CN113657345B - A non-contact heart rate variability feature extraction method based on real application scenarios - Google Patents

Abstract

The invention discloses a non-contact heart rate variability feature extraction method based on a realistic application scene, which comprises the following steps: 1) Collecting an image containing a human face; 2) Acquiring an image face area by using a strategy combining face detection and tracking; 3) Euler amplification enhances skin color variation; 4) Color space conversion is separated from the channel; 5) Adaptive threshold skin detection; 6) Extracting a source signal; 7) EEMD denoising; 8) Five-point sliding smooth filtering; 9) Peak value point detection and abnormal peak value point correction; 10 RR interval calculation and time domain, frequency domain, nonlinear HRV feature extraction. The method integrates a strategy for improving the feature extraction speed, a strategy for overcoming the influence of shaking and illumination and improving the feature extraction accuracy into the feature extraction method, can rapidly extract the HRV features in a non-contact mode in a practical application environment, and ensures consistency with a contact type extraction result.

Description

A non-contact heart rate variability feature extraction based on real application scenarios method

Technical field

The invention relates to the technical fields of computer vision and signal processing, and in particular to a non-contact heart rate variability feature extraction method.

Background technique

Heart rate variability (HRV) is an important indicator for evaluating autonomic nervous activity and the intrinsic dynamic mechanism of the heart. The phenomenon of heart rate variability originates from the regulatory effect of the autonomic nervous system on heart rate. The interaction between the sympathetic nervous system and the parasympathetic nervous system causes periodic changes in heart rate. Heart rate variability is widely used in research on heart disease, mental illness, emotion recognition, etc.

Heart rate variability features are mainly extracted from ECG signals collected by contact methods. In real-life application scenarios, using contact methods to obtain heart rate variability characteristics requires subjects to actively cooperate and wear relevant collection equipment, which brings a lot of inconvenience to real-life applications. Especially when heart rate variability features are used for emotion recognition and contact methods are used to collect heart rate variability features, the influence of human contact factors during the collection process will also have an impact on the extracted heart rate variability features of the subject.

Non-contact heart rate detection methods mainly include laser Doppler technology, microwave or millimeter wave Doppler radar and thermal imaging technology. However, the above-mentioned technology generally has high equipment cost, and long-term use of the equipment will have an impact on the human body, so it is not suitable for widespread use in practical applications. In recent years, imaging photoplethysmography (IPPG) developed based on photoplethysmography (PPG) has been able to obtain human heart rate information more accurately, which also makes non-contact heart rate variability measurement possible through this principle. .

The general process of extracting heart rate variability features through IPPG is to first collect a video containing a face through a camera, perform face detection on the video to delineate the face area, and then select a specific area from the face area as the ROI area, and then select this area Channel separation is performed through color space, the single-channel or multi-channel pixel mean of each frame is calculated, and the heart rate value is obtained by applying a signal processing method based on the difference between the pixel mean values. However, in real application scenarios, the problem of directly applying the above process to extract HRV features is that face detection needs to be carried out frame by frame, and the extraction speed is slow, which cannot meet the needs of practical applications. In addition, in real application scenarios, environmental factors such as shaking of the subject and different lighting conditions will also affect the accuracy of heart rate variability feature extraction, and the above extraction process cannot solve this problem well.

Contents of the invention

In order to solve the above problems, the present invention starts from two aspects: the non-contact heart rate variability feature extraction speed is slow, and different lighting conditions and shaking have a great impact on the accuracy of non-contact heart rate variability feature extraction. Based on the IPPG principle, the invention integrates image processing technology, signal Processing technology and feature extraction technology, a non-contact heart rate variability feature extraction method based on real application scenarios is proposed. The extracted features are consistent with the contact extraction results.

The non-contact heart rate variability feature extraction method proposed by the present invention proposes a solution strategy that combines face detection and face tracking to obtain the face area in order to improve the extraction speed of the heart rate variability feature. The face position is obtained through face detection. Then the face position is tracked, and at the same time during the tracking process, the face position is repositioned through face detection at fixed time intervals to prevent tracking offset, and the face offset caused by shaking is continuously corrected. Prevent incomplete extraction of the face area, thereby ensuring that the accuracy of face area acquisition is not affected while increasing the speed. In addition, the present invention also places image acquisition and image processing from the camera in two threads and relies on a shared queue to achieve simultaneous processing, thereby better improving the extraction speed.

In order to reduce the impact of different lighting conditions on the extraction results, the present invention proposes a solution strategy that combines channel separation, adaptive skin detection, and EEMD filtering. First, convert the face area image to the LUV color space, separate the L channel that reflects the brightness change, obtain the U channel image that reflects the chroma change, and convert the face area image to the YCrCb color space. According to the brightness under different lighting conditions The Y component is combined with the Cb component to adaptively determine the threshold to perform skin detection to obtain the skin part of the face area image, and calculate the skin detection results and the U channel image to obtain the original heart rate signal, initially reducing the impact of changes in light intensity, and applying the EEMD method to the original heart rate The signal is denoised to further reduce the impact of illumination changes.

In order to reduce the impact of the subject's shaking on the extraction results, the present invention proposes a peak point extraction strategy that can correct the peak points affected by shaking. First, calculate the signal peak points, and then determine whether each extracted peak point is affected by shaking and causes extraction abnormalities by using the slope between adjacent peak points of the signal and whether the distance between peak points is within the threshold range. For the peak points where extraction abnormalities occur, The abnormal peak point is corrected by averaging the positions of the normal peak points before the current abnormal peak point, thereby obtaining a relatively accurate signal peak point, thereby reducing the impact of shaking on feature extraction.

In order to achieve the above objectives, the collection process proposed by the present invention is:

Step 1. Collect images containing faces:

The subject faces the camera and collects face images at a fixed frame rate of 30FPS. The subject must continuously collect face images for at least 30 seconds to extract relatively accurate heart rate variability features.

Among them, the storage of collected images containing faces needs to be carried out in a sub-thread. That is to say, in the program, the collection of images and the reading of images for processing should be carried out in two threads at the same time. The two threads An image queue will be shared.

Step 2. Obtain the face area of the image:

Using the libfacedetection open source face detection library for face detection and the KLT (Kandade-Lucas-Tomasi) tracking method for face tracking, this method extracts the facial area, the face detection obtains the face position, and then performs the face position on the face. tracking, and at the same time, face detection is reused to locate the face position at a fixed time interval of 10 seconds during the tracking process and then the tracking is continued.

During the face tracking process, a minimum circumscribed rectangle is determined through the four vertices of the face area determined after face tracking, and the face affected by shaking is corrected through the center point and deflection angle of the rectangle to prevent Facial area extraction is incomplete.

Step 3. Euler amplification:

The Euler amplification method is used to enhance skin color changes in the face area and enhance the information related to physiological signals in the face image.

Step 4. Channel separation:

Convert the Euler-enlarged face image obtained in step 3 from the RGB color space to the LUV color space. The purpose is to separate the L channel that reflects the brightness change, and use the U channel that reflects the chromaticity change to extract the original heart rate signal. .

Step 5. Adaptive threshold skin detection:

An adaptive threshold is proposed for skin detection. The threshold is adaptively determined based on the brightness component Y combined with the Cb component under different lighting conditions. The range that meets the threshold is skin pixels, and its pixels are set to 255 white, and the rest are set to 0 black. The skin detection image is ANDed with the U channel to remove non-skin areas, and a U-channel face image with non-skin areas removed is obtained.

Step 6. Source signal extraction:

In a round of HRV feature extraction process, the pixel mean is calculated for all U-channel face images with non-skin areas removed in step 5, and a series of pixel mean points are obtained, and then the standardized calculation is performed on this series of pixel points to form Raw heart rate signal.

Step 7. EEMD denoising:

Use the EEMD (Ensemble Empirical Mode Decomposition) method to denoise the original heart rate signal to further reduce the impact of different lighting conditions.

Step 8. Five-point sliding:

Apply five-point sliding smoothing filter method to remove high-frequency noise still contained in the signal

Step 9. Peak point extraction and correction:

Calculate the signal peak point, and by setting the distance between the two peak points and the threshold range of the slope, find the peak point affected by shaking and correct it to obtain a relatively accurate signal peak point.

Step 10. HRV feature extraction:

Use the corrected peak point obtained in step nine to calculate the RR interval and R point time to extract HRV features. A total of 27 HRV features including time domain features, frequency domain features, and nonlinear features were extracted.

Among them, time domain features include: max, min, mean, median, SDNN, RMSSD, hr-mean, hr-sd, NN40, pNN40 or HRVti;

Among them, the frequency domain features are extracted by spectrum analysis using the Lomb-Scargle periodogram. The frequency domain features include: aVLF, aLF, aHF, aTotal, pVLF, pLF, pHF, nLF, nHF, LFHF, peakVLF, peakLF or peakHF;

Among them, nonlinear characteristics include: SD1, SD2 or SD1/SD2.

The advantages and positive effects of the present invention are:

The present invention proposes a non-contact heart rate variability feature extraction method based on real application scenarios. Specifically, the extraction method incorporates a strategy to improve the speed of non-contact heart rate variability feature extraction, and overcomes the different effects of shaking and lighting conditions. , thereby improving the accuracy of non-contact heart rate variability feature extraction. In real application scenarios, real application functions such as automatic switching of testers, automatic detection, and automatic calculation of HRV characteristics can be realized based on the present invention, which has strong practical application significance.

Description of the drawings

Figure 1 is a flow chart of the technical solution of the present invention.

Figure 2 is a flow chart of a strategy for improving the speed of feature extraction by combining face detection and face tracking to obtain face areas.

Figure 3 is a flow chart of a strategy that combines channel separation, adaptive skin detection, and EEMD filtering to reduce the impact of different lighting conditions.

Figure 4 is a flow chart of a peak point extraction strategy that can correct peak points affected by shaking.

Figure 5 shows the face correction rotation coordinate system.

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings.

The present invention provides a non-contact heart rate variability feature extraction method based on real application scenarios, which can quickly and more accurately extract heart rate variability features in real application scenarios, and can be used for emotion recognition to reflect the subject The level of psychological stress can be applied to real-life scenarios such as assisting customs staff in screening suspicious customs clearance personnel.

Referring to Figure 2, the present invention proposes a strategy for improving the speed of feature extraction by combining face detection and face tracking to obtain face areas. Obtain the face position through face detection and then track the face position. At the same time, during the tracking process, the face position is repositioned through face detection at fixed time intervals to prevent tracking offset, and the tracking deviation caused by shaking is continuously detected. Face offset performs face correction to prevent incomplete extraction of the face area, thereby ensuring that the accuracy of the face area acquisition is not affected while increasing the speed, and the present invention also separates image acquisition and image processing from the camera into two The threads rely on a shared queue to achieve simultaneous processing, thereby better improving the extraction speed.

Referring to Figure 3, the present invention proposes a strategy that combines channel separation, adaptive skin detection, and EEMD filtering to reduce the impact of different lighting conditions. First, convert the face area image to the LUV color space, separate the L channel that reflects the brightness change, obtain the U channel image that reflects the chroma change, and convert the face area image to the YCrCb color space. According to the brightness under different lighting conditions The Y component is combined with the Cb component to adaptively determine the threshold for skin detection to obtain the skin part of the face area image. The skin detection results and the U channel image are calculated to obtain the original heart rate signal to initially reduce the impact of different lighting conditions. The EEMD method is applied to the source signal. Perform noise reduction to further reduce the impact of different lighting conditions.

Referring to Figure 4, the peak point extraction strategy proposed by the present invention can correct the peak points affected by shaking. First, calculate the signal peak points, and then determine whether each extracted peak point is affected by shaking and causes extraction abnormalities by using the slope between adjacent peak points of the signal and whether the distance between peak points is within the threshold range. For the peak points where extraction abnormalities occur, The abnormal peak point is corrected by averaging the positions of the normal peak points before the current abnormal peak point, thereby obtaining a relatively accurate signal peak point, thereby reducing the impact of shaking on feature extraction.

Referring to Figure 1, the non-contact heart rate variability feature extraction method based on real application scenarios proposed by the present invention in combination with the above strategies mainly includes four parts: data acquisition part, image processing part, signal processing part, and HRV feature extraction part. It can be subdivided into 10 steps: collecting images containing faces, obtaining image facial areas, Euler amplification, channel separation, adaptive threshold skin detection, source signal extraction, EEMD denoising, five-point sliding, peak point extraction and correction , HRV feature extraction.

Specific steps are as follows:

Step 1. Collect images containing faces:

The subject faces the camera and collects face images at a fixed frame rate of the camera. Most of the frame rates of ordinary USB cameras on the market are now 30FPS, so the present invention assumes that face images are collected at 30FPS. The subject must continuously collect facial images for at least 30 seconds to extract relatively accurate heart rate variability features.

Since the present invention is based on real application scenarios, it does not meet the requirements if the face video is recorded first and then the heart rate variability features are extracted from the face video as in the general process. Therefore, when the present invention is applied, three threads are required to be executed simultaneously, one thread is responsible for collecting face images, one thread is responsible for processing face images, and one thread is responsible for signal processing and feature extraction.

Since the processing is divided into threads, one thread collects face images and another thread processes the face images, a shared space is needed to ensure that one thread stores the collected face images in the shared space in order, and the other thread collects the face images from the shared space. Face images are read in order for processing. Because access and read operations need to be performed in order and fast, it is most suitable to use the queue data structure to implement this shared space.

Step 2. Obtain the face area of the image:

The extraction speed of the traditional non-contact heart rate variability feature extraction process is relatively slow. Traditional non-contact feature extraction performs face detection on each frame of image. Although this method can stably extract the face area, it is very slow and inefficient. Facial area extraction is crucial in the overall process of HRV feature extraction, and facial area extraction processes images. Compared with pure numerical calculations such as signal processing and emotion classification, it itself will account for most of the time-consuming program running. , so optimizing this stage will greatly improve the overall operating speed of the system.

Starting from this problem, the present invention uses the libfacedetection open source face detection library for face detection and the KLT (Kandade-Lucas-Tomasi) tracking method for face tracking. This method extracts the facial area. This extraction method is fast. The facial area extraction is stable and does not affect the implementation of other functions (for example, in real application scenarios, key functions such as automatic switching of subjects need to be considered). Face detection in this extraction method can be regarded as face tracking to provide a tracking template, and in order to prevent The tracking position is offset relative to the face position. During the tracking process, the face will be re-detected and repositioned at a fixed time interval for face tracking.

In actual application scenarios, shaking should also be considered. Using face detection plus face tracking will result in incomplete facial area extraction when shaking, so face correction is required. Traditional face correction is generally First, facial feature points are extracted, and then the pupil position is located, and the face is corrected through the deflection angle of the line between the two pupils. Although this method can effectively correct the face position, the correction speed is slow. If used in the non-contact heart rate variability feature extraction process, it will seriously slow down the extraction speed. Therefore, the present invention determines a minimum circumscribed rectangle through the four vertices of the face area determined after face tracking, and performs correction through the center point and deflection angle of the rectangle. Referring to Figure 5, the deflection angle of the rectangle is determined by a coordinate system. When the rectangle is shifted to the right, in the first quadrant of the coordinate system, the deflection angle is based on the positive semi-axis of the x-axis as 0 degrees. When the rectangle is offset to the left, the angle is determined by the second quadrant, and the offset angle is based on the positive half-axis of the y-axis as 0 degrees. When the angle is between 0-45 degrees, use the current degree to determine the rotation angle clockwise. When the angle is between 45-90 degrees, use 90 minus the degree to determine the rotation angle counter-clockwise. Through the center point of the rectangle and the deflection angle of the rectangle, construct an affine transformation matrix, perform affine transformation on the entire image according to the affine transformation matrix, and re-intercept the image according to the center point of the rectangle and the length and width of the rectangle. That's it. Now comes the corrected face. Although this method is not as effective as the face feature point correction scheme, it is fast and the correction results can meet the application requirements.

Step 3. Euler amplification:

In the present invention, the Euler amplification method is used to enhance the skin color changes in the face area and enhance the information related to physiological signals in the face image.

Among them, the number of spatial decomposition layers in the Euler amplification method is 6, the time domain filtering frequency band is 1-2Hz, and the image amplification factor is 200.

Step 4. Channel separation:

In the present invention, in order to reduce the impact of different lighting conditions on HRV feature extraction, the Euler-amplified face image obtained in step 3 is converted from the RGB color space to the LUV color space. The purpose is to convert the L that reflects the brightness change into The channels are separated, and the U channel that reflects the chromaticity changes is used to extract the original heart rate signal.

Step 5. Adaptive threshold skin detection:

Because the face area image obtained through face detection and face tracking also contains face parts in non-skin areas, these parts will affect the accuracy of HRV feature extraction, so it is necessary to filter out non-skin areas of the face, and due to the reality The lighting conditions in application scenarios are different, so the skin detection effect of a single threshold is not good. The present invention proposes an adaptive threshold for skin detection. The threshold is determined adaptively based on the brightness component Y combined with the Cb component under different lighting conditions. The range that meets the threshold is the skin pixel. Set its pixels to 255 white and the rest to 0 black. The skin detection image is ANDed with the U channel to remove non-skin areas, and a U-channel face image with non-skin areas removed is obtained.

Define dynamic configuration rules in YcrCb color space as follows:

θ ₃ =6; θ ₄ =-8

if(Y≤128)θ ₁ =6; θ ₂ =12;

If the Cr value of a pixel meets the following conditions, the pixel is a skin pixel

c _r ≥-2(c _b +24); c _r ≥-(c _b +17);

c _r ≥-4(c _b +32); c _r ≥2.5(c _b +θ ₁ );

c _r ≥ θ ₃ ; c _r ≥ 0.5(θ ₄ -c _b );

Among them, Y is the brightness component, Cb is the blue chroma component, Cr is the red chroma component, and θ1 to θ4 are intermediate variables.

Step 6. Source signal extraction:

In the present invention, for one round of HRV feature extraction process, the pixel mean is calculated for all U-channel face images with non-skin areas removed in step 5, and a series of pixel mean points are obtained, and then this series of pixel points are standardized. The processing constitutes the original heart rate signal.

Step 7. EEMD denoising:

In real application scenarios, different lighting conditions have a great impact on HRV feature extraction. Experimental results show that the lower the illumination, the greater the error in the extracted HRV features. In order to reduce errors caused by different lighting conditions in HRV extraction, the present invention adopts the EEMD (Ensemble Empirical Mode Decomposition) method for noise reduction processing. Apply the EEMD method to adaptively obtain IMF components at different resolutions at each scale, calculate the instantaneous frequency through Hilbert transform, and distinguish the noise-dominated IMF component and the signal-dominated IMF component based on the instantaneous frequency. The noise-dominated IMF component is discarded and retained. The signal-dominated IMF component reconstructs the signal, thereby reducing the impact on HRV feature extraction under different lighting conditions.

Step 8. Five-point sliding:

The five-point moving average filter is a low-pass filter that can effectively remove high-frequency noise still contained in the signal. Apply five-point moving average filtering to further filter the signal after EEMD filtering in step 7 to make the signal smoother.

The calculation formula for the i-th new data of the five-point moving average is:

Among them, N is the signal length, f(j) is the signal value within the five-point sliding window, and y(i) is the new signal value determined by the five-point sliding average.

Step 9. Peak point extraction and correction:

When the subject's head shakes, it will cause the heart rate variability curve to jitter, which will lead to inaccurate peak point extraction of the heart rate variability curve. Therefore, the present invention proposes a peak point extraction strategy that can correct the peak points affected by shaking. First, the signal peak point is calculated for the smooth signal obtained after step eight, and then based on the slope between adjacent peak points and whether the distance between peak points is within the threshold. Within the range, it is determined whether the peak point extraction is affected by shaking and an extraction abnormality occurs. For the peak point where the extraction abnormality occurs, the abnormal peak point is corrected by averaging the positions of the normal peak points before the current abnormal peak point, and all abnormal peak points are corrected. This results in a relatively accurate signal peak point.

In the present invention, the formula for determining whether the slope between adjacent peak points is within the threshold range is as follows:

Among them, h _i represents the height of the i-th (i=1,2,...,n) peak point, and t _i represents the time corresponding to the i-th (i=1,2,...,n) peak point. If this is not satisfied, The formula represents that the i-th peak point is an abnormal peak point.

In the present invention, the formula for determining whether the distance between adjacent peak points is within the threshold range is as follows:

60/(HR-14)≤t _i -t _i-1 <60/(HR+14)，(i＝2,3,…,n)

Among them, t _i represents the time corresponding to the i-th (i=1,2,...,n) peak point, and HR represents the average heart rate. If this formula is not satisfied, it means that the i-th peak point is an abnormal peak point, where the average heart rate HR The calculation formula is as follows:

Among them, t _all represents the total duration of detection, and count represents the total number of peak points in detection.

For the detected abnormal peak points, the calculation formula for correcting them is as follows:

t _{F_new} ＝t _F-1 +[(t _F-1 -t _F-2 )+…+(t ₂ -t ₁ )]/(F-2)

Among them, t _{F_new} is the corrected result, (t _F-1 ,...,t ₁ ) is the normal peak point or the corrected peak point before the abnormal peak point to be corrected.

Step 10. HRV feature extraction:

Use the corrected peak point obtained in step nine to calculate the RR interval and R point time to extract HRV features. A total of 27 HRV features including time domain features, frequency domain features, and nonlinear features were extracted.

Among them, time domain features include: max, min, mean, median, SDNN, RMSSD, hr-mean, hr-sd, NN40, pNN40 or HRVti;

Among them, the frequency domain features are extracted by spectrum analysis using the Lomb-Scargle periodogram. The frequency domain features include: aVLF, aLF, aHF, aTotal, pVLF, pLF, pHF, nLF, nHF, LFHF, peakVLF, peakLF or peakHF;

Among them, nonlinear characteristics include: SD1, SD2 or SD1/SD2.

The above are the specific embodiments of the present invention.

Claims (9)

1. A non-contact heart rate variability feature extraction method based on real application scenarios, which is characterized by including the following steps: Step 1. Collect images containing faces: Collect images containing faces of the subject through the camera at a fixed frame rate; Step 2: Use face detection and face tracking to obtain the facial area of the image; Step 3: Apply the Euler amplification method to enhance skin color changes in the face area, and enhance the information related to physiological signals in the face image; Step 4: Adaptively determine the threshold value based on the brightness component Y combined with the Cb component under different lighting conditions for skin detection, and obtain a U-channel image with non-skin areas removed; Step 5: Convert the Euler-enlarged face image from RGB color space to LUV color space. The purpose is to separate the L channel that reflects brightness changes, and use the U channel that reflects chroma changes to extract the original heart rate signal; Step 6: Calculate the pixel average of the U channel image and standardize it to obtain the original heart rate signal; Step 7: Apply the EEMD method to denoise the original heart rate signal; Step 8: Apply a five-point sliding smoothing filter to remove the high-frequency noise still contained in the signal; Step 9: Calculate the signal peak point, and by setting the distance between the two peak points and the threshold range of the slope, find the peak point affected by shaking and correct it to obtain a relatively accurate signal peak point; Step 10: Use the corrected RR interval and R point time to extract HRV features, and extract a total of 27 HRV features including time domain features, frequency domain features, and nonlinear features; Step nine also includes: The formula for determining whether the slope between adjacent peak points is within the threshold range is as follows: Among them, h _i represents the height of the i-th peak point, t _i represents the time corresponding to the i-th peak point, if this formula is not satisfied, it means that the i-th peak point is an abnormal peak point; i=1,2,…,n ; The formula for determining whether the distance between adjacent peak points is within the threshold range is as follows: 60/(HR-14)≤t _i -t _i-1 <60/(HR+14), i＝2,3,…,n Among them, t _i represents the time corresponding to the i-th peak point, i = 1, 2,..., n, HR represents the average heart rate. If this formula is not satisfied, it means that the i-th peak point is an abnormal peak point, where the average heart rate HR Calculated as follows: Among them, t _all represents the total duration of detection, and count represents the total number of peak points in detection; For the detected abnormal peak points, the calculation formula for correcting them is as follows: t _{F_new} ＝t _F-1 +[(t _F-1 -t _F-2 )+…+(t ₂ -t ₁ )]/(F-2) Among them, t _{F_new} is the corrected result, t _F-1 ,...,t ₁ is the normal peak point or the corrected peak point before the abnormal peak point to be corrected. 2. A non-contact heart rate variability feature extraction method based on real-life application scenarios as claimed in claim 1, characterized in that in the second step, face detection and face tracking are used to obtain the face image. Regional specific steps include: 1) Use the libfacedetection open source face detection library for face detection and the KLT tracking method for face tracking to extract facial areas; 2) Face detection obtains the face position and then tracks the face position; 3) Correct faces affected by shaking during tracking; 4) During the tracking process, face detection is used again to locate the face position at a fixed time interval of 10 seconds before continuing tracking. 3. A non-contact heart rate variability feature extraction method based on real application scenarios as claimed in claim 2, characterized in that the specific steps of correcting faces affected by shaking include: 1) Determine a minimum circumscribed rectangle through the four vertices of the face area determined after face tracking, and correct the face through the center point and deflection angle of this rectangle; 2) The deflection angle of the rectangle is determined by a coordinate system. When the rectangle shifts to the right, in the first quadrant of the coordinate system, the deflection angle is based on the positive half-axis of the x-axis as 0 degrees. When the rectangle shifts to the left, The angle is determined by the second quadrant, and the offset angle is based on the positive half axis of the y-axis as 0 degrees; 3) When the deflection angle of the rectangle is between 0-45 degrees, use the current degree to determine the rotation angle as clockwise rotation. When the deflection angle of the rectangle is between 45-90 degrees, use 90 minus the degree to determine the rotation angle as counter-clockwise rotation; 4) Construct an affine transformation matrix through the center point of the rectangle and the deflection angle of the rectangle, and perform affine transformation on the entire image according to the affine transformation matrix; 5) Re-capture the image according to the center point of the rectangle and the length and width of the rectangle to get the corrected face. 4. A non-contact heart rate variability feature extraction method based on real application scenarios as claimed in claim 1, characterized in that the implementation of the steps one and two includes: The storage of collected images containing faces needs to be carried out in a sub-thread. That is to say, in the program, the collection of images and the reading of images for processing should be carried out simultaneously in two threads. The two threads will share a Image queue. 5. A non-contact heart rate variability feature extraction method based on real application scenarios as claimed in claim 1, characterized in that the step three applies the Euler amplification method to enhance skin color changes in the face area. The specific application parameters of the information related to physiological signals in the face image are: the number of spatial decomposition layers is 6, the time domain filtering frequency band is 1-2Hz, and the image amplification factor is 200. 6. A non-contact heart rate variability feature extraction method based on real application scenarios as claimed in claim 1, characterized in that the step four further includes: The threshold is determined adaptively based on the brightness component Y combined with the Cb component under different lighting conditions. The range that meets the threshold is the skin pixel, and its pixel points are set to 255 white, and the rest are set to 0 black. The skin detection image is ANDed with the U channel, so that Remove the non-skin areas and obtain the U-channel face image with the non-skin areas removed; Define dynamic configuration rules in YcrCb color space as follows: if(Y>128) θ ₃ =6; θ ₄ =-8; if(Y≤128)θ ₁ =6; θ ₂ =12; If the Cr value of a pixel meets the following conditions, the pixel is a skin pixel: c _r ≥-2(c _b +24); c _r ≥-(c _b +17); c _r ≥-4(c _b +32); c _r ≥2.5(c _b +θ ₁ ); c _r ≥ θ ₃ ; c _r ≥ 0.5(θ ₄ -c _b ); Among them, Y is the brightness component, Cb is the blue chroma component, Cr is the red chroma component, and θ1 to θ4 are intermediate variables. 7. A non-contact heart rate variability feature extraction method based on real application scenarios as claimed in claim 1, characterized in that the step seven of step 7 applies the EEMD method to denoise the original heart rate signal. The specific steps include: 1) Calculate the instantaneous frequency through Hilbert transform; 2) Distinguish the noise-dominated IMF component and the signal-dominated IMF component according to the instantaneous frequency; 3) Discard the noise-dominated IMF component obtained in 2) and reconstruct the signal-dominated IMF component. 8. A non-contact heart rate variability feature extraction method based on real application scenarios as claimed in claim 1, characterized in that the step eight further includes: The calculation formula for the five-point moving average is: Among them, N is the signal length, f(j) is the signal value within the five-point sliding window, and y(i) is the new signal value determined by the five-point sliding average. 9. A non-contact heart rate variability feature extraction method based on real application scenarios as claimed in claim 1, characterized in that the step ten further includes: Time domain features include: max, min, mean, median, SDNN, RMSSD, hr-mean, hr-sd, NN40, pNN40 or HRVti; Frequency domain features are extracted using spectrum analysis using the Lomb-Scargle periodogram. Frequency domain features include: aVLF, aLF, aHF, aTotal, pVLF, pLF, pHF, nLF, nHF, LFHF, peakVLF, peakLF or peakHF; Non-linear features include: SD1, SD2 or SD1/SD2.