection scenarios, further ensuring the consistency between the calculation results and the actual blood oxygen level, providing reliable support for the accurate acquisition of blood oxygen saturation in non-contact physiological detection.

[0079] In some implementations, the formula for calculating the AC component is:

[0080] ;

[0081] ;

[0082] Where AC represents the AC component and N is the sample size. This represents the original light intensity signal value at the i-th sampling point. This is the mean of the original light intensity signal values ​​at all sampling points.

[0083] By adopting the standard deviation form and using the mean of the original light intensity signal of all sampling points as the benchmark, the signal fluctuation amplitude can be accurately quantified, the DC component can be effectively separated, and the AC component related to physiological activities can be accurately extracted. Secondly, by introducing the number of samples N to cover the calculation of multiple sampling points, the random error of a single sampling point can be smoothed, the interference of extreme values ​​on the results can be reduced, and the stability and reliability of the AC component calculation can be improved.

[0084] In some implementations, the formula for calculating the DC component is:

[0085] ;

[0086] Where DC is the direct current component, N is the number of samples, and z t This is the estimated DC component of the original light intensity signal at time t.

[0087] z is obtained by using the sample size N for multiple time t. t The average value can smooth out individual z-axis values. t This reduces the interference of random errors and instantaneous fluctuations on the results, making the DC component more closely match the stable baseline level of the original light intensity signal, thus improving computational stability; t It directly corresponds to the DC component estimate of the original light intensity signal at time t. During calculation, it can accurately focus on the stationary part of the signal and effectively remove the AC component that changes with physiological activities.

[0088] In some implementations, the barrel effect model selects the worst-performing physiological indicator among all physiological indicators as the final suitability assessment level.

[0089] Based on the "weakest link" principle, this method takes a holistic approach, selecting the higher level of physiological indicators as the key criterion for determining job suitability. This comprehensive assessment method can fully and objectively consider an individual's health level, providing solid and reliable decision support for the rational allocation of positions and efficient personnel deployment, while ensuring both individual health and high-quality work performance.

[0090] By directly selecting the worst-performing physiological indicator from all indicators as the final suitability assessment level, the judgment logic is intuitive and objective, eliminating the need for complex multi-indicator weighting or subjective weighing, thus avoiding interference from human factors and ensuring the fairness and consistency of the assessment results. Secondly, it precisely focuses on physiological deficiencies, aligning with the core safety requirements of high-risk operations such as power, where any failure to meet the standard may lead to safety risks. Determining the level based on the worst indicator completely avoids the hidden danger of missing a single indicator, maximizing the physiological suitability of workers before they start their jobs. At the same time, the above rules are simple and easy to understand, allowing for rapid level determination without additional complex calculations, meeting the efficiency requirements of rapid pre-job screening and shortening the screening process time.

[0091] In some implementations, the graded warning signals include green (normal), yellow (warning), and red (prohibition from working).

[0092] By adopting a color-coded system of green, yellow, and red, the system aligns with the public's general understanding of safety warnings, is intuitive and easy to understand, and is suitable for the fast-paced screening scenarios in high-risk work sites such as power plants. This reduces the time spent interpreting information, facilitates rapid decision-making and intervention by management personnel, and ensures the efficiency and safety of pre-job screening. Secondly, the system has clear classifications that cover key scenarios: green corresponds to normal work, yellow to warning and attention, and red to prohibit work. This accurately matches the job suitability requirements under different physiological states, avoiding delays in screening normal personnel while promptly intercepting at-risk individuals and eliminating safety hazards caused by abnormal indicators.

[0093] In some implementations, the graded early warning signals are fed back in real time through an LED screen; if all physiological indicators are normal, the LED screen lights up green; if an indicator is at a critical value, the LED screen lights up yellow and prompts attention; if any indicator exceeds the safe range, the LED screen lights up red and issues an audible and visual alarm, prohibiting the employee from working.

[0094] The screening results are displayed instantly on the LED screen, providing strong real-time feedback without waiting for data transmission or manual summarization. This meets the timeliness requirements of rapid pre-job screening and significantly reduces decision-making time. When the light turns red, an audible and visual alarm is triggered simultaneously. Compared to a single visual prompt, this more strongly reminds managers to intercept unsuitable personnel, avoiding safety hazards caused by information omissions and effectively ensuring the effectiveness and safety of pre-job screening for high-risk operations.

[0095] In some embodiments, the end-to-end denoising processing of the illumination-corrected image sequence using a one-dimensional convolutional neural network autoencoder denoising algorithm specifically includes:

[0096] Each frame of the face is sliced, and each slice is considered as a sample of a remote photoplethysmography (PPG) signal. Recording for 10 seconds yields the sample track corresponding to each slice. The mathematical model for the average pixel value of each slice is as follows:

[0097] ;

[0098] Where I(i, t) is the average pixel value of the i-th slice at time t, m(i, t) is the specular reflection component, s(i, t) is the remote photoplethysmography pulsation signal component, n(i, t) is the motion noise component, and a and b are coefficients; each face slice contains a part of the signal, and different components are separated according to different sample tracks.

[0099] In some implementations, the input dimension of the one-dimensional convolutional neural network autoencoder denoising algorithm is [batch, H×W (total number of slices), 250 (length)]. After convolution and pooling, it is encoded as [batch, N (number of slices), L (encoding dimension)], and then reconstructed as [batch, 1, L] to achieve denoising.

[0100] During the encoding process, multiple channels are combined and superimposed to form a suitable reconstructed signal. The activation function is a non-linear thresholding process, and noise is eliminated in this part. Finally, only the signal reconstructed based on the encoded vector remains, thus completing the separation and noise reduction of the signal.

[0101] In some implementations, the heart rate variability is calculated as follows:

[0102] The obtained RR curves were analyzed in the time and frequency domains to extract heart rate variability.

[0103] 1) Time-domain analysis

[0104] ①PNN50: The number of times the difference between adjacent RR intervals is greater than 50 milliseconds;

[0105] ②RR Triangular Index: The ratio of the area of ​​the triangle formed by the RR interval sequence;

[0106] ;

[0107] Wherein, Number of RR intervals represents the total number of RR intervals, and Height of the R-Rinterval histogram represents the height of the RR interval histogram;

[0108] 2) Frequency Domain Analysis

[0109] The Fourier transform (FFT) model was used for frequency domain analysis. The frequency domain analysis indicators included ultra-low frequency power (VLF), low frequency power (LF), high frequency power (HF), low frequency / high frequency ratio (LF / HF Ratio), normalized low frequency power (NLF), and normalized high frequency power (NHF).

[0110] ① VLF: ;

[0111] ② LF: ;

[0112] ③ HF: ;

[0113] ④ NLF: Normalized Low-Frequency Power;

[0114] ;

[0115] ⑤ NHF: Normalized high-frequency power;

[0116] ;

[0117] ⑥ LF / HF Ratio: Low frequency / high frequency ratio, the ratio of low frequency (LF) and high frequency (HF) components, used to reflect the balance between the sympathetic and parasympathetic nervous systems;

[0118] ;

[0119] Wherein, LF power represents the power of the low-frequency components, and HF power represents the power of the high-frequency components.

[0120] Please see Figure 2 Another embodiment of the present invention provides a non-contact rapid physiological state screening system for implementing the aforementioned non-contact rapid physiological state screening method, comprising:

[0121] The perception layer includes a data acquisition module, which is used to acquire RGB three-channel spatiotemporal image sequences of facial regions of interest.

[0122] The processing layer includes an edge computing module, which runs an adaptive gamma correction algorithm and a one-dimensional convolutional neural network autoencoder denoising algorithm to calculate physiological indicators.

[0123] The decision-making level is used to run physiological indicator grading standards and the barrel effect model to generate graded early warning signals.

[0124] The data acquisition module of the perception layer focuses on the region of interest on the face to acquire RGB three-channel spatiotemporal image sequences. The non-contact acquisition does not require wearable devices, which can avoid the risks of skin allergies and cross-infection. At the same time, the region of interest is accurately focused on the area with effective physiological information, reducing interference from background irrelevant data.

[0125] Secondly, the processing layer relies on the edge computing module to run adaptive gamma correction algorithm and one-dimensional convolutional neural network autoencoder denoising algorithm. Edge computing does not rely on cloud transmission and can quickly perform adaptive gamma correction to eliminate light interference and one-dimensional convolutional neural network autoencoder denoising to separate the remote photoplethysmography pulsation signal from specular reflection and motion noise. It also efficiently calculates physiological indicators, greatly shortens data processing time, and meets the timeliness requirements of pre-job rapid screening. It can complete single-person screening within 10 seconds, supports multi-person parallel processing and data caching, and adapts to the needs of high-concurrency operation scenarios.

[0126] Meanwhile, the decision-making level generates graded early warning signals through physiological indicator grading standards and the barrel effect model. The judgment logic is objective, and the worst indicator is used to determine the grade to accurately avoid safety hazards. Moreover, the output results directly correspond to the screening purpose.

[0127] The three-layer architecture of perception, processing and decision-making ensures that each layer performs its function and connects efficiently, thus guaranteeing the accuracy of data collection, the real-time processing, and the security of decision-making.

[0128] In some embodiments, the sensing layer further includes a 4K high-definition industrial-grade visible light camera, which integrates an infrared fill light unit and an automatic white balance adjustment unit to ensure image quality in multi-light source scenarios.

[0129] The 4K high-definition industrial-grade visible light camera's high resolution can accurately capture subtle image details of facial regions of interest, ensuring high definition and high fidelity of RGB three-channel spatiotemporal image sequences. The integrated infrared illumination unit can supplement light in low-light and dimly lit scenarios, avoiding image blurring and data loss due to insufficient light, and ensuring continuous and stable acquisition in complex lighting environments. The automatic white balance adjustment unit can adapt to different light sources in real time, such as natural light and indoor lighting, correcting color deviations and ensuring accurate reproduction of RGB channel color information. This avoids interference from light source color deviations with remote photoplethysmography pulsation signal extraction, further improving the reliability of data acquisition.

[0130] In some implementations, please refer to Figure 3 The edge computing module includes an illumination correction unit, a one-dimensional convolutional neural network noise reduction unit, and a physiological indicator calculation unit.

[0131] The illumination correction unit dynamically calculates the optimal γ value through K-means clustering and uses an adaptive gamma correction algorithm to perform frame-by-frame illumination correction on the RGB three-channel spatiotemporal image sequence to obtain an illumination-corrected image sequence.

[0132] The one-dimensional convolutional neural network noise reduction unit is used to perform end-to-end noise reduction processing on the illumination-corrected image sequence using a one-dimensional convolutional neural network autoencoder noise reduction algorithm, separating the remote photoplethysmography pulsation signal from specular reflection and motion noise.

[0133] The physiological index calculation unit is used to calculate heart rate, heart rate variability and blood oxygen saturation based on the separated remote photoplethysmography pulse signal.

[0134] An adaptive gamma correction algorithm and a one-dimensional convolutional neural network autoencoder denoising algorithm are employed to dynamically eliminate interference from strong light, shadows, and mixed light sources. The illumination correction unit dynamically calculates the optimal γ value through K-means clustering, processes the RGB sequence frame by frame, and accurately eliminates overexposed / underexposed areas. The one-dimensional convolutional neural network denoising unit performs end-to-end denoising processing, which can effectively separate the remote photoplethysmography (PPG) pulse signal from specular reflection and motion noise, reduce non-physiological interference, and ensure signal purity. The physiological index calculation unit relies on the purified PPG pulse signal to simultaneously calculate heart rate, heart rate variability, and blood oxygen saturation, covering multi-dimensional physiological information. All units are integrated into the edge computing module, eliminating the need for cloud reliance, enabling fast processing speed, adapting to the needs of rapid pre-job screening, and providing accurate and efficient index data support for subsequent decision-making and early warning.

[0135] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned non-contact rapid screening method for physiological states.

[0136] The main functions of this invention are:

[0137] This invention utilizes a 4K high-definition industrial-grade visible light camera integrating an infrared illumination unit and an automatic white balance adjustment unit to ensure image quality in multi-light source scenarios. The integrated infrared illumination unit provides supplemental lighting in low-light and dimly lit environments, preventing image blurring and data loss due to insufficient light. The automatic white balance adjustment unit adapts to different light sources in real time, correcting color deviations and ensuring accurate reproduction of RGB channel color information. This prevents interference from light source color deviations with remote photoplethysmography (PPG) pulsation signal extraction, further improving the reliability of data acquisition. An adaptive gamma correction algorithm is also employed. With a one-dimensional convolutional neural network autoencoder denoising algorithm, it dynamically eliminates interference from strong light, shadows, and mixed light sources; through an edge computing module, it can complete single-person screening within 10 seconds, supports multi-person parallel processing and data caching, and adapts to the needs of high-concurrency work scenarios; it grades the calculated physiological indicators through a physiological indicator grading standard, and quantifies individual health deficiencies through the barrel effect model to determine the final job suitability assessment level, ensuring the scientific and objective nature of the assessment results; through the early warning response of LED screen colors (green / yellow / red) and audible and visual alarms, it enables real-time risk intervention and job suitability decisions.

[0138] The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions can be stored as one or more instructions or codes on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope and spirit of this invention and the appended claims. For example, due to the nature of software, the functions described above can be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units can be integrated into a single processing unit, or each unit can exist physically separately, or two or more units can be integrated into a single unit.

[0139] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0140] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0141] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0142] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A non-contact method for rapid screening of physiological states, characterized by: The method comprises the following steps: S1: acquiring a face video of a subject, and collecting an RGB three-channel spatiotemporal image sequence of a face region of interest from the face video; S2: performing frame-by-frame illumination correction on the RGB three-channel spatiotemporal image sequence by using an adaptive gamma correction algorithm to obtain an illumination correction image sequence; the adaptive gamma correction algorithm dynamically calculates an optimal gamma value by performing K-means clustering analysis on image pixel intensity distribution; S3: performing end-to-end noise reduction processing on the illumination correction image sequence by using a one-dimensional convolutional neural network autoencoder noise reduction algorithm to separate a remote photoplethysmography pulsatile signal; S4: calculating physiological indexes based on the separated remote photoplethysmography pulsatile signal; S5: performing grading determination on the calculated physiological indexes based on a physiological index grading standard, combining a bucket effect model to determine a final suitability for duty grade, and generating a grading warning signal; The calculation formula of the adaptive gamma correction algorithm is: ； wherein, correcting intensity values of the image sequence under illumination; correcting intensity values of the RGB three-channel spatio-temporal image sequence under overexposure or underexposure; γ is an optimal γ value dynamically calculated by K-means clustering analysis of image pixel intensity distribution; The dynamic calculation of the optimal gamma value by the K-means clustering analysis on the image pixel intensity distribution specifically comprises the following steps: normalizing the image pixel intensity to the range of [0, 1]; applying a K-means clustering algorithm to divide the pixels into K clusters and calculating the mean value of each cluster; dynamically adjusting the gamma value according to the distribution characteristics of the clusters, wherein the clusters with a mean value greater than 0.9 are overexposure clusters, and the clusters with a mean value less than 0.1 are underexposure clusters, and the optimal gamma value is dynamically calculated for the overexposure clusters and the underexposure clusters; The physiological indexes include heart rate, heart rate variability, and blood oxygen saturation, and the calculation formula of the heart rate is: HR = 60 / MeanRR; ； wherein HR is the heart rate, MeanRR is the average value of R-R intervals, Total duration represents the total duration of the electrocardiogram signal, and Number of R-R intervals represents the total number of heartbeats.

2. The non-contact physiological state rapid screening method according to claim 1, characterized in that: The calculation formula of the blood oxygen saturation is: ； ； ； wherein, SpO2 is the blood oxygen saturation, ΔA R is the signal AC / DC ratio for the R channel, ΔA G is the signal AC / DC ratio for the G channel, AC R is the AC component for the R channel, DC R is the DC component for the R channel, AC G is the AC component for the G channel, DC G is the DC component for the G channel, a, b are experimental calibration coefficients.

3. The non-contact physiological state rapid screening method of claim 1, wherein: The bucket effect model selects the worst one among all the physiological indexes as the final suitability for duty grade; and the grading warning signal includes green normal, yellow warning, and red prohibition of on-duty.

4. A non-contact physiological status rapid screening system for implementing a non-contact physiological status rapid screening method according to any one of claims 1-3, characterized in that, The method comprises the following steps: a perception layer comprising a data acquisition module for acquiring an RGB three-channel spatiotemporal image sequence of a face region of interest; a processing layer comprising an edge computing module for running an adaptive gamma correction algorithm and a one-dimensional convolutional neural network autoencoder noise reduction algorithm to calculate physiological indexes; a decision layer for running a physiological index grading standard and a bucket effect model to generate a grading warning signal; The edge computing module comprises an illumination correction unit, a one-dimensional convolutional neural network noise reduction unit, and a physiological index calculation unit; The illumination correction unit dynamically calculates an optimal gamma value by K-means clustering and performs frame-by-frame illumination correction on the RGB three-channel spatiotemporal image sequence by using an adaptive gamma correction algorithm to obtain an illumination correction image sequence; The one-dimensional convolutional neural network denoising unit is configured to perform end-to-end denoising processing on the illumination correction image sequence by using a one-dimensional convolutional neural network auto-encoder denoising algorithm, and separate the remote photoplethysmography pulsatile signal from the mirror reflection and motion noise. The physiological index calculation unit is configured to calculate the heart rate, heart rate variability and blood oxygen saturation based on the separated remote photoplethysmography pulsatile signal.

5. A computer-readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements a non-contact physiological state rapid screening method as claimed in any one of claims 1-3.

Images