A non-contact heart rate detection method for removing irregular motion artifacts

Abstract

The application discloses a non-contact heart rate detection method for removing motion artifacts, comprising the following steps: 1, tensorizing original video frame sequences, decomposing low rank and sparse components, thereby obtaining low rank video image frame sequences, and realizing image level denoising; 2, extracting pixel mean time series of a region of interest, and using a fully adaptive noise ensemble empirical mode decomposition method to obtain a target blood volume pulse signal; 3, performing spectrum analysis on the target blood volume pulse signal, obtaining a heart rate main frequency, and then performing corresponding operation to obtain a heart rate value. The application simultaneously removes irregular motion noise from the image level and the signal level, completes accurate heart rate detection in a real scene, provides a new direction for non-contact heart rate detection research, and further promotes the application of the non-contact detection technology in the real scene.

Description

Technical Field

[0001] This invention belongs to the field of video-based non-contact physiological parameter detection. Specifically, it is a motion-robust non-contact video heart rate detection method that integrates low-rank tensor factorization (LTF) and complementary ensemble empirical mode decomposition with adaptive noise (CEEMDAN), denoted as LFT-CEEMDAN. Background Technology

[0002] Heart rate monitoring has a wide range of applications, including healthcare, fitness, mental health, driving safety, human-computer interaction, and biometrics. It can help people understand their physical condition, adjust exercise intensity, relieve stress, improve safety, enhance interactive experiences, and protect privacy.

[0003] Currently, there are many methods for heart rate detection, mainly divided into two categories: contact and non-contact. Traditional contact heart rate measurement devices, such as electrocardiographs and pulse oximeters, typically require sensors to be attached to the skin or worn on the wrist, chest strap, or ear. They measure heart rate through principles such as electrical signals, sound waves, optics, or bioimpedance, which can cause discomfort to the individual. Non-contact methods do not require skin contact. Instead, they utilize technologies such as cameras, radar, microwaves, or infrared light to capture changes in heart rate. Among these, remote photoplethysmography (rPPG) detects changes in blood volume in facial capillaries in video images in conjunction with the physiological activity of the heartbeat, thereby extracting blood volume pulse signals that reflect heart rate information. Due to the widespread use, low cost, and lack of need for any worn devices, cameras are widely used, typically in newborn health monitoring, home health monitoring, and motion monitoring.

[0004] The blood volume pulse signal amplitude contained in rPPG is very weak. However, in monitoring scenarios simulating daily life, there is noise interference such as irregular movement or sudden changes in lighting. Furthermore, the amplitude of this noise interference is much larger than that of the blood volume pulse (BVP) signal because movement can cause changes in facial color, brightness, and position, affecting the performance of rPPG heart rate detection. Therefore, remote photoplethysmography (RPPG) is highly susceptible to motion noise interference, posing a significant challenge to the accuracy of heart rate measurement. Studies have shown that the green channel contains strong pulsation information; therefore, single-channel signal decomposition processing techniques can be used to improve the quality of the blood volume pulse signal. To address the noise reduction problem, the components containing pulse information can be identified through source separation, such as blind signal separation methods. The most representative blind signal separation methods are independent component analysis (ICA) and principal component analysis (PCA). With the development of artificial intelligence technology, deep learning has also been applied to remote photoplethysmography. In addition, existing research has proposed combining different types of methods to improve pulse signal quality.

[0005] Principal component analysis (PCA) is a fundamental method in data analysis. It leverages the low-dimensional structure within high-dimensional data, a structure commonly found in various data types such as images, text, video, and bioinformatics. It is computationally efficient and powerful for data instances slightly corrupted by minor noise. However, a major drawback of PCA is its vulnerability to severely corrupted or anomalous observations, which are ubiquitous in real-world data. To date, many robust versions of PCA have been proposed, but many suffer from high computational costs.

[0006] Robust principal component analysis (RPCA) is the first polynomial-time algorithm with a strong recovery guarantee. Suppose we are given an observation matrix... It can be decomposed into D = A0 + E0, where E0 is low-rank sparse. If the singular vectors of E0 satisfy some incoherent conditions, such as E0 being low-rank and sufficiently sparse, then E0 and E0 can be recovered with high probability by solving the following convex problem.

[0007] stD = A + E

[0008] Among them, ||A|| *Let |E|| denote the nuclear norm (the sum of singular values ​​of E), and |E||1 denote the 1-norm (the sum of the absolute values ​​of all terms in E). However, a major drawback of RPCA is that it can only handle two-dimensional data. For example, grayscale images; however, grayscale video is indexed by two spatial variables and one temporal variable, making it three-dimensional data. Therefore, to use RPCA for the video we want to process, the multi-channel data must first be reconstructed into a matrix. Such preprocessing often leads to information loss and performance degradation. Summary of the Invention

[0009] This invention aims to address the shortcomings of existing technologies by proposing a non-contact heart rate detection method that removes irregular motion artifacts. This method is intended to improve the accuracy of video heart rate detection in irregular motion scenarios, thereby providing a new direction for non-contact heart rate detection research and promoting the application of non-contact detection technology in everyday life.

[0010] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0011] The non-contact heart rate detection method for removing motion artifacts in this invention is characterized by the following steps:

[0012] Step 1: Obtain the low-rank video image frame sequence;

[0013] Step 1.1: Use a camera device to acquire an I-frame of facial image video, denoted as F = [f1, f2, ..., f i ,…,f I ], where f i Let i represent the face image of the i-th frame, i = 1, 2, ..., I, and let the width of each face image be W and the height be H;

[0014] Step 1.2: Construct tensors using equation (1) Where n1 = I, representing the number of frames in the face image video F; n2 = W × H, representing the number of pixels in a single frame of the face image; and n3 = 3, representing the number of RGB channels.

[0015]

[0016] In equation (1), represents a low-rank face image video, and ε is the sparse noise tensor;

[0017] Step 1.3: Solve for the tensor Medium and low rank facial image video in, This represents the low-rank face image of the i-th frame, where i = 1, 2, ..., E;

[0018] Step 2: Filter the highest quality regions of interest;

[0019] Step 2.1: Locating and tracking the region of interest:

[0020] The facial feature point detection and tracking method is used to detect and track facial feature points in I-frame face video images F, thereby locating the bounding rectangles of M regions of interest in each frame of face image, i = 1, 2, ..., I, M ≥ 2;

[0021] Step 2.2: Extraction of the time series of pixel mean values ​​in the region of interest:

[0022] Calculate the low-rank face image of the i-th frame. The average pixel value of each region of interest is used to obtain a low-rank face image video. The pixel mean of each sub-region of interest is used to form a time series of pixel mean values ​​for each region of interest in an I-frame low-rank face image;

[0023] Let the pixel time series of the m-th region of interest in the green channel of the I-frame low-rank face image be denoted as... in, This represents the m-th region of interest in the low-rank face image of the i-th frame. The average pixel value of the green channel;

[0024] Step 2.3: Calculate G [m] The signal-to-noise ratio index is used to obtain [G] [1] G [2] …,G [m] ,…,G [M] The signal-to-noise ratio (SNR) of [G] is then sorted in descending order to select the [G] index. [1] G [2] ...,G [m] ,…,G [M] The region of interest with the highest signal-to-noise ratio in the i-th frame is taken as the low-rank face image. The highest quality region of interest is selected, and the low-rank face image of the i-th frame is selected. Let X be the pixel mean sequence of the highest quality region of interest. i Thus, the pixel mean sequence of the highest quality region of interest in the I-frame low-rank face image is obtained, denoted as X = [X1, X2, ..., X]. i ,…,X I ];

[0025] Step 3: Obtain the target blood volume pulse signal:

[0026] Step 3.1: Decompose X using the fully adaptive noise ensemble empirical mode decomposition method to obtain L intrinsic mode components S = [S1, S2, ..., S...] to reflect the heart rate.l ,…S L ], S l S represents the l-th intrinsic mode component. l =[S l,1 ,S l,2 ,…,S l,i ,…S l,I ], S l,i This represents the i-th low-rank face image in the l-th intrinsic mode component. The average pixel value on the surface;

[0027] Step 3.2: Define the candidate intrinsic mode component set S sad and initialize S sad Empty; initialize l = 1;

[0028] Step 3.3: Calculate the l-th intrinsic mode component S l The spectrum diagram is analyzed, and it is determined whether the frequency corresponding to the largest amplitude in the l-th spectrum diagram is within the set frequency range [LPF, HPF] of the blood volume pulse signal. If it is, the l-th intrinsic mode component S is... l Add candidate intrinsic mode component set S sad If the cutoff frequency is positive, proceed directly to step 3.4; where LPF represents the lowest cutoff frequency and HPF represents the highest cutoff frequency.

[0029] Step 3.4: Assign l+1 to l; and return to step 3.3 until l>L, thus obtaining the candidate intrinsic mode component set S. sad ;

[0030] Step 3.5: Compare the candidate intrinsic mode component sets S sad The maximum amplitude corresponding to the frequency of each intrinsic mode component is determined, and the intrinsic mode component corresponding to the maximum value is selected as the target blood volume pulse signal, denoted as L=[L1,2,…,L…]. i ,…,L I ]; where L i This represents the i-th low-rank face image in the intrinsic mode components corresponding to the maximum value. The average pixel value on the surface;

[0031] Step 4: Heart rate calculation;

[0032] After filtering L, the power spectral density distribution of the filtered target blood volume pulse signal is calculated using Fast Fourier Transform. The source component with the most frequencies falling within the set frequency range [LPF, HPF] in the power spectral density distribution is identified as the heart rate signal, and the dominant frequency f of the heart rate signal is then calculated. max And obtain the heart rate value.

[0033] The non-contact heart rate detection method for removing motion artifacts described in this invention is also characterized in that step 1.3 includes:

[0034] Step 1.3.1: Construct the objective function using equation (2):

[0035]

[0036] In equation (2), It is a low-rank facial image video. The nuclear norm of ε, ||ε||1 is the sparse noise tensor ε. Norm, These are the weighting coefficients;

[0037] Step 1.3.2: Construct the augmented Lagrangian function using equation (3)

[0038]

[0039] In equation (3), μ is a positive parameter; Represents the Lagrange multipliers. express and The inner product, Denotes the square of the Frobenius norm;

[0040] Step 1.3.3: Define the current iteration number as k and initialize k = 0; define and initialize the Lagrange multipliers in the k-th iteration. Low-rank face image video in the kth iteration The sparse noise tensor ε under the k-th iteration k The positive parameter μ in the k-th iteration k ;

[0041] Step 1.3.4: Calculate the low-rank face image video in the (k+1)th iteration using equation (4).

[0042]

[0043] Step 1.3.5: Calculate the sparse noise tensor ε in the (k+1)th iteration using equation (5). k+1 :

[0044]

[0045] Step 1.3.6: Calculate the Lagrange multipliers in the (k+1)th iteration using equation (6).

[0046]

[0047] Step 1.3.7: Calculate the positive parameter μ in the (k+1)th iteration using equation (7). k+1 :

[0048] μ k+1 =min{μ min ,ρ×μ k} (7)

[0049] In equation (7), ρ represents the relational parameter;

[0050] Step 1.3.8: Assign k+1 to k, and return to step 1.3.4 to continue execution until... And ||ε k+1 -ε k || ∞ ≤e, and Until then, thus making the final As Where e represents the threshold.

[0051] An electronic device is invented, comprising a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the non-contact heart rate detection method, and the processor is configured to execute the program stored in the memory.

[0052] The invention relates to a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the non-contact heart rate detection method.

[0053] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0054] 1. Unlike traditional rPPG technology which mainly focuses on denoising at the signal level, this invention combines both image and signal levels to denoise rPPG signals, resulting in better performance.

[0055] 2. In step one of this invention, when dealing with noise issues such as irregular motion or sudden and drastic changes in illumination during rPPG detection, unlike traditional rPPG techniques which primarily denoise at the signal level, this invention directly denoises at the image level by tensorizing the video image sequence. This ensures that the rPPG signal extracted from the processed video contains only low-rank signals. The objective function can be solved by constructing the augmented Lagrangian function using a polynomial-time algorithm, such as the Alternating Direction Method of Multipliers (ADMM), without requiring parameter adjustments to guarantee tensor decomposition. This saves the effort and time spent on parameter adjustments and ensures the performance of the algorithm.

[0056] 3. Step 2.5 of this invention, the fully adaptive noise set empirical mode decomposition method, decomposes the non-stationary time series into a finite number of intrinsic mode components, thereby further eliminating the influence of other noises and possible residual noise from ambient light changes, and obtaining accurate video heart rate detection results. Attached Figure Description

[0057] Figure 1 This is a flowchart of the present invention;

[0058] Figure 2 This is a diagram showing the size of a frame extracted from the original video and a diagram showing the division of the region of interest.

[0059] Figure 3 The rPPG signals before and after processing in step one;

[0060] Figure 4 The intrinsic mode components S1 to S9 generated by performing EEMD on the green channel in this invention, and the spectrum obtained by performing FFT transformation on each intrinsic mode component, are respectively the spectrum of intrinsic mode components S1 to S9 from top to bottom. Detailed Implementation

[0061] In this embodiment, a method for remote photoplethysmography (PPG) heart rate detection to remove irregular motion artifacts involves: first, capturing a low-rank video image sequence; second, tensing the original video frame sequence and constructing an augmented Lagrangian function to decompose the low-rank and sparse components from the tensor data, retaining the low-rank signal video frame sequence; third, extracting the pixel mean time series of the optimal color channel, mainly including locating, dividing, and filtering the region of interest, tracking the region of interest, and extracting the pixel mean time series; and fourth, using a fully adaptive noise set empirical mode decomposition method to further denoise the pixel mean time series from a signal perspective; finally, calculating the heart rate by performing spectral analysis on the pulse signal and then performing corresponding calculations to obtain the heart rate value. By combining video image processing and signal analysis, robustness to changes in illumination and irregular motion is improved.

[0062] To verify the noise reduction effect of the present invention experimentally, in this embodiment, a Logitech C920 RGB camera was used to record video at a frame rate of 30fps. The invention will be further explained in conjunction with the accompanying drawings and specific implementation details. Figure 1 As shown, the method includes:

[0063] Step 1: Obtain the low-rank video image frame sequence;

[0064] Step 1.1: Use a camera device to acquire an I-frame of a face image, and denote it as F = [f1, f2, ..., f i ,…,f I ], f i This represents the acquired i-th frame of the face image, where i = 1, 2, ..., I, and the image size of a single frame is W (width) and H (height). Figure 2 The image shows a schematic diagram of the size of one frame extracted from the original video.

[0065] Step 1.2: Obtain the tensor using equation (1) Where n1 = I, representing the number of frames in the face image video F; n2 = W × H, representing the number of pixels in a single frame of the face image; and n3 = 3, representing the number of RGB channels.

[0066]

[0067] In equation (1), ε represents a low-rank face image video, where ε is the sparse noise tensor.

[0068] Step 1.3: Solve for the tensor Medium and low rank facial image video in, Let i represent the low-rank face image of the i-th frame, where i = 1, 2, ..., I.

[0069] Step 1.3.1: Construct the function:

[0070]

[0071] In equation (2), It is a tensor The low-rank part of the middle, It is a low-rank tensor The nuclear norm, ε is the tensor The sparse part in the sparse noise tensor ε, ||ε||1 is the sparse noise tensor ε Norm, These are the weighting coefficients;

[0072] Step 1.3.2: Solve the optimization problem in equation (2) using the alternating direction multiplier method by introducing Lagrange multipliers. The quadratic penalty term can be used to construct the augmented Lagrangian function using equation (3).

[0073]

[0074] In equation (3), μ is a positive parameter; Represents the Lagrange multipliers. express and The inner product, This represents the square of the Frobenius norm;

[0075] Step 1.3.3: Define the current iteration number as k and initialize k = 0; define and initialize the Lagrange multipliers in the k-th iteration. Low-rank face image video in the kth iteration The sparse noise tensor ε under the k-th iteration k The positive parameter μ in the k-th iteration k ;initialization

[0076] Step 1.3.4: Calculate the low-rank face image video in the (k+1)th iteration using equation (4).

[0077]

[0078] Step 1.3.5: Calculate the sparse noise tensor ε in the (k+1)th iteration using equation (5). k+1 :

[0079]

[0080] Step 1.3.6: Calculate the Lagrange multipliers in the (k+1)th iteration using equation (6).

[0081]

[0082] Step 1.3.7: Calculate the positive parameter μ in the (k+1)th iteration using equation (7). k+1 :

[0083] μ k+1 =min{μ min ,ρ×μ k} (7)

[0084] In equation (7), ρ represents the parameter used to balance the relationship between the primal problem and the dual problem. In this embodiment, ρ = 1.1.

[0085] Step 1.3.8: Assign k+1 to k, and return to step 1.3.4 to continue execution until... And ||ε k+1 -ε k || ∞ ≤e, and Until then, thus making the final As Where e represents the threshold, in this example, e = 10. -8 ,like Figure 3 As shown, this example illustrates the signal state of a data segment before and after processing in step one.

[0086] Step 2: Extract the pixel mean time series of the optimal color channel;

[0087] Step 2.1: Locating and tracking the region of interest:

[0088] A facial feature point detection and tracking method is used to detect and track facial feature points in I-frame face video images F, thereby locating the bounding rectangles of M regions of interest in each frame of face image. Let W be the width and height of the bounding rectangle of the m-th region of interest in each frame of face image. m H m i = 1, 2, ..., I, M ≥ 2; This example uses the convolutional expert-constrained local model provided by the OpenFace toolbox to detect and track facial feature points in I-frame video images. During movement, the region of interest (ROI) will change in area and position as the subject moves back, forth, left, and right. To ensure the stability of the ROI, this embodiment selects relatively stable feature points from OpenFace's feature points to locate the bounding rectangle of the ROI in each frame of video images. The schematic diagram of the ROI division is shown below. Figure 2 As shown, these are the forehead and left cheek, respectively, and the number of regions of interest M = 2.

[0089] Step 2.2: Extraction of the time series of pixel mean values ​​in the region of interest:

[0090] Calculate the low-rank face image of the i-th frame. The average pixel value of each region of interest is used to obtain a low-rank face image video. The pixel mean of each sub-region of interest is calculated, and a time series of the pixel mean of each region of interest in the I-frame low-rank face image is formed. In rPPG heart rate detection calculation, the green channel is selected as the best choice. This is because green light has a longer wavelength, resulting in a better signal and a better signal-to-noise ratio than other light sources. In practical research and applications, the green channel contains the strongest heart rate signal, which is consistent with the PPG principle that green light can obtain a better signal. In addition, research shows that among the RGB three channels, the green channel has the highest correlation with the heart rate signal, and using the green channel can improve the accuracy and robustness of remote heart rate detection. Therefore, this invention selects the optimal color channel as the green channel signal as the signal for screening the highest quality region of interest. The pixel time series of the green channel of the m-th region of interest in the I-frame low-rank face image is denoted as [the green channel signal is then used to select the highest quality region of interest]. in, This represents the m-th region of interest in the low-rank face image of the i-th frame. The average pixel value of the green channel.

[0091] Step 2.3: Calculate G [m] The signal-to-noise ratio index is used to obtain [G] [1] G [2] …,G [m] ,…,G [M] The signal-to-noise ratio (SNR) of [G] is then sorted in descending order to select the [G] index. [1] G [2] …,G [m] ,…,G [M] The region of interest with the highest signal-to-noise ratio in the i-th frame is taken as the low-rank face image. The highest quality region of interest is selected, and the low-rank face image of the i-th frame is selected. Let X be the pixel mean sequence of the highest quality region of interest. i Thus, the pixel mean sequence of the highest quality region of interest in the I-frame low-rank face image is obtained, denoted as X = [X1, X2, ..., X]. i ,…,X I In this embodiment, the signal-to-noise ratio of the sub-region of interest in the forehead region is the highest, so the set of pixel mean values ​​of the green channel in the forehead region is selected for subsequent processing.

[0092] Step 3: Obtain the target blood volume pulse signal:

[0093] Step 3.1: Decompose X using the fully adaptive noise ensemble empirical mode decomposition method to obtain L intrinsic mode components S = [S1, S2, ..., S...] to reflect the heart rate. l ,…S L ], S l S represents the l-th intrinsic mode component. l =[S l,1 ,S l,2 ,…,S l,i ,…S l,I ], S l,i This represents the i-th low-rank face image in the l-th intrinsic mode component. The average pixel value; in this embodiment, based on the signal length, the number of intrinsic mode components is 9, and the specific form of each component is as follows. Figure 4 As shown.

[0094] Step 3.2: Define the candidate intrinsic mode component set S sad and initialize S sad Empty; initialize l = 1;

[0095] Step 3.3: Calculate the l-th intrinsic mode component S l The spectrum diagram is analyzed, and it is determined whether the frequency corresponding to the largest amplitude in the l-th spectrum diagram is within the set frequency range [LPF, HPF] of the blood volume pulse signal. If it is, the l-th intrinsic mode component S is... l Add candidate intrinsic mode component set S sad If the cutoff frequency is not specified, proceed directly to step 3.4; where LPF represents the lowest cutoff frequency and HPF represents the highest cutoff frequency.

[0096] Step 3.4: Assign l+1 to l; and return to step 2.7 until l>L, thus obtaining the candidate intrinsic mode component set S. sad In this embodiment, the normal heart rate range is set to [0.7, 2.5] Hz, corresponding to a heart rate of 42 beats per minute (bpm) to 150 bpm. Figure 4 It is not difficult to find that the frequencies corresponding to the maximum amplitudes of the intrinsic mode components S1 and S2 are within the normal heart rate range. However, in the fully adaptive noise set empirical mode decomposition algorithm, S1 is the same as the original data, so only S2 constitutes the candidate intrinsic mode component set.

[0097] Step 3.5: Compare the candidate intrinsic mode component sets S sad The maximum amplitude corresponding to the frequency of each intrinsic mode component is determined, and the intrinsic mode component corresponding to the maximum value is selected as the target blood volume pulse signal, denoted as L=[L1,2,…,L…]. i ,…,LI ]; where L i This represents the i-th low-rank face image in the intrinsic mode components corresponding to the maximum value. The average pixel value.

[0098] Step 4: Heart rate calculation;

[0099] After filtering L using a Butterworth filter, the power spectral density distribution of the filtered target blood volume pulse signal is calculated using a Fast Fourier Transform. The source component with the most frequencies falling within the set frequency range [LPF, HPF] in the power spectral density distribution is identified as the heart rate signal, and the dominant frequency f of the heart rate signal is then calculated. max And obtain the heart rate value. Thus, the estimated heart rate value is time × f. max , where time represents 1 minute.

[0100] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0101] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0102] To verify the effectiveness of the method of this invention, its performance was validated on the publicly available databases UBFC-RPPG and ECG-Fitness. This invention uses conventional evaluation metrics to assess heart rate detection performance, specifically including mean absolute error (HR). mae (mean Absolute Error, MAE), Root Mean Square Error (HR) rmse (root mean square error, RMSE) and Pearson's Correlation Coefficient (r).

[0103] The UBFC-RPPG database used videos of subjects in a resting state, with a sampling rate of 30 frames per second and PPG heart rate reference signal. The ECG-Fitness database consists of 204 fitness-themed videos, each 1 minute long. ECG signals were simultaneously acquired using a dual-lead electrocardiograph. This database recruited 17 subjects (aged 20 to 53 years) who performed four activities (talking, rowing, stationary cycling, and elliptical training) under three lighting settings, with each video lasting 1 minute. This database was used to validate the applicability of this method in real-world scenarios.

[0104] Table 1. Results of the comparative experiment using the UBFC-RPPG database

[0105]

[0106] Table 2 Comparison of heart rate measurement results in speaking scenarios using the ECG-Fitness database.

[0107]

[0108] Table 1 presents the heart rate detection results of the proposed method and several other methods on the UBFC-RPPG database. As can be seen from Table 1, the LFT-CEEMDAN method proposed in this invention achieves the best results in all indicators compared to the other four methods, with HR... mae Reduced to 2.91 bpm, HR rmse The heart rate was reduced to 4.42 bpm, and the r-value increased to 0.97. Under static or minimally dynamic conditions, the algorithm combining LTF and EEMD demonstrates strong separation of regular, periodic heart rate signals, enabling it to extract heart rate signals more accurately than model-based methods and traditional blind source separation methods. Overall, the proposed method achieves near-best results on the UBFC-RPPG database, indicating its superior performance on the relatively static UBFC-RPPG database.

[0109] Table 2 presents the experimental results of the proposed method and several other methods in a speaking scenario using the ECG-Fitness database. As can be seen from Table 2, the LFT-CEEMDAN method proposed in this invention achieves the best results across all metrics compared to the other four methods, with HR... mae Reduced to 8.95 bpm, HR rmse The error was reduced to 9.82 bpm, and r was improved to 0.71. Overall, the experimental results demonstrate the superiority of the proposed method in removing irregular motion artifacts, i.e., in real-world scenarios.

[0110] In summary, experiments have demonstrated that the method proposed in this invention is more suitable for heart rate detection under irregular motion conditions. The method proposed in this invention, which integrates low-rank tensor decomposition and fully adaptive noise set empirical mode decomposition, improves the accuracy of video heart rate detection in scenarios with irregular motion and changing illumination, and exhibits good motion robustness.

Claims

1. A non-contact heart rate detection method for removing motion artifacts, characterized in that, The procedure is as follows: Step 1: Obtain the low-rank video image frame sequence; Step 1.1: Obtain using camera equipment A frame of facial image video, denoted as ,in, Indicates the first Frame of face images, Let the width of each frame of the face image be W and the height be H; Step 1.2: Construct tensors using equation (1) ,in, , indicating facial image video F Frame rate This represents the number of pixels in a single frame of a face image. 3 indicates the number of RGB channels; (1) In equation (1), It represents low-rank facial image video. For sparse noise tensors; Step 1.3: Solve for the tensor Medium and low rank facial image video ,in, Indicates the first Low-rank face images, ; Step 1.3.1: Construct the objective function using equation (2), and use the alternating multiplier method to construct the augmented Lagrangian function to solve for the objective function: (2) In equation (2), It is a low-rank facial image video. nuclear norm number, It is a sparse noise tensor of Norm, These are the weighting coefficients; Step 2: Filter the highest quality regions of interest; Step 2.1: Locating and tracking the region of interest: Facial landmark detection and tracking methods are used to analyze I-frame face video images. F Facial feature points are detected and tracked to locate the bounding rectangles of M regions of interest in each frame of the face image. M≥2; Step 2.2: Extraction of the time series of pixel mean values ​​in the region of interest: Calculate the first Low-rank face images The average pixel value of each region of interest is used to obtain a low-rank face image video. The pixel mean of each sub-region of interest is used to form a time series of pixel mean values ​​for each region of interest in an I-frame low-rank face image; The first The pixel time series of the green channel of a region of interest in an I-frame low-rank face image is denoted as follows: ;in, Indicates the first The region of interest is in the first Low-rank face images The average pixel value of the green channel; Step 2.3: Calculation The signal-to-noise ratio index is thus obtained. The signal-to-noise ratio (SNR) metrics are then sorted in descending order to select the appropriate value. The region of interest with the highest signal-to-noise ratio is taken as the first... Low-rank face images The highest quality region of interest, and the first Low-rank face images The pixel mean sequence of the highest quality region of interest is denoted as Thus, the pixel mean sequence of the highest quality region of interest in the I-frame low-rank face image is obtained, denoted as X = ; Step 3: Obtain the target blood volume pulse signal: Step 3.1: Use the fully adaptive noise ensemble empirical mode decomposition method to... X The decomposition yields a value used to reflect heart rate. Each intrinsic mode component , Indicates the first Each intrinsic mode component , Indicates the first The th eigenmode component Low-rank face images The average pixel value on the surface; Step 3.2: Define the candidate intrinsic mode component set and initialize Empty; initialize ; Step 3.3: Calculate the first... Each intrinsic mode component The spectrum diagram, and determine the first Does the frequency corresponding to the largest amplitude in each spectrum fall within the frequency range of the set blood volume pulse signal? If inside, then the first Each intrinsic mode component Add to candidate intrinsic mode component set Otherwise, proceed directly to step 3.4; where, Indicates the lowest cutoff frequency. Indicates the highest cutoff frequency; Step 3.4: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require Assign to ; and return to step 3.3 until Thus, the candidate intrinsic mode component set is obtained. ; Step 3.5: Compare candidate intrinsic mode component sets The maximum amplitude value corresponding to the frequency of each intrinsic mode component is determined, and the intrinsic mode component corresponding to the maximum value is selected as the target blood volume pulse signal, and denoted as... ;in, This represents the eigenmode component corresponding to the maximum value. Low-rank face images The average pixel value on the surface; Step 4: Heart rate calculation; Using filters After filtering, the power spectral density distribution of the filtered target blood volume pulse signal is calculated using Fast Fourier Transform, and the frequencies in the power spectral density distribution are within a set frequency range. The signal with the most abundant source component was determined to be the heart rate signal, and the dominant frequency of the heart rate signal was then calculated. And obtain the heart rate value.

2. The non-contact heart rate detection method for removing motion artifacts according to claim 1, characterized in that, Step 1.3 also includes: Step 1.3.2: Construct the augmented Lagrangian function using equation (3) : (3) In equation (3), It is a positive parameter; Represents the Lagrange multipliers. express and The inner product, Denotes the square of the Frobenius norm; Step 1.3.3: Define the current iteration number as... and initialize =0; Define and initialize the first k Lagrange multipliers in the next iteration , No. k Low-rank face image video in the next iteration , No. k Sparse noise tensor under the next iteration No. k Positive parameters in the next iteration ; Step 1.3.4: Calculate the first step using equation (4). Low-rank face image video after +1 iteration : (4) Step 1.3.5: Calculate the first step using equation (5). Sparse noise tensor under the next iteration : (5) Step 1.3.6: Calculate the first step using equation (6). Lagrange multipliers in the next iteration : (6) Step 1.3.7: Calculate the first step using equation (7). Positive parameters in the next iteration : (7) In equation (7), Indicates relational parameters; Step 1.3.8: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Assign to And return to step 1.3.4 to execute until... Until then, thus making the final As ,in, This represents the threshold.

3. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store a program that supports the processor in executing the non-contact heart rate detection method of claim 1 or 2, and the processor is configured to execute the program stored in the memory.

4. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program, when run by a processor, executes the steps of the non-contact heart rate detection method according to claim 1 or 2.

Images