A non-contact driver heart rate detection method

Abstract

The present invention relates to the field of remote photoelectric volume scanning, and more particularly to a non-contact driver heart rate detection method. The method acquires a facial image of the driver and obtains an rPPG signal matrix using a KTL algorithm, guided filtering, and the like. The rPPG signal matrix is ​​subjected to dimensionality reduction, ICA, and fast Fourier transform to ultimately obtain a current heart rate measurement value. A Kalman filter is used to calculate a heart rate prediction value at the previous moment, and a current heart rate prediction value is obtained based on the current heart rate measurement value and the previous heart rate prediction value.

Description

Technical Field

[0001] The present invention relates to the field of remote photoelectric volume scanning, and in particular to a non-contact driver heart rate detection method. Background Art

[0002] Heart rate refers to the number of heartbeats per minute and is one of the indicators of physical and mental health. A person's heart rate when they are quiet and not engaging in strenuous exercise is called their resting heart rate. Heart rate monitoring can be categorized as either contact or non-contact, depending on the detection technology. Contact monitoring uses a photoplethysmography (PPG) method using smart devices such as ECGs. This requires close contact with the skin, which can cause significant discomfort, irritation, or pain, affecting test results. It can also be more irritating for infants, burn victims, and other groups.

[0003] With the widespread adoption of computer technology and cameras, researchers have proposed remote photoplethysmography (rPPG). Like traditional PPG, rPPG uses the same principle: each heartbeat causes periodic changes in the blood and microvasculature in human skin tissue, leading to periodic variations in light absorption and reflection. While these periodic variations are invisible to the human eye, they can be analyzed from video images captured by high-definition cameras, enabling heart rate detection. Therefore, this method is also known as super-sensitive heart rate detection. Its advantages extend beyond the fact that the person being tested does not need to wear any equipment; it can be used for long-term physiological monitoring. Traffic accidents are frequent in today's society. Besides well-known factors like drunk driving and speeding, the driver's health is also a major contributing factor. Heart rate, a golden indicator of physiological health, can be effectively assessed through contactless heart rate estimation based on video streams. However, this method is susceptible to motion artifacts caused by head movement and the effects of lighting. Summary of the Invention

[0004] To solve the above problems, the present invention provides a non-contact driver heart rate detection method, which is characterized by comprising the following steps:

[0005] S1. Install a camera directly in front of the driver's head and use the camera to capture the driver's facial image video;

[0006] S2. Use the dlib face detector in OpenCV to detect the face rectangle of each face image in the face image video frame, and locate the coordinates of 68 facial feature points of the face in the face rectangle;

[0007] S3. Use the KLT algorithm to track the driver's head movement by analyzing the displacement of 68 facial feature points in the facial image video, and output a stabilized facial image video.

[0008] S4. Perform guided filtering on each frame of the stabilized face image in the video to obtain a filtered face video;

[0009] S5. Calculate the average vector of the forehead, left cheek, right cheek, and nose of each frame of the filtered face image video, and perform a weighted average of the average vectors of the four regions of each frame of the filtered face image to obtain the rPPG intensity signal of each frame of the filtered face image;

[0010] S6. Summarize all rPPG intensity signals to obtain an rPPG signal matrix, and perform dimensionality reduction processing on the rPPG signal matrix to obtain a low-dimensional rPPG signal matrix;

[0011] S7. Separate the low-dimensional rPPG signal matrix by ICA to obtain the blood static and dynamic signal vectors, and perform fast Fourier transform on the blood dynamic signal vectors to obtain the current heart rate measurement value;

[0012] S8. Calculate the heart rate prediction value at the previous moment through the Kalman filter, and obtain the heart rate prediction value at the current moment based on the heart rate measurement value at the current moment and the heart rate prediction value at the previous moment.

[0013] Furthermore, in step S4, guided filtering is performed on each face image frame in the stabilized face image video to obtain a filtered face image. The local linear relationship between the guided image and the filtered face image is expressed as:

[0014] q i =a k I i +b k

[0015] In order to determine the linear coefficients of the guided image and the filtered face image while minimizing the difference between the stabilized face image and the filtered face image, the guided filtering process is transformed into an optimization problem, which can be expressed as:

[0016]

[0017] Solving the optimization problem, we get:

[0018]

[0019]

[0020] Among them, I i represents the gray value of the i-th pixel of the guide image I in the local window, pi represents the gray value of the i-th pixel of the stable face image p in the local window, q i represents the grayscale value of the i-th pixel of the filtered face image q in the local window, ω k represents a local window of length k, a k represents the local window ω k The first constant linear coefficient, b k represents the local window ω k The second constant linear coefficient, ω represents the local window ω k The total number of pixels in ,∈ represents the regularization coefficient, represents the local window ω k The mean of all second constant linear coefficients in μ k Indicates that the image I is guided in the local window ω k The mean value calculated from all pixels in Indicates that the image I is guided in the local window ω k The variance calculated for all pixels within Indicates that the face image p is stabilized in the local window ω k The average gray value calculated for all pixels in the .

[0021] Furthermore, according to solving the optimization problem, the linear coefficient between the filtered face image and the guide image is obtained, which is expressed as:

[0022]

[0023] in, Represents the average grayscale value of the i-th pixel in the local window of the filtered face image q.

[0024] Furthermore, in step S6, the rPPG signal matrix is ​​multiplied by the dimensionality reduction matrix to obtain a low-dimensional rPPG signal matrix. The dimensionality reduction matrix is ​​expressed as:

[0025]

[0026] Where T represents the transpose operation of the matrix.

[0027] Furthermore, step S7 specifically includes:

[0028] S11. Define the low-dimensional rPPG signal matrix as S and calculate the mean-reduced low-dimensional rPPG signal matrix The calculation formula is:

[0029]

[0030] Where E{S} represents the identity matrix of the low-dimensional rPPG signal matrix S;

[0031] S12. Calculate the covariance matrix of the de-meaned low-dimensional rPPG signal matrix and perform eigenvalue decomposition on it to obtain a whitening matrix;

[0032] S13. The whitening matrix is ​​processed by ICA to obtain a blood static and dynamic signal vector; the blood static and dynamic signal vector includes two different components: a blood static signal component and a blood dynamic signal component;

[0033] S14. Perform a fast Fourier transform on the blood static and dynamic signal vector. The component with the lowest peak value in the transformed blood static and dynamic signal vector is the blood static signal component.

[0034] S15. Use a bandpass filter to process the blood dynamic signal component, and use the peak frequency of the processed blood dynamic signal component as the heart rate measurement value at the current moment.

[0035] Furthermore, step S12 calculates the covariance matrix of the mean-removed low-dimensional rPPG signal matrix, including:

[0036] Cov(S)=QΛQ T

[0037] Cov(S)=SS T =PRov(S)P T =E

[0038]

[0039] Wherein, Q represents an orthogonal matrix composed of eigenvectors, Λ represents a symmetric matrix composed of eigenvalues, Cov(S) represents the covariance matrix of the low-dimensional rPPG signal matrix S, P represents the exchange matrix, Rov(S) represents the autocorrelation matrix of the low-dimensional rPPG signal matrix S, and E represents the unit matrix.

[0040] Furthermore, the formula for calculating the heart rate prediction value using Kalman filtering is:

[0041]

[0042] E n =E n-1 +k n ·(M n -x n-1 )

[0043] P n =(1-k n )·P n +Q

[0044] Among them, P n represents the error covariance matrix, k nrepresents the Kalman gain, R represents the measurement noise covariance matrix, E n Represents the optimal estimated value at time n, M n represents the measured value, x n-1 represents the prior estimate, and Q represents the measurement process covariance.

[0045] Beneficial effects of the present invention:

[0046] The accuracy of heart rate detection based on facial videos is easily affected by factors such as illumination changes and rigid head motion. To address these issues, this paper combines face detection and video tracking algorithms to capture facial images, effectively reducing the impact of head rigidity. When extracting rPPG signals from the face, the strength of the rPPG signal is related to the selection of the facial region of interest and the illumination distribution of the region of interest. Considering that the selection of a single region of interest may be affected by other factors, thereby reducing the strength of the rPPG signal, for example, the forehead is easily affected by hair and glasses. At the same time, the effective area of ​​the selected region of interest also affects the strength of the rPPG signal, so this paper uses a combination of the forehead, left cheek, right cheek, and nose as the region of interest to address the impact of a single region of interest on heart rate calculation. Irregular facial contours can cause skin pigmentation, resulting in uneven illumination distribution and reducing the signal-to-noise ratio of rPPG signal extraction. This paper uses guided filtering to smooth the facial skin, effectively addressing the uneven illumination distribution caused by irregular facial contours, thereby improving the signal-to-noise ratio of the rPPG signal. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is an overall flow chart of the heart rate measurement method of the present invention;

[0048] Figure 2 This is a comparison diagram of the image before and after guided filtering of the present invention;

[0049] Figure 3 The present invention obtains an initial estimated heart rate value diagram by performing fast Fourier transform on the blood dynamic signal vector;

[0050] Figure 4 This is a diagram of the estimated value of the heart rate obtained after Kalman filtering in the present invention. DETAILED DESCRIPTION

[0051] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0052] The present invention provides a non-contact driver heart rate detection method, such as Figure 1 Shown, including:

[0053] S1. Install a camera directly in front of the driver's head and use the camera to capture the driver's facial image video;

[0054] S2. Use the dlib face detector in OpenCV to detect the face rectangle of each face image in the face image video frame, and locate the coordinates of 68 facial feature points of the face in the face rectangle;

[0055] Specifically, the facial key point detection based on the 68-feature principle is used to realize facial feature point marking, which can automatically find the locations of the eyes, nose, mouth and facial contour landmark features on the face based on the detected face.

[0056] S3. Use the KLT algorithm to track the driver's head movement by analyzing the displacement of 68 facial feature points in the facial image video, and output a stabilized facial image video.

[0057] Specifically, early rPPG-based heart rate detection research was usually conducted when the measured object was stationary. This is because movement can cause changes in the distance and angle between the region of interest (ROI) and the camera and light source, thereby introducing greater interference and noise. The present invention is aimed at filming drivers in motion, and the acquired images will be disturbed by movement. The KLT algorithm can effectively track the 68 facial feature points of the face, thereby eliminating the impact of motion artifacts on the rPPG signal and obtaining a better rPPG signal in the motion state.

[0058] S4. Perform guided filtering on each frame of the stabilized face image video to obtain a filtered face image video; Figure 2 As shown: The concave areas of the face are removed by guided filtering to obtain a smooth face image.

[0059] S5. Calculate the average vector of the forehead, left cheek, right cheek, and nose of each frame of the filtered face image video, and perform a weighted average of the average vectors of the four regions of each frame of the filtered face image to obtain the rPPG intensity signal of each frame of the filtered face image;

[0060] Specifically, the human face has other tissues besides skin, and only the reflected light passing through the skin contains physiological information. If the effective area of ​​the selected skin region is too small, it will affect the accuracy of the heart rate estimation. Due to direct sunlight, some skin areas on the face will have uneven light distribution, which will also affect the accuracy of heart rate calculation if involved in the calculation. Therefore, four areas are selected: forehead, left cheek, right cheek, and nose to calculate their respective average value vectors.

[0061] S6. Summarize all rPPG intensity signals to obtain an rPPG signal matrix, and perform dimensionality reduction processing on the rPPG signal matrix to obtain a low-dimensional rPPG signal matrix;

[0062] S7. Separate the low-dimensional rPPG signal matrix by ICA to obtain the blood static and dynamic signal vector, and perform fast Fourier transform on the blood dynamic signal vector to obtain the current heart rate measurement value; Figure 3 As shown: By performing fast Fourier transform on the static and dynamic signal vectors of the blood, the heart rate measurement value of the human body at the current moment of image acquisition is obtained.

[0063] S8. Calculate the heart rate prediction value of the previous moment through the Kalman filter, and obtain the heart rate prediction value of the current moment based on the heart rate measurement value of the current moment and the heart rate prediction value of the previous moment. Figure 4 As shown: The human heart rate prediction value is obtained accurately as the image is collected through Karl filtering.

[0064] Specifically, when a camera sensor receives reflected light from the face, it detects irregular bumps and pigmentation, as the human face is not an ideal reflector. Scars and acne marks, for example, can create noise when collecting rPPG signals, affecting the accuracy of heart rate estimation. Guided filtering effectively reduces noise other than illumination and motion artifacts, facilitating more efficient rPPG signal extraction.

[0065] In one embodiment, it is assumed that the input image (stabilized face image) is p, the output image (filtered face image) is q, the guide image is I, and the subscript i Represented as a pixel in the image.

[0066] In this embodiment, an image with feature points as boundaries is selected as the guide image I. The local linear model assumes that a point on a function has a linear relationship with its neighboring points. A complex function can be represented by many local linear functions. When the value of a point on the function is required, it is only necessary to calculate the values ​​of all the linear functions containing the point and take the average. We can regard the image as a two-dimensional function, then q and I are in the local window ω centered at pixel k. k There is a local linear relationship in:

[0067] q i =a k I i +b k

[0068] Among them, a k represents the local window ω k The first constant linear coefficient of the k-th pixel, b k represents the local window ω k The second constant linear coefficient of the k-th pixel in .

[0069] At the same time, the output image q is the input image p after noise removal, and the two have the following relationship:

[0070] q i =p i -n i

[0071] Among them, n i Represents the noise at the i-th pixel of the output image q.

[0072] In order to determine the linear coefficients of the output image and the guide image and to minimize the difference between the input image and the output image, it is converted into an optimization problem, which can be expressed as:

[0073]

[0074] Solving the optimization problem, we get:

[0075]

[0076]

[0077] According to the local linear relationship between the filtered face image and the guide image, it is calculated:

[0078]

[0079]

[0080]

[0081] Among them, I i represents the gray value of the i-th pixel in the local window of the guidance image I, p i represents the gray value of the i-th pixel in the local window of the stable face image p, q i represents the grayscale value of the i-th pixel in the local window of the filtered face image q, ω k represents a local window of length k, a k represents the local window ω kThe first constant linear coefficient, b k represents the local window ω k The second constant linear coefficient, ω represents the local window ω k The total number of pixels within, ∈ represents the regularization coefficient; in the guided filtering process, the size of the local window changes continuously and finally converges. In this process, multiple b k value, represents the local window ω k The mean of all second constant linear coefficients in μ k Represents the guidance image I in the local window ω k The mean within Represents the guidance image I in the local window ω k The variance within Represents the stable face image p in the local window ω k The average gray value within.

[0082] Human skin absorbs and reflects visible light, and a camera's photosensor receives this reflected light. The photosensor converts the received reflected light into an electrical signal, which is then converted to a digital signal via AD processing. As light passes through skin tissue and then reflects back to the sensor, it attenuates. While muscle, bone, veins, and other connective tissues absorb light in a largely constant manner, blood, unlike other tissues, experiences non-uniform blood flow in arteries, causing hemoglobin concentration to fluctuate periodically with heart movement. Consequently, light absorption also fluctuates periodically, reflecting heart rate information.

[0083] Specifically, the illumination reflection model assumes that the image is composed of illumination intensity, a static blood signal component, and a dynamic blood signal component. The dynamic blood signal component contains heart rate information, and illumination intensity is irrelevant to the static blood signal component and the dynamic blood signal component. Therefore, this embodiment performs dimensionality reduction processing on the rPPG signal matrix by filtering out illumination intensity through dot multiplication of the dimensionality reduction matrix to obtain a low-dimensional rPPG signal matrix. The dimensionality reduction matrix is ​​expressed as:

[0084]

[0085] Specifically, the low-dimensional rPPG signal matrix is ​​processed by ICA to separate the blood static signal component and the blood dynamic signal component. The specific process is as follows:

[0086] S11. Define the low-dimensional rPPG signal matrix as S and calculate the mean-reduced low-dimensional rPPG signal matrix The calculation formula is:

[0087]

[0088] Where E{S} represents the identity matrix of the low-dimensional rPPG signal matrix S;

[0089] S12. Calculate the covariance matrix of the de-meaned low-dimensional rPPG signal matrix and perform eigenvalue decomposition on it to obtain a whitening matrix;

[0090] Specifically, the process of calculating the covariance matrix of the mean-reduced low-dimensional rPPG signal matrix is:

[0091] Cov(S)=QΛQ T

[0092] Cov(S)=SS T =PRov(S)P T =E

[0093]

[0094] Wherein, Q represents an orthogonal matrix composed of eigenvectors, Λ represents a symmetric matrix composed of eigenvalues, Cov(S) represents the covariance matrix of the low-dimensional rPPG signal matrix S, P represents the exchange matrix, Rov(S) represents the autocorrelation matrix of the low-dimensional rPPG signal matrix S, and E represents the unit matrix.

[0095] Specifically, the specific process of obtaining the whitening matrix is:

[0096] First, the data is centered, that is, the observation vector is subtracted from its mean:

[0097] C=Cov(S)-E1

[0098] Where E1 is the mean of the covariance matrix Cov(S) of the low-dimensional rPPG signal matrix S, and C is the matrix after the low-dimensional rPPG signal matrix S is centralized;

[0099] Whitening refers to the process of making the observation vectors uncorrelated and having unit variance through linear transformation.

[0100] Let the matrix of linear transformation be X, then the whitening matrix V is expressed as:

[0101] V=XC

[0102] So that every vector in V are irrelevant, and E{VV T}=I,where X=E -1 / 2 U T ,

[0103] E=diag(d1,d2,....,d n ), d i is the correlation matrix R X =E{XX TIf the signals extracted by ICA from the mixed signal are independent of each other and do not obey the Gaussian distribution, then such extracted signals are called target signals;

[0104] Assume that the static blood signal is Y. According to the basic design requirements of ICA, the target signals are independent of each other and do not obey the Gaussian distribution. Then:

[0105] Y=W T1 V

[0106] Where V represents the whitening matrix, W represents the coefficient matrix of the mixed signal decomposition, and T1 represents the periodic length of the decomposition data.

[0107] Among them, W should be offset from the kurtosis of the Gaussian distribution as much as possible. The farther the kurtosis of the V value is from the kurtosis of the Gaussian distribution, the more it can satisfy the maximization of the non-Gaussian degree, and then the target signal Y can be obtained.

[0108] S14. Perform a fast Fourier transform on the blood static and dynamic signal vector. The component with the lowest peak value in the transformed blood static and dynamic signal vector is the blood static signal component.

[0109] S15. Use a bandpass filter to process the blood dynamic signal component, and use the peak frequency of the processed blood dynamic signal component as the heart rate measurement value at the current moment.

[0110] Specifically, considering that the measurement value at a single moment is still affected by slight movements of the subject such as shaking the head and rocking, there will be certain deviations. Therefore, this embodiment gradually converges the heart rate value to an accurate value through Kalman filtering while ensuring that the subject's heart rate does not change suddenly and drastically.

[0111] The formula for calculating the heart rate prediction value using Kalman filtering is:

[0112]

[0113] E n =E n-1 +k n ·(M n -x n-1 )

[0114] P n =(1-k n )·P n +Q

[0115] Among them, P n represents the error covariance matrix, k n represents the Kalman gain, R represents the measurement noise covariance, E n Represents the optimal estimate at time n, M n represents the measured value, x n-1Represents the prior estimate, and Q represents the covariance of the measurement process.

[0116] In the present invention, unless otherwise clearly stipulated and limited, the terms "installation", "setting", "connection", "fixation", "rotation" and the like should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be the internal connection of two elements or the interaction relationship between two elements. Unless otherwise clearly defined, ordinary technicians in this field can understand the specific meanings of the above terms in the present invention according to the specific circumstances.

[0117] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A non-contact driver heart rate detection method, characterized in that: The following steps are involved: S1. Install a camera directly in front of the driver's head and use the camera to capture the driver's facial image video; S2. Use the dlib face detector in OpenCV to detect the face rectangle of each face image in the face image video frame, and locate the coordinates of 68 facial feature points of the face in the face rectangle; S3. Use the KLT algorithm to track the driver's head movement by analyzing the displacement of 68 facial feature points in the facial image video, and output a stabilized facial image video. S4. Perform guided filtering on each frame of the stabilized face image in the video to obtain a filtered face video; S5. Calculate the average vector of the forehead, left cheek, right cheek, and nose of each frame of the filtered face image in the filtered face image video, and perform a weighted average of the average vectors of these four regions to obtain the rPPG signal strength of each frame of the filtered face image; S6. Summarize all rPPG intensity signals to obtain an rPPG signal matrix, and perform dimensionality reduction processing on the rPPG signal matrix to obtain a low-dimensional rPPG signal matrix; S7. Separate the low-dimensional rPPG signal matrix by ICA to obtain the blood static and dynamic signal vectors, and perform fast Fourier transform on the blood dynamic signal vectors to obtain the current heart rate measurement value; S8. Calculate the heart rate prediction value at the previous moment through the Kalman filter, and obtain the heart rate prediction value at the current moment based on the heart rate measurement value at the current moment and the heart rate prediction value at the previous moment.

2. A non-contact driver heart rate detection method according to claim 1, characterized in that: In step S4, the guide image is used to perform guided filtering on each frame of the stabilized face image in the stabilized face image video to obtain a filtered face image. The local linear relationship between the guide image and the filtered face image is expressed as: q i =a k I i +b k In order to determine the linear coefficients of the guided image and the filtered face image while minimizing the difference between the stabilized face image and the filtered face image, the guided filtering process is transformed into an optimization problem, which can be expressed as: Solving the optimization problem, we get: Among them, I i represents the gray value of the i-th pixel of the guide image I in the local window, p i represents the gray value of the i-th pixel of the stable face image p in the local window, q i represents the grayscale value of the i-th pixel of the filtered face image q in the local window, ω k represents a local window of length k, a k represents the local window ω k The first constant linear coefficient, b k represents the local window ω k The second constant linear coefficient, ω represents the local window ω k The total number of pixels in ,∈ represents the regularization coefficient, represents the local window ω k The mean of all second constant linear coefficients in μ k Indicates that the image I is guided in the local window ω k The mean value calculated from all pixels in Indicates that the image I is guided in the local window ω k The variance calculated for all pixels within Indicates that the face image p is stabilized in the local window ω k The average gray value calculated for all pixels in the .

3. The non-contact driver heart rate detection method according to claim 2, characterized in that: According to solving the optimization problem, the linear coefficient between the filtered face image and the guide image is obtained, which is expressed as: in, Represents the average grayscale value of the i-th pixel in the local window of the filtered face image q.

4. The non-contact driver heart rate detection method according to claim 1, characterized in that: In step S6, the rPPG signal matrix is ​​multiplied by the dimensionality reduction matrix to obtain a low-dimensional rPPG signal matrix. The dimensionality reduction matrix is ​​expressed as: Where T represents the transpose operation of the matrix.

5. The non-contact driver heart rate detection method according to claim 1, characterized in that: Step S7 specifically includes: S11. Define the low-dimensional rPPG signal matrix as S and calculate the mean-reduced low-dimensional rPPG signal matrix The calculation formula is: Where E{S} represents the identity matrix of the low-dimensional rPPG signal matrix S; S12. Calculate the covariance matrix of the de-meaned low-dimensional rPPG signal matrix and perform eigenvalue decomposition on it to obtain a whitening matrix; S13. The whitening matrix is ​​processed by ICA to obtain a blood static and dynamic signal vector; the blood static and dynamic signal vector includes two different components: a blood static signal component and a blood dynamic signal component; S14. Perform a fast Fourier transform on the blood static and dynamic signal vector. The component with the lowest peak value in the transformed blood static and dynamic signal vector is the blood static signal component. S 15. Use a bandpass filter to process the blood dynamic signal component, and use the peak frequency point of the processed blood dynamic signal component as the heart rate measurement value at the current moment.

6. The non-contact driver heart rate detection method according to claim 5, characterized in that: Step S12 calculates the covariance matrix of the mean-removed low-dimensional rPPG signal matrix, including: Cov(S)=QΛQ T Cov(S)=SS T =PRov(S)P T =E Wherein, Q represents an orthogonal matrix composed of eigenvectors, Λ represents a symmetric matrix composed of eigenvalues, Cov(S) represents the covariance matrix of the low-dimensional rPPG signal matrix S, P represents the exchange matrix, Rov(S) represents the autocorrelation matrix of the low-dimensional rPPG signal matrix S, and E represents the unit matrix.

7. The non-contact driver heart rate detection method according to claim 1, characterized in that: The formula for calculating the heart rate prediction value using Kalman filtering is: E n =E n-1 +k n ·(M n -x n-1 ) P n =(1-k n )·P n +Q Among them, P n represents the error covariance matrix, k n represents the Kalman gain, R represents the measurement noise covariance matrix, E n represents the optimal estimated value at time n, M n represents the measured value, x n-1 represents the prior estimate, and Q represents the covariance matrix of the measurement process.

Images