Remote physiological signal detection method based on adaptive filter, terminal and medium

Abstract

The present application relates to the technical field of non-contact physiological signal, and discloses a remote physiological signal detection method based on an adaptive filter, a terminal and a medium.The method acquires a face video to be detected, and obtains a noise rPPG signal through noise processing; feature extraction is performed on the face video to obtain a multi-scale space-time graph; the noise rPPG signal and the multi-scale space-time graph are fused to obtain fused data; the fused data is denoised by using a denoising module to obtain denoised data; the denoising module is an adaptive filter denoising module, which is composed of multiple layers of filter structures inside, and each layer of filter structure includes a fixed spectrum shaping filter, a scene adaptive shaping filter and a spectrum period feature extraction filter connected in sequence; and a predicted rPPG signal is output according to the denoised data.The present application uses an adaptive filter to deeply fuse and denoise noise and multi-scale space-time feature data, and realizes high-precision rPPG signal estimation.

Description

Technical Field

[0001] This invention relates to the field of non-contact physiological signal detection technology, specifically to a remote physiological signal detection method, terminal, and medium based on adaptive filters. Background Technology

[0002] Currently, physiological signal detection technologies based on facial videos mainly include the following different model architectures:

[0003] 1. Non-end-to-end model: This type of model typically requires complex feature preprocessing steps to extract information from multiple color channels such as RGB and YUV, and improves the accuracy of rPPG estimation through feature decoupling strategies. Although this type of model ensures accurate signal extraction to a certain extent, it suffers from cumbersome preprocessing, poor real-time performance, and limited noise suppression capabilities.

[0004] 2. End-to-end model-based: In recent years, end-to-end models (such as the video Transformer architecture) have gradually emerged. These models can adaptively aggregate local and global spatiotemporal features, thereby reducing intermediate preprocessing steps and enhancing the representation of rPPG signals. However, these models usually have a large computational load and high hardware performance requirements, making it difficult to achieve efficient operation in resource-constrained environments.

[0005] 3. Unsupervised models based on contrastive learning: These models utilize spatial similarity between different regions and temporal similarity of rPPG signals within a short timeframe for feature learning, reducing reliance on labeled data. However, their model stability and detection accuracy still need improvement.

[0006] While existing non-contact physiological signal detection technologies have made some progress in practical applications, some shortcomings still exist. First, due to external interference such as changes in lighting, facial occlusion (e.g., hair, beard, glasses), and head movements, the accuracy of signal detection results is often difficult to guarantee. Second, optical differences caused by different races, ages, and skin characteristics limit the generalization ability of the models. In addition, although some end-to-end models based on deep networks such as 3D CNNs can improve detection accuracy, their high computational complexity also limits their application in low-resource scenarios. Summary of the Invention

[0007] To address the technical problems existing in the prior art, this invention provides a remote physiological signal detection method, terminal, and medium based on an adaptive filter. The adaptive filter is used to perform deep fusion and denoising of noise and multi-scale spatiotemporal feature data, thereby achieving high-precision rPPG signal estimation.

[0008] To achieve the above objectives, the present invention provides the following technical solution:

[0009] This invention discloses a remote physiological signal detection method based on an adaptive filter, comprising:

[0010] Acquire the face video to be detected, and obtain the noise rPPG signal through noise processing;

[0011] Feature extraction is performed on the face video to obtain a multi-scale spatiotemporal map;

[0012] The noisy rPPG signal and the multi-scale spatiotemporal map are fused to obtain fused data;

[0013] The fused data is denoised using a denoising module to obtain denoised data. This denoising module employs an adaptive filter denoising module, internally composed of a multi-layer filter structure. Each layer includes a fixed-spectrum shaping filter, a context-adaptive shaping filter, and a spectral periodicity feature extraction filter connected sequentially. The fixed-spectrum shaping filter performs initial frequency domain filtering using randomly initialized fixed parameters. The context-adaptive shaping filter uses the output of the fixed-spectrum shaping filter to generate data-dependent filtering parameters, achieving context-adaptive frequency domain filtering to reconstruct the periodicity and error components in the time series. The spectral periodicity feature extraction filter performs frequency domain reconstruction on the output of the context-adaptive shaping filter to recover the periodicity and error components in the time series.

[0014] Based on the denoised data, the predicted rPPG signal is output.

[0015] As a further improvement to the above scheme, the processing procedure of the fixed spectrum shaping filter includes:

[0016] Perform a Fourier transform on the input fused data, i.e., the time series T, to obtain the frequency domain representation of the time series.

[0017]

[0018] In the formula, Fourier transform processing;

[0019] Using pre-randomly initialized and fixed filter parameters H φ Frequency domain representation Element-wise multiplication yields the pre-shaped frequency domain features.

[0020]

[0021] In the formula, and These are all parameters of the channel-shared filter; ⊙L This is element-wise multiplication along the dimension of sequence length L;

[0022] right Perform an inverse Fourier transform to obtain the output T of the fixed-spectrum shaping filter. plain :

[0023]

[0024] In the formula, This represents the inverse Fourier transform.

[0025] As a further improvement to the above scheme, the processing procedure of the scene-adaptive shaping filter includes:

[0026] For the output T of the fixed spectrum shaping filter plain Perform a Fourier transform to obtain the frequency domain representation after the Fourier transform.

[0027]

[0028] Generate adaptive filter parameters using a data-dependent neural network.

[0029]

[0030] In the formula, NN(·) represents neural network processing;

[0031] Frequency domain representation With adaptive filter parameters Element-wise multiplication yields the shaped frequency domain features.

[0032]

[0033] Frequency domain features after shaping Perform an inverse Fourier transform to obtain the output T of the scene-adaptive shaping filter. ctx :

[0034]

[0035] As a further improvement to the above scheme, the processing procedure of the spectral periodic feature extraction filter includes:

[0036] The output T of the scene-adaptive shaping filter ctx Perform a Fourier transform to extract amplitude and phase information:

[0037]

[0038] In the formula, A is the amplitude, Φ is the phase, and ∠ represents the angle;

[0039] Select the top K important frequency components from amplitude A. The frequencies and phases corresponding to the K important frequency components are respectively... and Constructing a frequency domain reconstructed signal through weighted synthesis:

[0040]

[0041] In the formula, T syn For frequency domain reconstructed signal; A k Let A be the amplitude of the kth significant frequency component in amplitude A; t is a time variable; t is a time variable;

[0042] Reconstruct the frequency domain signal T syn This is the output of the entire noise reduction process.

[0043] As a further improvement to the above scheme, the expression for noise processing is as follows:

[0044]

[0045] In the formula, x0 is the original rPPG signal; x t’ Let q(x) be the noise rPPG signal after the t'th noise processing; t' = 1, 2, ..., N, where N is the total number of noise processing iterations; t’ ||x0) represents the state x given an initial state x0. t’ The conditional probability distribution; As the attenuation factor, The value gradually decreases as the number of processing iterations t' increases during the processing; ∈ represents noise.

[0046] As a further improvement to the above scheme, the specific process of feature extraction for facial features includes:

[0047] The face video is divided into multiple segments, and each segment yields F frames of video images.

[0048] Face detection is performed on video images, and the face is divided into n regions of interest (ROIs).

[0049] Combining n ROIs yields two combinations: single ROIs and combinations of different ROIs. n -1 combination;

[0050] For each combination, calculate the average pixel value of all ROIs in the combination across each color channel; there are a total of 6 color channels, namely RGB and YUV;

[0051] Finally, the dimension is 6×(2) n The vector of -1)×F is used as a multi-scale spatiotemporal graph.

[0052] As a further improvement to the above scheme, the specific process of fusing the noisy rPPG signal and the multi-scale spatiotemporal map includes:

[0053] The dimension of the noise rPPG signal is 1×F. Expanding the dimension of the noise rPPG signal to 1×(2F) will increase the dimension of the noise rPPG signal to 1×(2F). n -1)×F, the expanded noisy rPPG signal is stitched with the multi-scale spatiotemporal map along the channel dimension to obtain fused data, the dimension of which is 7×(2). n -1)×F.

[0054] As a further improvement to the above scheme, the noise rPPG signal is Gaussian noise that follows a standard normal distribution.

[0055] The present invention also discloses a remote physiological signal detection system based on an adaptive filter, which is applied to the remote physiological signal detection method based on an adaptive filter as described above; the detection system includes: a noise processing module, an MSTmap calculation module, a fusion module and a denoising module.

[0056] The noise processing module is used to acquire the face video to be detected and obtain the noise rPPG signal through noise processing;

[0057] The MSTmap calculation module is used to extract features from face videos to obtain multi-scale spatiotemporal maps;

[0058] The fusion module is used to fuse the noisy rPPG signal and the multi-scale spatiotemporal map to obtain fused data;

[0059] The denoising module is used to denoise the fused data to obtain denoised data. This denoising module employs an adaptive filter denoising module, internally composed of a multi-layer filter structure. Each layer includes a fixed-spectrum shaping filter, a context-adaptive shaping filter, and a spectral periodic feature extraction filter connected in sequence. The fixed-spectrum shaping filter performs initial frequency domain filtering using randomly initialized fixed parameters. The context-adaptive shaping filter uses the output of the fixed-spectrum shaping filter to generate data-dependent filtering parameters, achieving context-adaptive frequency domain filtering, thereby reconstructing the periodicity and error components in the time series. The spectral periodic feature extraction filter performs frequency domain reconstruction on the output of the context-adaptive shaping filter to recover the periodicity and error components in the time series.

[0060] The prediction module is used to output a predicted rPPG signal based on the denoised data.

[0061] The present invention also discloses a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the remote physiological signal detection method based on an adaptive filter as described above.

[0062] Compared with the prior art, the beneficial effects of the present invention are:

[0063] 1. This invention utilizes adaptive filter technology, employing a multi-layer structure (including a fixed spectrum shaping filter, a scene-adaptive shaping filter, and a spectrum periodic feature extraction filter) to achieve layered denoising of noisy rPPG signals and accurate reconstruction of real signals, significantly improving the model's generalization ability. The adaptive filter fully leverages the periodicity and spectral characteristics of the rPPG signal during the reconstruction process, making the final output physiological signal more closely resemble the real human body state. Through the collaborative work of multiple filters, this invention can effectively cope with external interference such as changes in lighting, facial occlusion, and head movement, ensuring the stability and robustness of the detection results and achieving high-precision rPPG signal estimation.

[0064] 2. Compared to traditional deep convolutional networks or end-to-end Transformer models, adaptive filters use layer-by-layer denoising and feature fusion to reduce computational complexity while maintaining high-precision signal detection, making them suitable for resource-constrained real-time scenarios.

[0065] 3. Extracting effective region features using multi-scale spatiotemporal maps (MSTmap): During MSTmap calculation, a face detection tool is used to segment the effective face region in the video, allowing the model to focus on high-information areas, thereby significantly improving the accuracy of signal estimation. MSTmap calculation not only uses the original RGB channels but also introduces additional YUV channels, enriching visual feature information and providing a more comprehensive basis for model training and inference. By arranging and combining multiple regions of interest, MSTmap can display richer spatial information, enabling the subsequent adaptive filter module to extract more effective features during denoising, further improving detection accuracy.

[0066] 4. This invention uses a splicing method to fuse one-dimensional noise rPPG signals with multi-channel MSTmaps, maintaining the independence of their respective characteristics. This not only learns the distribution characteristics of the rPPG signals during the noise restoration process, but also guides the mapping of the MSTmaps to the real signals, avoiding mutual interference. Attached Figure Description

[0067] Figure 1 This is a framework diagram of the remote physiological signal detection system based on an adaptive filter in Embodiment 1 of the present invention.

[0068] Figure 2 This is a flowchart of the remote physiological signal detection method based on an adaptive filter according to Embodiment 1 of the present invention.

[0069] Figure 3 This is a flowchart of the training and estimation processes of the adaptive filter model in Embodiment 1 of the present invention.

[0070] Figure 4 This is a schematic diagram of the structure of the computer terminal in Embodiment 2 of the present invention. Detailed Implementation

[0071] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0072] Example 1

[0073] Please see Figure 1 This embodiment provides a remote physiological signal detection system based on an adaptive filter, including: a noise processing module, an MSTmap (Multi-scale Stmap) calculation module, a fusion module, a denoising module, a prediction module, and a sampling module. The entire system's operation is divided into two parts: a training process and an estimation process.

[0074] During training:

[0075] The noise processing module is used to add noise to the original rPPG signal to obtain a noisy rPPG signal.

[0076] The MSTmap calculation module is used to process face videos to obtain multi-scale spatiotemporal maps containing visual features;

[0077] The fusion module is used to fuse noisy rPPG signals with multi-scale spatiotemporal maps to obtain fused data;

[0078] The denoising module is used to denoise the fused data to obtain denoised data;

[0079] The prediction module is used to predict the rPPG signal based on the denoised data output.

[0080] During the estimation process:

[0081] The sampling module is used to randomly sample and generate noisy rPPG signals;

[0082] The MSTmap calculation module is used to process the face video to be estimated and obtain a multi-scale spatiotemporal map containing visual features.

[0083] The fusion module is used to fuse noisy rPPG signals with multi-scale spatiotemporal maps to obtain fused data;

[0084] The denoising module is used to denoise the fused data to obtain denoised data;

[0085] The prediction module is used to predict the rPPG signal based on the denoised data output.

[0086] Depend on Figure 2 and Figure 3 As shown in the figure, this embodiment also provides a remote physiological signal detection method based on an adaptive filter. First, the collected sample data is input to train the adaptive filter model during the training process. Then, the trained adaptive filter model is used to estimate the remote physiological signal. During the estimation process, the input is the face video of the person to be detected, and the output is the corresponding predicted rPPG signal.

[0087] The training process of the adaptive filter model is as follows:

[0088] Step 1.1: Obtain the sample set, which includes face videos and the corresponding raw rPPG signals;

[0089] Step 1.2 involves adding noise to the original rPPG signal in the noise processing module. This transforms the ordered original rPPG signal into a disordered noisy rPPG signal, gradually transitioning from an initial state with almost no noise to a state primarily composed of noise. This process provides a rich dataset for model training, including all intermediate states from noise-free to highly noisy, which is crucial for the subsequent denoising process as it relies on these intermediate states to learn how to gradually restore the original noise-free state. The mathematical expression for noise processing is as follows:

[0090]

[0091] Where x0 represents the original rPPG signal; x t’ Let x represent the noise rPPG signal after the t'th noise processing step; t' = 1, 2, ..., N, and finally obtain the noise rPPG signal x after N noise processing steps. N N represents the total number of noise processing iterations; in this embodiment, N = 1000; q(x t’ ||x0) represents the state x given an initial state x0. t’ The conditional probability distribution is used to describe the interval from x0 to x. t’ The evolutionary process; It is a predefined attenuation factor. It gradually decreases during the processing. The proportion of the original rPPG signal retained in each noise processing step is controlled; ∈ represents random noise drawn from the standard normal distribution N(0,I); The proportion of noise in each noise processing step was determined, and as the number of noise processing steps t' increases, Gradually increase.

[0092] Step 1.3: Input the face video corresponding to the original rPPG signal into the MSTmap calculation module to calculate the multi-scale spatiotemporal map. The specific method is as follows:

[0093] The input face video is divided into segments, with each segment consisting of F frames and 10 frames between each segment. Each segment yields an F-frame video image. In this embodiment, F = 200.

[0094] Face detection was performed using the facial detection tool OpenFace, and 68 facial feature points were extracted. Based on these facial feature points, six ROIs (Regions of Interest) were defined for the face: forehead, left cheek, right cheek, left triangle area, right triangle area, and chin.

[0095] Combining n ROIs yields two combinations: single ROIs and combinations of different ROIs. n -1 combinations, in this embodiment n=6.

[0096] For each combination, calculate the average pixel value of all ROIs in the combination across each color channel. There are a total of 6 color channels: RGB and YUV.

[0097] Finally, the dimension is 6×(2) n The vector of -1)×F is used as a multi-scale spatiotemporal graph. In this embodiment, the dimension of the multi-scale spatiotemporal graph is 1×63×200.

[0098] Step 1.4: Input the noisy rPPG signal obtained in step S12 and the multi-scale spatiotemporal map obtained in step S13 into the fusion module for fusion to obtain fused data.

[0099] Noise rPPG signal after noise processing N The dimension is 1×300. First, the noise rPPG signal x N The dimensions are expanded to 1×63×300. The expanded noisy rPPG signal is then concatenated with the multi-scale spatiotemporal map along the channel dimension using the concat function to obtain the fused data y. N , data y NThe dimensions are 7×63×200.

[0100] Step 1.5: Input the fused data into the denoising module for denoising processing to obtain denoised data, thereby obtaining a stable and pure (i.e., noise-free) rPPG feature vector.

[0101] The denoising module employs an MFD (Adaptive Filter Denoising Module), which consists of a c-layer filter structure. Each layer includes a fixed spectrum shaping filter, a context-adaptive shaping filter, and a spectrum periodic feature extraction filter connected in sequence. Multiple filter modules adaptively extract features and fuse multi-layer embedded features to better understand data relationships. In this embodiment, c = 6.

[0102] Fixed-spectrum shaping filters use pre-randomly initialized and fixed parameters to perform element-by-element filtering on the frequency domain representation of the input signal, directly suppressing noise and preserving key spectral features.

[0103] Context-adaptive shaping filters generate data-dependent filtering parameters based on the input signal. By adaptively adjusting the frequency domain filtering process, they achieve context-adaptive spectral shaping, thereby reconstructing the periodic and error components in the time series.

[0104] The spectral periodic feature extraction filter extracts the amplitude and phase information of the input signal through Fourier transform, selects the top K important frequency components and performs weighted synthesis, thereby realizing frequency domain reconstruction and restoring the periodicity and error components in the time series.

[0105] The mathematical expression for the denoising module during training is as follows:

[0106] z = MFD(y N )

[0107] Among them, y N For the fused data, z represents the denoised data; MFD(·) is the processing function of the adaptive filter denoising module.

[0108] MFD(y N )=Fourier(Contextual(Plain(y N )))

[0109] Where Plain(·) is the processing function of the fixed spectrum shaping filter, and the processing method is as follows:

[0110] First, perform a Fourier transform on the input fused data, i.e., the time series T, to obtain the frequency domain representation of the time series.

[0111]

[0112] In the formula, Fourier transform processing;

[0113] Then, using pre-randomized and fixed filter parameters H φ Frequency domain representation Element-wise multiplication yields the pre-shaped frequency domain features.

[0114]

[0115] In the formula, and These are all parameters of the channel-shared filter; ⊙ L This is element-wise multiplication along the dimension of sequence length L;

[0116] Finally, Perform an inverse Fourier transform to obtain the output T of the fixed-spectrum shaping filter. plain :

[0117]

[0118] In the formula, This represents the inverse Fourier transform.

[0119] Contextual(·) is the processing function for the context-adaptive shaping filter, and the processing method is as follows:

[0120] First, the output T of the fixed spectrum shaping filter plain Perform a Fourier transform to obtain the frequency domain representation after the Fourier transform.

[0121]

[0122] Then, adaptive filter parameters are generated using a data-dependent neural network.

[0123]

[0124] In the formula, NN(·) represents neural network processing;

[0125] Next, the frequency domain representation will be... With adaptive filter parameters Element-wise multiplication yields the shaped frequency domain features.

[0126]

[0127] Finally, the frequency domain features after shaping Perform an inverse Fourier transform to obtain the output T of the scene-adaptive shaping filter. ctx :

[0128]

[0129] Fourier(·) is the processing function of the spectral periodic feature extraction filter, and the processing method is as follows:

[0130] First, the output T of the scene-adaptive shaping filter... ctx Perform a Fourier transform to extract amplitude and phase information:

[0131]

[0132] In the formula, A is the amplitude, Φ is the phase, and ∠ represents the angle;

[0133] Then, the top K important frequency components are selected from the amplitude A. The frequencies and phases corresponding to the K important frequency components are respectively... and Constructing a frequency domain reconstructed signal through weighted synthesis:

[0134]

[0135] In the formula, T syn For frequency domain reconstructed signal; A k Let be the amplitude of the kth significant frequency component in amplitude A; t is the time variable;

[0136] Finally, the frequency domain reconstructed signal T syn As the output of the entire noise reduction process.

[0137] Step 1.6: Input the denoised feature vector z obtained in step 1.5 into the prediction module, and finally obtain the predicted rPPG signal.

[0138] The model is trained by minimizing the difference between the predicted rPPG signal and the original rPPG signal.

[0139] First, average pooling is performed on the second dimension of the denoised feature vector z (7×63×200) to obtain intermediate features (7×1×200). Then, a linear layer is passed to reduce the channel dimension, and finally the predicted rPPG signal (1×1×200) is obtained.

[0140] The trained adaptive filter model is used to estimate remote physiological signals. The specific estimation process is as follows:

[0141] Step 2.1, use the sampling module to sample samples from those following a normal distribution. An initial noise rPPG signal x' is obtained by sampling in Gaussian noise. N Then, a randomly generated noise rPPG signal x' is used. N As the starting point of the denoising process, the mathematical expression for sampling is as follows:

[0142]

[0143] Among them, the randomly generated noise rPPG signal x' N The dimension is 1×200.

[0144] Step 2.2: Following the method in Step 1.3, input the face video into the MSTmap calculation module to calculate the multi-scale spatiotemporal map.

[0145] Step 2.3, following the method in Step 1.4, process the randomly generated noise rPPG signal x' N The multi-scale spatiotemporal graph is input into the fusion module for fusion to obtain fused data y'. N .

[0146] Step 2.4, merge the data y' N The data is input into the denoising module for denoising, resulting in denoised data.

[0147] The denoising method during estimation differs from that during training. The feature vector output after denoising by the denoising module during estimation is then used as input data again for further denoising. This process is repeated K times before the final output is obtained. The expression is shown below:

[0148] z k =MFD(z) k+1 )+z k+1

[0149] MFD(z k+1 )=Fourier(Contextual(Plain(z k+1 ))×c

[0150] Where k is the number of iterations in the denoising process, k = 1, 2, ..., K; z k Let z0 = y' be the feature vector of the k-th denoising process. N Finally, the feature vector z after K denoising processes is obtained. K That is, the data after noise reduction, z K The dimensions are 7×63×200.

[0151] z1=y 800 =y 1000 +MFD(y1000 ), k=1

[0152] z2=y 600 =y 800 +MFD(y 800 ), k=2

[0153] z3=y 400 =y 600 +MFD(y 600 k=3

[0154] z4=y 200 =y 400 +MFD(y 400 ), k=4

[0155] Step 2.5: Following the method in Step 1.6, the feature vector z that has undergone K denoising processes is... K The signal is input into the prediction module, and finally the predicted rPPG signal is obtained.

[0156] This invention uses a multi-scale spatiotemporal map (MSTmap) instead of directly inputting video, which has the following advantages:

[0157] During MSTmap calculation, using facial recognition tools to accurately locate faces effectively improves the model's recognition accuracy. If video is directly input into the model, errors are likely to occur because the model struggles to accurately distinguish the effective areas in the image, such as mistaking the background for a face. The use of facial recognition tools allows the model to focus on the effective facial areas, thereby improving recognition accuracy.

[0158] In addition to the use of facial recognition tools, MSTmap innovatively introduces YUV channels on top of the original RGB channels. These three additional channels have been proven effective by previous research, providing the model with richer feature information and thus helping the model to perform better training and inference.

[0159] Furthermore, MSTmap cleverly arranges and combines facial ROIs. The original 6 regions are combined to create 63 different combinations. These rich combinations allow the multi-filter module of the MFD layer in Denoser to extract more valuable information, resulting in more accurate calculation results.

[0160] This invention fuses the rPPG signal with the color channel dimensions of the MSTmap. On one hand, during channel fusion, the one-dimensional noisy rPPG signal is combined with the six-dimensional MSTmap color channels. Since the MSTmap itself dominates in terms of dimension, it is not affected by the interference of the noisy rPPG signal, thus ensuring the model's prediction performance. On the other hand, the noisy rPPG signal is fused with the MSTmap using a concat method. This process does not alter the numerical values ​​of the noisy rPPG signal and the MSTmap; it simply connects their dimensions, and the features they represent are independent of each other. This unique fusion method allows the model to learn the data distribution of the rPPG signal from the noisy rPPG signal, and simultaneously utilizes this data distribution to provide powerful guidance for mapping the MSTmap to the true rPPG signal.

[0161] Example 2

[0162] This embodiment provides a computer terminal, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the remote physiological signal detection method based on an adaptive filter as described in Embodiment 1.

[0163] like Figure 4 As shown, the computer terminal provided in this embodiment includes: at least one processor 101, and a memory 102 connected to at least one processor 101. This embodiment does not limit the specific connection medium between the processor 101 and the memory 102. Figure 4 The example shown is the connection between processor 101 and memory 102 via bus 100. Bus 100 is... Figure 4 The connections between other components are shown in bold lines and are for illustrative purposes only, not as limiting information. Bus 100 can be divided into address bus, data bus, control bus, etc., for ease of representation. Figure 4 The bus is represented by a single thick line, but this does not indicate that there is only one bus or one type of bus. Alternatively, the processor 101 may also be called a controller; there is no restriction on the name.

[0164] In this embodiment, the memory 102 stores instructions that can be executed by at least one processor 101. The at least one processor 101 can execute the aforementioned method by executing the instructions stored in the memory 102.

[0165] The processor 101 is the control center of the device. It can connect to various parts of the control device through various interfaces and lines. By running or executing instructions stored in memory 102 and calling data stored in memory 102, the processor can perform various functions and process data, thereby monitoring the device as a whole.

[0166] In one possible design, processor 101 may include one or more processing units. Processor 101 may integrate an application processor and a modem processor, wherein the application processor mainly handles the operating system, user interface, and applications, and the modem processor mainly handles wireless communication. It is understood that the modem processor may also not be integrated into processor 101. In some embodiments, processor 101 and memory 102 may be implemented on the same chip; in some embodiments, they may also be implemented on separate chips.

[0167] Processor 101 can be a general-purpose processor, such as a central processing unit (CPU), digital signal processor, application-specific integrated circuit, field-programmable gate array or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the remote physiological signal detection method based on adaptive filters disclosed in Embodiment 1 can be directly implemented by the hardware processor, or implemented by a combination of hardware and software modules in processor 101.

[0168] Memory 102, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules. Memory 102 may include at least one type of storage medium, such as flash memory, hard disk, multimedia card, card-type memory, random access memory (RAM), static random access memory (SRAM), programmable read-only memory (PROM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), magnetic storage, magnetic disk, optical disk, etc. Memory 102 can be any other medium capable of carrying or storing desired program code in the form of instructions or data structures that can be accessed by a computer, but is not limited thereto. In this embodiment, memory 102 can also be a circuit or any other device capable of implementing storage functions for storing program instructions and / or data.

[0169] By designing and programming the processor 101, the code corresponding to the security verification method described in the foregoing embodiments can be embedded into the chip, thereby enabling the chip to execute the code during operation. Figure 2 The steps of the remote physiological signal detection method based on adaptive filters are shown. How to design and program the processor 101 is a technique well-known to those skilled in the art and will not be described further here.

[0170] Example 3

[0171] This embodiment provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a processor, it implements the steps of the remote physiological signal detection method based on an adaptive filter as described in Embodiment 1.

[0172] The computer-readable storage medium may include flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the storage medium may be an internal storage unit of a computer device, such as the hard disk or memory of the computer device. In other embodiments, the storage medium may also be an external storage device of the computer device, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., provided on the computer device. Of course, the storage medium may include both internal storage units and external storage devices of the computer device. In this embodiment, the memory is typically used to store the operating system and various application software installed on the computer device. In addition, the memory can also be used to temporarily store various types of data that have been output or will be output.

[0173] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A remote physiological signal detection method based on adaptive filters, characterized in that, include: Acquire the face video to be detected, and obtain the noise rPPG signal through noise processing; Feature extraction is performed on the face video to obtain a multi-scale spatiotemporal map; The noisy rPPG signal and the multi-scale spatiotemporal map are fused to obtain fused data; The fused data is denoised using a denoising module to obtain denoised data. This denoising module employs an adaptive filter denoising module, internally composed of a multi-layer filter structure. Each layer includes a fixed-spectrum shaping filter, a context-adaptive shaping filter, and a spectral periodicity feature extraction filter connected sequentially. The fixed-spectrum shaping filter performs initial frequency domain filtering using randomly initialized fixed parameters. The context-adaptive shaping filter uses the output of the fixed-spectrum shaping filter to generate data-dependent filtering parameters, achieving context-adaptive frequency domain filtering to reconstruct the periodicity and error components in the time series. The spectral periodicity feature extraction filter performs frequency domain reconstruction on the output of the context-adaptive shaping filter to recover the periodicity and error components in the time series. Based on the denoised data, the predicted rPPG signal is output.

2. The remote physiological signal detection method based on adaptive filters according to claim 1, characterized in that, The processing steps of the fixed-spectrum shaping filter include: Perform a Fourier transform on the input fused data, i.e., the time series T, to obtain the frequency domain representation of the time series. In the formula, Fourier transform processing; Using pre-randomly initialized and fixed filter parameters H φ Frequency domain representation Element-wise multiplication yields the pre-shaped frequency domain features. In the formula, and These are all parameters of the channel-shared filter; ⊙ L This is element-wise multiplication along the dimension of sequence length L; right Perform an inverse Fourier transform to obtain the output T of the fixed-spectrum shaping filter. plain : In the formula, This represents the inverse Fourier transform.

3. The remote physiological signal detection method based on adaptive filters according to claim 2, characterized in that, The processing steps of the scenario-adaptive shaping filter include: For the output T of the fixed spectrum shaping filter plain Perform a Fourier transform to obtain the frequency domain representation after the Fourier transform. Generate adaptive filter parameters using a data-dependent neural network. In the formula, NN(·) represents neural network processing; Frequency domain representation With adaptive filter parameters Element-wise multiplication yields the shaped frequency domain features. Frequency domain features after shaping Perform an inverse Fourier transform to obtain the output T of the scene-adaptive shaping filter. ctx :

4. The remote physiological signal detection method based on an adaptive filter according to claim 3, characterized in that, The processing procedure of the spectral periodic feature extraction filter includes: The output T of the scene-adaptive shaping filter ctx Perform a Fourier transform to extract amplitude and phase information: In the formula, A is the amplitude, Φ is the phase, and ∠ represents the angle; Select the top K important frequency components from amplitude A. The frequencies and phases corresponding to the K important frequency components are respectively... and Constructing a frequency domain reconstructed signal through weighted synthesis: In the formula, T syn For frequency domain reconstructed signal; A k Let be the amplitude of the kth significant frequency component in amplitude A; t is the time variable; Reconstruct the frequency domain signal T syn As the output of the entire noise reduction process.

5. The remote physiological signal detection method based on an adaptive filter according to claim 1, characterized in that, The expression for noise processing is as follows: In the formula, x0 is the original rPPG signal; x t’ Let q(x) be the noise rPPG signal after the t'th noise processing; t' = 1, 2, ..., N, where N is the total number of noise processing iterations; t’ |x0) represents the state x given an initial state x0. t’ The conditional probability distribution; As the attenuation factor, The value gradually decreases as the number of processing iterations t' increases during the processing; ∈ represents noise.

6. The remote physiological signal detection method based on an adaptive filter according to claim 1, characterized in that, The specific process of feature extraction for facial features includes: The face video is divided into multiple segments, and each segment yields F frames of video images. Face detection is performed on video images, and the face is divided into n regions of interest (ROIs). Combining n ROIs yields two combinations: single ROIs and combinations of different ROIs. n -1 combination; For each combination, calculate the average pixel value of all ROIs in the combination across each color channel; there are a total of 6 color channels, namely RGB and YUV; Finally, the dimension is 6×(2) n The vector of -1)×F is used as a multi-scale spatiotemporal graph.

7. The remote physiological signal detection method based on an adaptive filter according to claim 6, characterized in that, The specific process of fusing the noisy rPPG signal and the multi-scale spatiotemporal map includes: The dimension of the noise rPPG signal is 1×F. Expanding the dimension of the noise rPPG signal to 1×(2F) will increase the dimension of the noise rPPG signal to 1×(2F). n -1)×F, the expanded noisy rPPG signal is stitched with the multi-scale spatiotemporal map along the channel dimension to obtain fused data, the dimension of which is 7×(2). n -1)×F.

8. The remote physiological signal detection method based on an adaptive filter according to claim 1, characterized in that, The noise rPPG signal is Gaussian noise that follows a standard normal distribution.

9. A remote physiological signal detection system based on an adaptive filter, characterized in that, The method for remote physiological signal detection based on an adaptive filter as described in any one of claims 1 to 8; the detection system comprises: The noise processing module is used to acquire the face video to be detected and obtain the noise rPPG signal through noise processing. The MSTmap calculation module is used to extract features from face videos to obtain multi-scale spatiotemporal maps; The fusion module is used to fuse the noisy rPPG signal and the multi-scale spatiotemporal map to obtain fused data; A denoising module is used to denoise the fused data to obtain denoised data. This denoising module employs an adaptive filter denoising module, internally composed of a multi-layer filter structure. Each layer includes a fixed-spectrum shaping filter, a context-adaptive shaping filter, and a spectral periodic feature extraction filter connected in sequence. The fixed-spectrum shaping filter performs preliminary frequency domain filtering using randomly initialized fixed parameters. The context-adaptive shaping filter uses the output of the fixed-spectrum shaping filter to generate data-dependent filtering parameters, achieving context-adaptive frequency domain filtering, thereby reconstructing the periodicity and error components in the time series. The spectral periodic feature extraction filter performs frequency domain reconstruction on the output of the context-adaptive shaping filter to recover the periodicity and error components in the time series. The prediction module is used to output a predicted rPPG signal based on the denoised data.

10. A computer terminal, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the remote physiological signal detection method based on an adaptive filter as described in any one of claims 1 to 8.

Images