A lightweight identity authentication method based on remote photoelectric volume pulse wave signal

Abstract

This invention belongs to the field of biometric recognition technology, specifically disclosing a lightweight identity authentication method based on remote photoplethysmography (rPPG) signals. This method constructs a dual-color fusion spatiotemporal map by fusing color fidelity information in the RGB space with luminance-chrominance decoupled representations in the YUV space, thereby enhancing the robustness of physiological signal representation. Furthermore, this invention also constructs a lightweight identity authentication model based on rPPG signals. The model includes an rPPG signal estimation module that extracts features through a heterogeneous convolutional network. This network extracts global rhythmic features and local morphological features of the rPPG signal using depthwise separable convolution and grouped convolution, and fuses these two features through an adaptive gating mechanism. The model also includes an rPPG morphological feature enhancement module that explicitly models detailed features through periodic segmentation and a lightweight gated recurrent unit encoder. Finally, the biometric authentication module outputs the identity authentication result.

Description

Technical Field

[0001] This invention belongs to the field of biometric identification technology, specifically relating to a lightweight identity authentication method based on remote photoplethysmography pulse wave signals, which is particularly suitable for scenarios such as contactless identity verification, intelligent security monitoring, and contactless access control management. Background Technology

[0002] With the widespread adoption of digital services such as financial payments, access control and security, and mobile terminal unlocking, biometric identification technologies based on inherent physiological or behavioral characteristics of the human body have become an important technical means in the field of identity authentication. Existing biometric identification technologies mainly include fingerprint recognition, iris recognition, facial recognition, and palmprint recognition. These technologies typically collect stable anatomical or textural features of the human body and use pattern recognition or deep learning models to complete identity identification. While these methods can achieve high recognition rates under controlled environments, they still have the following problems: First, most solutions rely on dedicated acquisition equipment or close-range interaction. For example, fingerprint and palmprint recognition require users to contact the sensors, and long-term wear or frequent contact may cause hygiene problems or reduce the user experience. Second, visible physical features such as faces are at risk of being forged by photos, videos, or 3D masks, usually requiring additional liveness detection modules to ensure security. Third, biometric features such as faces and irises are unique and unchangeable; once leaked, they will cause permanent identity security risks.

[0003] Compared to traditional static geometric features, physiological signals such as electrocardiograms (ECG) and photoplethysmography (PPG) directly reflect an individual's cardiovascular and circulatory system characteristics, exhibiting resistance to forgery and good intra-individual stability. Existing research has used contact-based PPG sensors to collect pulse wave signals from fingers or earlobes, extracting their time-domain, frequency-domain, and morphological features for identity recognition or authentication; some solutions have already been applied to access control and wearable devices. However, these methods require close contact between the photoelectric probe and the skin, and place high demands on user cooperation and specific scene conditions, making them difficult to meet the application needs of long-distance, contactless scenarios, and also hindering seamless integration with existing camera systems.

[0004] Therefore, researchers began exploring the application of remote photoplethysmography (rPPG) signals in biometrics, attempting to authenticate individuals by analyzing differences in rPPG signal waveform morphology and spectral distribution. Unlike face recognition, which relies on visible appearance features, rPPG-based authentication relies on subtle skin color changes in videos to capture physiological rhythms, offering a degree of anti-spoofing and privacy protection. Since the introduction of remote photoplethysmography, various algorithmic systems have been developed for rPPG signal extraction and enhancement, including traditional blind source separation techniques (ICA / PCA), physiological model-based methods (CHROM, POS), and end-to-end estimation methods based on deep learning that have emerged in recent years. Traditional methods have relatively low computational complexity, but their robustness is often difficult to guarantee in complex scenarios such as head movements and lighting changes. Deep learning methods, through convolutional networks, temporal networks, or Transformer structures, can automatically learn pulse-related spatiotemporal features from face videos. While this improves the accuracy of rPPG signal extraction to some extent, it faces challenges such as large model parameters, high computational resource consumption, and limited deployment of edge devices.

[0005] Despite numerous studies on rPPG signals in heart rate monitoring, their application in biometrics remains in its early stages. Extending rPPG technology from heart rate monitoring to high-precision biometric authentication faces significant challenges: First, existing rPPG research largely focuses on coarse-grained physiological indicators such as heart rate estimation, lacking systematic exploration of fine-grained morphological features in rPPG signal waveforms that can be used for identity authentication (such as systolic peaks, diastolic peaks, dicrotic notches, and time-frequency domain structures). This results in the rPPG signal's individual discriminative power not being fully utilized, making it difficult to meet the authentication accuracy required for practical applications. Second, to cope with complex scenarios such as changes in lighting and motion interference, existing methods tend to employ large-scale deep learning models to improve robustness. However, this significantly increases the number of model parameters and computational overhead, creating a prominent contradiction with the limited resources of edge devices and hindering the practical deployment of the technology. Summary of the Invention

[0006] This invention proposes a lightweight identity authentication method based on remote photoplethysmography (rPPG) signals. Addressing the shortcomings of existing technologies in rPPG signal morphological feature mining and excessive model complexity, this invention enhances the robustness of physiological signal representation by constructing a dual-color fusion spatiotemporal graph. The method establishes a lightweight identity authentication model based on rPPG signals, incorporating an rPPG signal estimation module and an rPPG morphological feature enhancement module to process the rPPG signal. While deeply mining the fine-grained morphological features of the rPPG signal waveform, it significantly reduces the number of model parameters and computational overhead, achieving a balance between authentication accuracy and edge deployment efficiency, thus meeting the practical needs of contactless identity authentication scenarios.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] A lightweight authentication method based on remote photoplethysmography (PPG) signals includes the following steps:

[0009] Step 1. Obtain raw video data and preprocess it to build a training dataset;

[0010] During the preprocessing process, for each frame of the face image in the original video, a spatiotemporal mapping method that fuses RGB and YUV dual color spaces is used to generate a dual-color fused spatiotemporal map corresponding to the video data.

[0011] Step 2. Build a lightweight identity authentication model based on remote photoplethysmography (rPPG) signals. The model includes an rPPG signal estimation module, an rPPG morphological feature enhancement module, and a biometric authentication module.

[0012] The rPPG signal estimation module is used to extract features from a three-stage cascaded heterogeneous convolutional network (HGB) for a dual-color fusion spatiotemporal map, thereby obtaining the remote photoplethysmography (rPPG) signal.

[0013] Each stage of HGB consists of a downsampled heterogeneous convolutional block HGB_down and a basic heterogeneous convolutional block HGB_base concatenated together; both HGB_down and HGB_base adopt a dual-branch processing path that includes a large kernel perception path and a small kernel aggregation path.

[0014] The large kernel sensing path captures the global rhythmic features of rPPG signals through depthwise separable convolution; the small kernel aggregation path extracts the local morphological features of rPPG signals through grouped convolution.

[0015] The output features of the large kernel perception path and the small kernel aggregation path are fused through an adaptive gating mechanism;

[0016] The rPPG morphological feature enhancement module explicitly models the detailed features in the rPPG signal through periodic segmentation and a lightweight gated cyclic encoder to obtain the morphological feature vector corresponding to the rPPG signal.

[0017] The biometric authentication module uses a classifier to determine identity based on the morphological feature vector.

[0018] Step 3. Train the model based on the training dataset, and use the trained model for identity feature matching and authentication.

[0019] Furthermore, based on the aforementioned lightweight authentication method based on remote photoplethysmography (PPG) signals, this invention also proposes a corresponding lightweight authentication system based on remote PPG signals, which adopts the following technical solution:

[0020] A lightweight identity authentication system based on remote photoplethysmography (PPG) signals includes the following modules:

[0021] The preprocessing module is used to acquire raw video data and perform preprocessing to build a training dataset;

[0022] During the preprocessing process, for each frame of the face image in the original video, a spatiotemporal mapping method that fuses RGB and YUV dual color spaces is used to generate a dual-color fused spatiotemporal map corresponding to the video data.

[0023] And an identity authentication module, used to build a lightweight identity authentication model based on remote photoplethysmography (rPPG) pulse wave signals. This model includes an rPPG signal estimation module, an rPPG morphological feature enhancement module, and a biometric authentication module.

[0024] The rPPG signal estimation module is used to extract features from a three-stage cascaded heterogeneous convolutional network (HGB) for a dual-color fusion spatiotemporal map, thereby obtaining the remote photoplethysmography (rPPG) signal.

[0025] Each stage of HGB consists of a downsampled heterogeneous convolutional block HGB_down and a basic heterogeneous convolutional block HGB_base concatenated together; both HGB_down and HGB_base adopt a dual-branch processing path that includes a large kernel perception path and a small kernel aggregation path.

[0026] The large kernel sensing path captures the global rhythmic features of rPPG signals through depthwise separable convolution; the small kernel aggregation path extracts the local morphological features of rPPG signals through grouped convolution.

[0027] The output features of the large kernel perception path and the small kernel aggregation path are fused through an adaptive gating mechanism;

[0028] The rPPG morphological feature enhancement module explicitly models the detailed features in the rPPG signal through periodic segmentation and a lightweight gated cyclic encoder to obtain the morphological feature vector corresponding to the rPPG signal.

[0029] The biometric authentication module uses a classifier to determine identity based on the morphological feature vector.

[0030] The model is trained based on the training dataset, and the trained model is used for identity feature matching and authentication.

[0031] The present invention has the following advantages:

[0032] As described above, this invention discloses a lightweight identity authentication method based on remote photoplethysmography (rPPG) signals. This method constructs a lightweight identity authentication model based on rPPG signals, which includes an rPPG signal estimation module and an rPPG morphological feature enhancement module. This invention employs a spatiotemporal mapping method that fuses two color spaces, combining the original color fidelity of the RGB space with the robustness of the YUV space under varying illumination to form a complementary spatiotemporal signal representation. This dual-color fusion spatiotemporal representation enhances the adaptability of input features to illumination fluctuations and motion interference, providing rich and environmentally robust foundational data for subsequent high-precision rPPG signal estimation and identity determination. Furthermore, the rPPG signal estimation module, through the design of heterogeneous gated convolutional blocks, including large-kernel perception paths and small-kernel aggregation paths, fully preserves the global rhythm and local morphological features of the rPPG signal while ensuring model lightweightness, achieving an optimal balance between authentication accuracy and resource consumption. The rPPG morphological feature enhancement module explicitly models key details in the rPPG signal waveform, such as contraction peaks, diastolic peaks, and diphtheria notches, through periodic segmentation and a lightweight gated cyclic encoder, significantly enhancing the individual discriminativeness and stability of identity features. Simultaneously, during model training, the total loss function constructed in this invention employs a composite loss function consisting of signal estimation loss and identity authentication loss for joint training. Through the synergistic optimization of physiological waveform fidelity constraints and identity discrimination objectives, it effectively avoids performance losses caused by inconsistent module objectives in traditional cascaded architectures, thereby improving the model's cross-scenario robustness. Finally, by fusing dual-color space spatiotemporal graph representation, heterogeneous gated convolutional block multi-scale feature extraction, and an end-to-end lightweight signal enhancement and authentication network, this invention effectively reduces the number of model parameters and computational resources. While ensuring identity authentication accuracy, it improves inference efficiency, achieving end-to-end lightweight modeling. It possesses good versatility and deployment potential in resource-constrained environments or mobile devices, making it suitable for scenarios such as contactless identity verification, intelligent security monitoring, and contactless access control management. Attached Figure Description

[0033] Figure 1 This is a general framework diagram of the lightweight identity authentication method based on remote photoplethysmography pulse wave signal in an embodiment of the present invention;

[0034] Figure 2 This is a flowchart of obtaining a dual-color fusion spatiotemporal map in an embodiment of the present invention;

[0035] Figure 3 This is a structural diagram of the rPPG signal estimation module in an embodiment of the present invention;

[0036] Figure 4This is a structural diagram of HGB_down and HGB_base in the HGB module of this embodiment of the invention;

[0037] Figure 5 This is a flowchart of the large-core perception path of the HGB module in an embodiment of the present invention;

[0038] Figure 6 This is a flowchart of the HGB module small core aggregation path in an embodiment of the present invention;

[0039] Figure 7 This is a structural diagram of the rPPG morphological feature enhancement module in an embodiment of the present invention;

[0040] Figure 8 This is a flowchart of the shrinkage peak detection and period segmentation in an embodiment of the present invention;

[0041] Figure 9 This is a flowchart of the SE attention mechanism in an embodiment of the present invention;

[0042] Figure 10 This is a structural diagram of a lightweight gated loop unit encoder in an embodiment of the present invention. Detailed Implementation

[0043] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0044] Example 1

[0045] like Figure 1 As shown in Embodiment 1, a lightweight identity authentication method based on remote photoplethysmography (PPG) signals is described, including the following steps:

[0046] Step 1. Obtain the raw video data, preprocess the raw video data, and construct the training dataset.

[0047] The original video data includes multiple publicly available and authoritative datasets, covering different recording conditions, subject activity states, and real measurement methods. The videos in the dataset are preprocessed to construct a dataset for training a lightweight identity authentication model for remote photoplethysmography (PPG) signals in step 2. The preprocessed dataset is then divided into training and testing sets.

[0048] During the preprocessing process, for each frame of the face image in the original video, a spatiotemporal mapping method that fuses RGB and YUV dual color spaces is used to generate a dual-color fused spatiotemporal map corresponding to the video data.

[0049] like Figure 2 As shown, the process for obtaining the dual-color fusion spatiotemporal map is as follows:

[0050] Set the input video sequence ;

[0051] in Indicates the number of frames. Indicates the height of the original video. This indicates the width of the original video.

[0052] First, a face alignment algorithm, such as FaceAlignment, is used to locate and crop the regions of interest (ROIs) of faces in each frame of the input video sequence, resulting in the cropped video sequence. ;

[0053] in Indicates the height of the cropped video. This indicates the width of the video after cropping.

[0054] Then, the cropped video frames and their corresponding real rPPG reference signals are uniformly resampled to a fixed sampling frequency. In this embodiment, This is to eliminate timing distortion caused by differences in the frame rate of the dataset.

[0055] The real rPPG reference signal is obtained through a contact sensor and is acquired synchronously with facial video.

[0056] Each frame of a face image is divided into multiple uniform spatial sub-blocks.

[0057] The average value of each channel in the RGB and YUV color spaces for each color space sub-block is calculated separately to fully explore the detail expression capability of the RGB color space and the robustness of the YUV color space under illumination changes, and the corresponding RGB color time series and YUV color time series for the color space sub-block are constructed accordingly.

[0058] The values ​​of the RGB color time series in frame t are defined as follows:

[0059] ;

[0060] This represents the feature vector of the RGB color time series at frame t. , , The first The average value of the R, G, and B channels of this sub-block in the RGB color space. It represents a three-dimensional real vector space.

[0061] The values ​​of the YUV color time series in frame t are defined as follows:

[0062] ;

[0063] This represents the feature vector of the YUV color time series at frame t. , , The first The average value of the Y, U, and V channels of the sub-block in the YUV color space.

[0064] The RGB color space consists of three independent channels: Red, Green, and Blue, defined as R, G, and B channels respectively. Each channel is responsible for storing the brightness information of the corresponding color.

[0065] The YUV color space consists of three channels: Y, U, and V.

[0066] Y represents luminance, which is the brightness or darkness of an image. The human eye is most sensitive to changes in brightness, therefore the Y channel carries the main outline and details of an image;

[0067] U represents Blue-Difference, which is the difference between the blue component and the brightness.

[0068] V represents Red-Difference, which is the difference between the red component and the brightness.

[0069] The RGB and YUV color sequences of each spatial sub-block are concatenated frame by frame along the channel dimension to obtain the enhanced spatiotemporal signal sequence of that spatial sub-block. The value of this enhanced sequence in frame t is:

[0070] ;

[0071] in ( ) indicates a splicing operation. This represents a six-dimensional real vector space.

[0072] Finally, the enhanced spatiotemporal signal sequences of all spatial sub-blocks are arranged according to their corresponding spatial locations and stacked along the time dimension to generate a dual-color fused spatiotemporal map. Where C represents the number of fusion channels, H represents the height of the dual-color fusion spatiotemporal map, and W represents the width of the dual-color fusion spatiotemporal map.

[0073] The generated dual-color fused spatiotemporal map samples and their corresponding identity labels are divided into training and testing sets for subsequent model training and evaluation.

[0074] This invention integrates the decoupled representations of brightness and chromaticity in the RGB and YUV color spaces to form complementary spatiotemporal signals, effectively enhancing the robustness of features to illumination fluctuations and motion interference. This provides a complete and environmentally robust data foundation for subsequent high-precision rPPG signal estimation and identity recognition.

[0075] Step 2. Build a lightweight identity authentication model based on remote photoplethysmography (rPPG) pulse wave signals. The model includes an rPPG signal estimation module, an rPPG morphological feature enhancement module, and a biometric authentication module. Each module achieves gradient sharing and feature transfer through end-to-end joint training.

[0076] The rPPG signal estimation module is used to extract features from a three-stage cascaded heterogeneous convolutional network (HGB) for a dual-color fusion spatiotemporal map, thereby obtaining the remote photoplethysmography (rPPG) signal.

[0077] Each stage of HGB consists of a downsampled heterogeneous convolutional block HGB_down and a basic heterogeneous convolutional block HGB_base concatenated together; both HGB_down and HGB_base adopt a dual-branch processing path that includes a large kernel perception path and a small kernel aggregation path.

[0078] The large kernel sensing path captures the global rhythmic features of rPPG signals through depthwise separable convolution; the small kernel aggregation path extracts the local morphological features of rPPG signals through grouped convolution.

[0079] The output features of the large kernel perception path and the small kernel aggregation path are fused through an adaptive gating mechanism;

[0080] The rPPG morphological feature enhancement module explicitly models the detailed features in the rPPG signal through periodic segmentation and a lightweight gated cyclic encoder to obtain the morphological feature vector corresponding to the rPPG signal.

[0081] The biometric authentication module uses a classifier composed of a multi-layer fully connected network to perform identity determination based on morphological feature vectors.

[0082] In this embodiment, the rPPG signal estimation module is used to decouple high-quality rPPG signals from the dual-color fused spatiotemporal map sequence, such as... Figure 3 As shown, the processing flow of the rPPG signal estimation module is as follows:

[0083] First, receive the dual-color fusion spatiotemporal map generated in step 1. ;

[0084] Further through the initial feature extraction layer Spatial downsampling and channel expansion are performed to reduce the resolution of the feature map and increase the number of channels, laying the foundation for subsequent processing and obtaining the initial feature map. The calculation formula is as follows:

[0085] ;

[0086] in ( ) represents a 3×3 convolution operation. ( ) indicates batch normalization operation. ( ) represents the activation function. ( ) indicates a max pooling operation with a step size of 2. Represents the initial feature map The number of channels.

[0087] Specifically, the initial feature extraction layer's processing flow is as follows: First, the input features are spatially downsampled using a 3×3 convolutional layer with a stride of 2, while simultaneously expanding the number of channels from C to C0. Next, batch normalization is applied to stabilize the training process, and non-linearity is introduced through the ReLU activation function to enhance the model's expressive power. Subsequently, a 3×3 max-pooling layer with a stride of 2 further reduces the feature map resolution, achieving a second 2x downsampling. This series of operations not only preserves important physiological details but also reduces the feature resolution to H / 4×W / 4, significantly reducing the computational complexity of subsequent heterogeneous convolutional networks and laying the foundation for lightweight network design.

[0088] Subsequently, the initial feature map enters a three-stage cascaded heterogeneous convolutional network; HGB_down refers to a heterogeneous gated convolutional block with a stride of 2, which downsamples the feature map to reduce spatial resolution; HGB_base refers to a heterogeneous gated convolutional block with a stride of 1, which is used to extract features without changing the spatial resolution of the feature map, thereby further enhancing the expressive power of the features; although HGB_down and HGB_base differ in stride settings, they are both built on the same core heterogeneous convolutional block HGB architecture, ensuring the consistency of the internal structure of the module and the efficiency of feature extraction.

[0089] The initial feature map is processed through the first-stage heterogeneous convolutional network as follows:

[0090] ; ;

[0091] in This represents the features of the initial feature map after processing by the first-stage downsampling heterogeneous convolutional block. express Features after the first stage of basic heterogeneous convolutional block processing This indicates a heterogeneous gated convolution operation with a stride of 2 in the first stage. This indicates a heterogeneous gated convolution operation with a stride of 1 in the first stage. Representation of feature map The number of channels.

[0092] The computation process of the second-stage heterogeneous convolutional network is as follows:

[0093] ; ;

[0094] in express Features after the second stage of downsampling heterogeneous convolutional block processing express Features after the second stage of basic heterogeneous convolutional block processing This indicates a heterogeneous gated convolution operation with a stride of 2 in the second stage. This indicates a heterogeneous gated convolution operation with a stride of 1 in the second stage. Representation of feature map The number of channels.

[0095] The computation process of the third-stage heterogeneous convolutional network is as follows:

[0096] ; ;

[0097] in express Features after the third stage of downsampling heterogeneous convolutional block processing express Features after the third stage of basic heterogeneous convolutional block processing This represents the heterogeneous gated convolution operation with a stride of 2 in the third stage. This represents the heterogeneous gated convolution operation with a stride of 1 in the third stage. Representation of feature map The number of channels.

[0098] Through this hierarchical design that progressively reduces resolution and moderately increases the number of channels, the network can progressively extract multi-scale rPPG features from local details to global rhythms, enhancing the ability to model the spatiotemporal structure of rPPG waveforms while controlling computational overhead.

[0099] Finally, the output features of the third-stage heterogeneous convolutional network are fed into the global feature aggregation and rPPG regression layer for processing to obtain continuous rPPG signals. Its expression is as follows:

[0100] ;

[0101] in ( ) represents a 1×1 convolution operation. Indicates batch normalization, ( ) represents the activation function. ( ) indicates a global average pooling operation. ( ) represents the linear mapping operation of the fully connected layer.

[0102] Specifically, the global feature aggregation and rPPG regression layer aims to regress high-quality rPPG signals. The processing flow is as follows: First, 1×1 convolution operations are used to integrate information between channels and adjust the number of channels. Next, batch normalization is performed to stabilize the training process, reduce dependence on initialization, and accelerate network training. Then, an exponential linear unit (ELU) activation function is introduced to enhance the model's non-linear expressive power while automatically adjusting negative values ​​to prevent neuron "death." To achieve global feature aggregation, global average pooling is used to compress the feature map of each channel into a single value, reducing the number of model parameters and mitigating the risk of overfitting. Finally, a fully connected layer maps the aggregated global features to the output space of the rPPG signal, completing the conversion from high-dimensional features to physiological signals, providing accurate and reliable physiological feature information for subsequent identity authentication.

[0103] In this embodiment, the HGB module achieves dynamic feature modulation through heterogeneous collaboration between the large-core sensing path and the small-core aggregation path. This effectively suppresses redundant noise while reducing the number of parameters, significantly improving the ability to characterize individual waveform details. Figure 4 As shown, the processing flow of HGB_down and HGB_base in the HGB module is as follows:

[0104] Define the feature map of the input HGB module as follows ;

[0105] in This represents the number of channels in the input feature map. Indicates the height of the input feature map. This indicates the width of the input feature map.

[0106] The input feature map is processed in two paths: one path is the large kernel perception path, which captures the global rhythmic features of the remote photoplethysmography (PPG) signal through depthwise separable convolution; the other path is the small kernel aggregation path, which extracts the local morphological features of the remote PPG signal through grouped convolution.

[0107] The processing flow of the large kernel sensing path is as follows: Figure 5 As shown, the specific processing flow is as follows:

[0108] First, the channel dimension is compressed using 1×1 convolution, and then batch normalization and ReLU activation are performed.

[0109] Then, 7×7 depthwise separable convolution is used to capture long-range spatial dependencies.

[0110] Finally, a dynamic weight map is generated sequentially through batch normalization, ReLU activation, and 1×1 convolution mapping. Its expression is as follows;

[0111] ;

[0112] in, This represents the number of feature channels output by each group. Indicates the number of groups. Indicates the height of the dynamic weighted graph. Indicates the width of the dynamic weighted graph; ( ) represents a 1×1 pointwise convolution. ( ) indicates batch normalization operation. ( ) represents the activation function. ( ) indicates a 7×7 depth separable convolution operation.

[0113] To make dynamic weighted graph Channel number and static characteristics Alignment, dynamic weight graph The intermediate weighted graph is obtained by averaging along the grouping dimension. Its expression is as follows:

[0114] ;

[0115] in Indicates the first Group weighting chart.

[0116] Then, the Sigmoid activation function is used to... Mapping to the [0,1] interval yields a normalized gated weight graph. Its expression is as follows:

[0117] ;

[0118] in ( ) represents the Sigmoid activation function.

[0119] The processing flow of the small core aggregation path is as follows: Figure 6 As shown, the specific processing flow is as follows:

[0120] For the input feature map Perform 3×3 grouped convolution processing to directly obtain static features. Its expression is as follows;

[0121] ;

[0122] in Indicates the height of the feature map. Indicates the width of the feature map. ( ) indicates a 3×3 grouped convolution operation.

[0123] Finally, feature fusion is completed through an adaptive gating mechanism, resulting in a normalized gating weight map. Static features output by the small kernel aggregation path The formula for element-wise multiplication is as follows:

[0124] ;

[0125] in This indicates the fusion characteristics after modulation. This represents element-wise multiplication. This operation is performed using a gated weight graph. By performing channel-by-channel spatial weighting on the static feature V, we can achieve dynamic screening of global physiological rhythms and enhance key local morphological details.

[0126] Modulated fusion features After batch normalization and ReLU activation, the enhanced features after gating are obtained. Its expression is as follows:

[0127] ;

[0128] Gating weight graph in large kernel perception path It is dynamically generated, and the gating weight map is adjusted for different input samples, or even different spatial locations of the same input. The values ​​are all different, and the network can automatically adjust (adaptively) the static features in the small kernel aggregation path based on the content of the current features. The system prioritizes each channel and spatial location feature (weighting) to achieve adaptive control. The adaptive gating mechanism directly filters key physiological features, reducing the number of parameters while suppressing redundant noise, and significantly improving the ability to characterize individual-specific waveform details.

[0129] This invention innovatively designs a lightweight HGB module that achieves multi-scale feature collaborative extraction and dynamic fusion through two heterogeneous paths: a large-kernel sensing path focuses on capturing global rhythm trends such as heart rate variability; and a small-kernel aggregation path finely extracts local morphological details such as systolic peaks and dicrotic notches. Both are modulated and dynamically filtered element-wise through an adaptive gating mechanism, significantly reducing computational complexity while enhancing the ability to characterize individual-specific cardiovascular morphological features.

[0130] In this embodiment, the rPPG morphological feature enhancement module aims to transform the original rPPG signal into a stable feature representation with strong identity discrimination and cross-period consistency, such as... Figure 7 As shown, the processing flow of the rPPG morphological feature enhancement module is as follows:

[0131] First, receive the continuous rPPG signal output from the rPPG signal estimation module. ;

[0132] Then, the continuous long-range photoplethysmography (PPG) signal was segmented using shrinkage peak detection and period segmentation. The system consists of several periodic segments, and then the following processing procedure is performed on each periodic segment:

[0133] First, deep convolution operations are performed on periodic segments to capture local detail features in the rPPG signal, and an adaptive regularized smoothing feature learning process is introduced to reduce dependence on parameter initialization, thereby improving the overall stability of the model. Then, an SE attention mechanism is introduced to adaptively adjust the weights of each channel, thus obtaining the optimized feature vectors for each periodic segment; for example... Figure 9 As shown, the specific processing flow of the SE attention mechanism is as follows: Local features in the signal are processed sequentially through global average pooling, a fully connected layer, a ReLU activation function, another fully connected layer, and a Sigmoid activation function. Then, optimized features are obtained through channel multiplication. The combination of global average pooling and fully connected layers adaptively adjusts the weights of each channel, highlighting important features and suppressing irrelevant features, thus enhancing feature representation. The ReLU activation function, the Sigmoid activation function, and the channel-wise multiplication operation further optimize the feature selection process.

[0134] Next, The optimized feature vectors of each periodic segment are arranged according to the original occurrence order of their corresponding rPPG signal periods to form a periodic-level temporal feature sequence. ,in Indicates the first The eigenvector corresponding to each rPPG signal period.

[0135] To improve feature extraction efficiency, the periodic temporal features are further processed through depthwise separable convolution to obtain temporal features, which significantly reduces the computational cost while maintaining feature extraction capabilities.

[0136] Subsequently, the encoder of the lightweight gated loop unit was used to... The temporal features of each periodic segment are encoded to output a set of hidden state vectors. , Indicates the first The encoder utilizes the dynamic mechanism of its update and reset gates to model the rhythmic correlation across cycles and the phase evolution of systole-diastole within a single cycle in the rPPG signal, thereby extracting rhythmic information that characterizes individual cardiovascular activity. This allows for the detailed depiction of morphological features within waveforms, such as the rate of systolic peak rise and dicrotic notch, providing a robust basis for subsequent identity authentication.

[0137] The set of hidden state vectors is integrated and mapped in terms of features through a fully connected layer, and the final output is... A highly discriminative rPPG signal morphological feature vector This is used for subsequent identity verification matching, where Indicates the first The morphological feature vector of each periodic segment.

[0138] like Figure 8 As shown, the specific processing flow for shrinkage peak detection and period segmentation is as follows:

[0139] First, the cutoff frequency is adopted as... Zero-phase low-pass filter for rPPG signal Filtering is performed to obtain a smooth signal. ; and on Through a window length of The moving average is used to remove baseline drift, resulting in the drift-free signal value, expressed as:

[0140] ;

[0141] in, Indicates the smoothed signal at the 1st The instantaneous amplitude at each sampling point This indicates the smoothed signal in the summation index variable. The amplitude at the corresponding sampling point, To smooth the length of the window, Indicates the smoothed signal at the 1st The signal value after removing drift at each sampling point.

[0142] Then calculate First-order difference The calculation formula is as follows: ; and detect The set of candidate contraction peak positions is obtained from the zero-crossing point where the positive peak transitions to the negative peak. .

[0143] Subsequently, multiple constraints were applied to screen candidates based on amplitude thresholds and physiological intervals; specifically, the mean amplitude corresponding to each candidate contraction peak was calculated. with standard deviation Retain those that meet the conditions The candidate contraction peaks are identified, and the set of candidate contraction peak positions is obtained after screening. ,in An adjustable threshold coefficient is used; then interval filtering is performed, and calculations are performed. The time interval between adjacent candidate contraction peaks is retained only if it falls within the physiologically reasonable range. Candidate contraction peaks within the range are identified, thus yielding the final set of effective contraction peak locations. ,in The signal sampling rate, Indicates the first A total of 1 effective contraction peak positions were detected. One effective contraction peak, corresponding to One complete rPPG signal cycle.

[0144] Then, periodic segmentation is performed, using adjacent contraction peaks as boundaries to divide the continuous rPPG signal into segments. The first complete periodic segment, the... One original periodic segment The expression is:

[0145] ;

[0146] Each original periodic segment is resampled to a fixed length using linear interpolation. To obtain the set of periodic segments ,in Indicates the first Each periodic segment.

[0147] like Figure 10 As shown, the specific process of encoding timing features by the lightweight gated cyclic encoder is as follows:

[0148] First, the rPPG feature vector of the current period segment k. The hidden state vector of the previous period segment Input to the encoder.

[0149] Next, calculate the update gate output. This gate is used to control the transmission ratio of historical rPPG signal rhythm information and maintain the consistency of rPPG signal rhythm across cycles. Its calculation formula is as follows:

[0150] ;

[0151] in , These are the weight matrices for the input and hidden states corresponding to the update gate, respectively. To update the gate's bias vector, ( ) is the Sigmoid activation function.

[0152] Simultaneously, calculate the reset gate output. This gate controls whether to reset the historical state to accommodate local waveform changes from the systolic peak to the diastolic end within a single rPPG signal cycle. Its calculation formula is as follows:

[0153] ;

[0154] in , These are the weight matrices for the input and hidden states corresponding to the reset gate, respectively. This is the offset vector for resetting the gate.

[0155] Subsequently, the candidate hidden states are calculated. This state, which combines the current pulse characteristics with the reset historical state, is used to capture waveform details such as the location of the dicrotic notch. Its calculation formula is as follows:

[0156] ;

[0157] in , These are the input and weight matrices corresponding to the candidate hidden states, respectively. Let be the bias vector of the candidate hidden state. ( ) is the hyperbolic tangent activation function.

[0158] Then, update the hidden state of the gate pair for the previous period segment. Candidate hidden states of the current periodic segment Adaptive fusion is performed to obtain the hidden state of the current periodic segment, thereby enhancing the temporal consistency of identity discrimination features; the calculation formula is as follows: ;

[0159] in, For the first The hidden state of the output of the i-th period segment, and also as the i-th period segment output. The input state of each periodic segment.

[0160] After processing all K periodic segments sequentially, the lightweight gated cyclic encoder outputs a set of hidden state vectors. .

[0161] In this embodiment, the biometric authentication module receives the morphological feature vector output by the rPPG morphological feature enhancement module and completes identity authentication based on it. The core of this module is a classifier composed of a multi-layer fully connected network, which uses the ReLU activation function to introduce non-linearity. After processing the input morphological feature vector, the classifier outputs the matching probability of its belonging to each preset identity and completes identity discrimination based on the highest probability.

[0162] Step 3. Train the identity authentication model built based on the training dataset in Step 1, and use the trained identity authentication model to perform identity feature matching and authentication on real-time videos.

[0163] During model training, a composite loss function is constructed, which includes signal estimation loss and identity authentication loss. The calculation formula is as follows:

[0164] ;

[0165] in Indicates the signal estimation loss. This indicates a loss related to identity verification. , This represents the loss balance coefficient.

[0166] Because rPPG signals have significant periodicity, to optimize the network's performance in the rPPG signal estimation task, the negative Pearson correlation coefficient (NPCC) is used to calculate the signal estimation loss, which measures the consistency between the predicted waveform and the true blood volume pulse reference signal in terms of morphology and temporal timing. The formula for calculating the signal estimation loss is as follows:

[0167] ;

[0168] in To predict the total number of sampling points of the rPPG signal and the real rPPG signal, and The first The model-predicted rPPG signal value and the actual rPPG reference signal value at each sampling point. This represents the mean of the predicted rPPG signal. The mean of the real rPPG signal is represented; the NPCC loss significantly improves the estimation accuracy of periodic physiological signals by penalizing waveform morphology distortion and temporal offset.

[0169] To enhance the model's ability to discriminate different individual identity features, a cross-entropy loss based on periodic segments is used to classify identities for each rPPG signal period. The formula for calculating the identity authentication loss is as follows:

[0170] ;

[0171] in Indicates the total number of periodic segments. This indicates that the identity authentication model is for the first... Each periodic segment belongs to the real identity tag The predicted probability, , This represents the total number of identities in the training dataset. The identity authentication loss enables the model to extract individual-specific cardiovascular features from the rPPG morphology, improving cross-cycle consistency.

[0172] During the model training phase, the Adam optimizer was used to optimize the model parameters, with the initial learning rate set to 1×10⁻⁶. - ³, and dynamically adjust the learning rate using a cosine annealing strategy to improve training stability and convergence efficiency. During training, model performance is periodically evaluated on the validation set, comprehensively considering signal estimation accuracy (e.g., mean squared error, Sinsen correlation coefficient), identity authentication effectiveness (e.g., equal error rate, area under the ROC curve), and model efficiency (e.g., number of parameters, inference speed), and the optimal model parameters are saved. In the inference phase, the input face video to be authenticated first undergoes a preprocessing step to generate a dual-color fused spatiotemporal map. The data is then input into the trained model, and sequentially passes through the rPPG signal estimation, morphological feature enhancement, and biometric matching modules to finally output the identity determination result, achieving end-to-end lightweight biometric authentication.

[0173] Furthermore, to verify the lightweight characteristics of the method of this invention, the following experiments were conducted. In the experiments, RhythmNet, PhysFormer, and rPPG-MAE were selected as comparison methods, and the specific comparison data are shown in Table 1.

[0174] Table 1 Comparison of the method of this invention with mainstream methods

[0175]

[0176] The experimental results shown in Table 1 demonstrate that this invention significantly outperforms several comparable methods, including RhythmNet, PhysFormer, and rPPG-MAE, in terms of model lightweighting. The number of parameters in this invention is only 1.45M, representing reductions of approximately 91.8%, 80.4%, and 98.2% compared to RhythmNet, PhysFormer, and rPPG-MAE, respectively. This substantial reduction in model size facilitates storage and deployment on resource-constrained terminal devices. Furthermore, the floating-point operations (Flops) of this invention are only 2.39G, far lower than other comparable models. This translates to lower computational overhead and faster response times during the inference phase, providing a reliable guarantee for real-time identity authentication.

[0177] This invention organically integrates the rPPG signal estimation module, morphological feature enhancement module, and identity authentication module, and uses a composite loss function for joint optimization. It stably extracts high-quality rPPG signals from non-contact facial videos. While mining the morphological features of rPPG signals for identity recognition, it also takes into account the lightweight design of the model and real-time deployment on edge devices. It significantly reduces the computational overhead while ensuring authentication accuracy, and achieves efficient deployment on resource-constrained edge devices.

[0178] Example 2

[0179] This embodiment 2 describes a lightweight identity authentication system based on remote photoplethysmography (PPG) signals. This lightweight identity authentication system based on PPG signals is based on the same inventive concept as the lightweight identity authentication method based on PPG signals in embodiment 1 above.

[0180] A lightweight identity authentication system based on remote photoplethysmography (PPG) signals includes the following modules:

[0181] The preprocessing module is used to acquire raw video data and perform preprocessing to build a training dataset;

[0182] During the preprocessing process, for each frame of the face image in the original video, a spatiotemporal mapping method that fuses RGB and YUV dual color spaces is used to generate a dual-color fused spatiotemporal map corresponding to the video data.

[0183] And an identity authentication module, used to build a lightweight identity authentication model based on remote photoplethysmography (rPPG) pulse wave signals. This model includes an rPPG signal estimation module, an rPPG morphological feature enhancement module, and a biometric authentication module.

[0184] The rPPG signal estimation module is used to extract features from a three-stage cascaded heterogeneous convolutional network (HGB) for a dual-color fusion spatiotemporal map, thereby obtaining the remote photoplethysmography (rPPG) signal.

[0185] Each stage of HGB consists of a downsampled heterogeneous convolutional block HGB_down and a basic heterogeneous convolutional block HGB_base concatenated together; both HGB_down and HGB_base adopt a dual-branch processing path that includes a large kernel perception path and a small kernel aggregation path.

[0186] The large kernel sensing path captures the global rhythmic features of rPPG signals through depthwise separable convolution; the small kernel aggregation path extracts the local morphological features of rPPG signals through grouped convolution.

[0187] The output features of the large kernel perception path and the small kernel aggregation path are fused through an adaptive gating mechanism;

[0188] The rPPG morphological feature enhancement module explicitly models the detailed features in the rPPG signal through periodic segmentation and a lightweight gated cyclic encoder to obtain the morphological feature vector corresponding to the rPPG signal.

[0189] The biometric authentication module uses a classifier to determine identity based on the morphological feature vector.

[0190] The model is trained based on the training dataset, and the trained model is used for identity feature matching and authentication.

[0191] It should be noted that any content not mentioned in the above-described functional modules of the system described in Embodiment 2 can be referred to the step description of the corresponding method in Embodiment 1 above, and will not be repeated in detail here.

Claims

1. A lightweight identity authentication method based on remote photoplethysmography (PPG) signals, characterized in that, Includes the following steps: Step 1. Obtain raw video data and preprocess it to build a training dataset; During the preprocessing process, for each frame of the face image in the original video, a spatiotemporal mapping method that fuses RGB and YUV dual color spaces is used to generate a dual-color fused spatiotemporal map corresponding to the video data. Step 2. Build a lightweight identity authentication model based on remote photoplethysmography (rPPG) signals. The model includes an rPPG signal estimation module, an rPPG morphological feature enhancement module, and a biometric authentication module. The rPPG signal estimation module is used to extract features from a three-stage cascaded heterogeneous convolutional network (HGB) for a dual-color fusion spatiotemporal map, thereby obtaining the remote photoplethysmography (rPPG) signal. Each stage of HGB consists of a downsampled heterogeneous convolutional block HGB_down and a basic heterogeneous convolutional block HGB_base concatenated together; both HGB_down and HGB_base adopt a dual-branch processing path that includes a large kernel perception path and a small kernel aggregation path. The large kernel sensing path captures the global rhythmic features of rPPG signals through depthwise separable convolution; The small kernel aggregation path extracts local morphological features of the rPPG signal through grouped convolution; The output features of the large kernel perception path and the small kernel aggregation path are fused through an adaptive gating mechanism; The rPPG morphological feature enhancement module explicitly models the detailed features in the rPPG signal through periodic segmentation and a lightweight gated cyclic encoder to obtain the morphological feature vector corresponding to the rPPG signal. The biometric authentication module uses a classifier to determine identity based on the morphological feature vector. Step 3. Train the model based on the training dataset, and use the trained model for identity feature matching and authentication.

2. The lightweight identity authentication method based on remote photoplethysmography (PPG) signals according to claim 1, characterized in that, In step 1, the process for obtaining the dual-color fusion spatiotemporal map is as follows: Set the input video sequence ; in Indicates the number of frames. Indicates the height of the original video. Indicates the width of the original video; First, a face alignment algorithm is used to locate and crop the regions of interest (ROIs) of faces in each frame of the input video sequence, resulting in the cropped video sequence. ; in Indicates the height of the cropped video. Indicates the width of the cropped video; Then, the cropped video frames and their corresponding real rPPG reference signals are uniformly resampled to a fixed sampling frequency. This is to eliminate timing distortion caused by differences in the frame rate of the dataset; Each frame of a face image is divided into multiple uniform spatial sub-blocks; Calculate the average pixel value of each channel in the RGB and YUV color spaces of each spatial sub-block in each frame, and construct the RGB color time series and YUV color time series corresponding to the spatial sub-block accordingly. The values ​​of the RGB color time series in frame t are defined as follows: ； in This represents the feature vector of the RGB color time series at frame t. , , The first The average value of the R, G, and B channels of this sub-block in the RGB color space. Represents a three-dimensional real vector space; The values ​​of the YUV color time series in frame t are defined as follows: ； in This represents the feature vector of the YUV color time series at frame t. , , The first The average value of the Y, U, and V channels of the sub-block in the YUV color space; The RGB color time series and YUV color time series of each spatial sub-block are concatenated frame by frame along the channel dimension to obtain the enhanced spatiotemporal signal sequence of that spatial sub-block. The value of this enhanced spatiotemporal signal sequence in frame t is... for: ； in ( ) indicates a splicing operation. Represents a six-dimensional real vector space; Finally, the enhanced spatiotemporal signal sequences of all spatial sub-blocks are arranged according to their corresponding spatial locations and stacked along the time dimension to generate a dual-color fused spatiotemporal map. ; Where C represents the number of fusion channels, H represents the height of the dual-color fusion spatiotemporal map, and W represents the width of the dual-color fusion spatiotemporal map.

3. The lightweight identity authentication method based on remote photoplethysmography (PPG) signals according to claim 1, characterized in that, In step 2, the processing flow of the rPPG signal estimation module is as follows: First, receive the dual-color fusion spatiotemporal map generated in step 1. Further, through the initial feature extraction layer... Spatial downsampling and channel expansion are performed to obtain the initial feature map. The calculation formula is as follows: ； in ( ) represents a 3×3 convolution operation. ( ) indicates batch normalization operation. ( ) represents the activation function. ( ) indicates a max pooling operation with a step size of 2. Represents the initial feature map The number of channels; Initial feature map The computation process of the first stage of the heterogeneous convolutional network is as follows: ； ； in This represents the features of the initial feature map after the first stage of HGB_down1 processing. ( ) indicates a heterogeneous gated convolution operation with a stride of 2 in the first stage; express Features after the first stage of HGB_base1 processing ( ) indicates a heterogeneous gated convolution operation with a stride of 1 in the first stage. Representation of feature map The number of channels; The computation process of the second-stage heterogeneous convolutional network is as follows: ； ； in express Features after the second stage of HGB_down2 processing ( ) indicates a heterogeneous gated convolution operation with a stride of 2 in the second stage; express Features after the second stage of HGB_base2 processing ( ) indicates a heterogeneous gated convolution operation with a stride of 1 in the second stage. Representation of feature map The number of channels; The computation process of the third-stage heterogeneous convolutional network is as follows: ； ； in express Features after the third stage of HGB_down3 processing ( ) indicates a heterogeneous gated convolution operation with a stride of 2 in the third stage; express Features after the third stage of HGB_base3 processing ( ) indicates a heterogeneous gated convolution operation with a stride of 1 in the third stage. Representation of feature map The number of channels; Finally, the output features of the third-stage heterogeneous convolutional network are fed into the global feature aggregation and rPPG regression layer, and then sequentially pass through 1×1 convolution, batch normalization, ELU activation function, global average pooling, and fully connected layer operations to obtain continuous rPPG signals. .

4. The lightweight identity authentication method based on remote photoplethysmography (PPG) signals according to claim 1, characterized in that, In step 2, the input feature map of input HGB is defined as follows: ; in The number of channels in the input feature map. The height of the input feature map, The width of the input feature map; The processing flow of the large kernel sensing path is as follows: Input feature map First, the channel dimension is compressed using 1×1 convolution, and then batch normalization and ReLU activation are performed. Then, 7×7 depthwise separable convolutions are used to capture long-range spatial dependencies; Finally, a dynamic weight map is generated sequentially through batch normalization, ReLU activation, and 1×1 convolution mapping. The calculation formula is as follows: ； in This represents the number of feature channels output by each group. Indicates the number of groups. Indicates the height of the dynamic weighted graph. This indicates the width of the dynamic weighted graph. ( ) indicates a 7×7 depth-separable convolution operation; Then the dynamic weight graph The intermediate weighted graph is obtained by averaging along the grouping dimension. The calculation formula is as follows: ； in Indicates the first Group weighting chart; Then, the Sigmoid activation function is used to... Mapping to the [0,1] interval yields a normalized gated weight graph. Its expression is as follows: ;in ( ) represents the Sigmoid activation function; The processing flow for small-core aggregation paths is as follows: For the input feature map Perform 3×3 grouped convolution processing to obtain static features. Its expression is as follows: ;in Indicates the height of the feature map. Indicates the width of the feature map. ( ) indicates a 3×3 grouped convolution operation; Finally, the gating weight graph output by the big kernel perception path is... Static features output by the small kernel aggregation path Element-wise multiplication is used to achieve feature modulation, and the calculation formula is as follows: ; in This indicates the fusion characteristics after modulation. This indicates element-wise multiplication; Modulated fusion features After batch normalization and ReLU activation, the enhanced features after gating are obtained. Its expression is as follows: .

5. The lightweight identity authentication method based on remote photoplethysmography (PPG) signals according to claim 1, characterized in that, In step 2, the processing flow of the rPPG morphological feature enhancement module is as follows: First, receive the continuous rPPG signal output from the rPPG signal estimation module. Then, the continuous rPPG signal is separated using shrinkage peak detection and period segmentation. Divided into A periodic segment; Then, perform the following processing procedure on each periodic segment: First, deep convolution is performed to capture local morphological features in the rPPG signal, and then the feature learning process is smoothed through adaptive regularization. Subsequently, the SE attention mechanism is introduced to adaptively adjust the weights of each channel of the multi-channel feature map output by the deep convolution, thereby obtaining the optimized feature vector of each periodic segment. Next, the optimized feature vectors of the K periodic segments are arranged according to the original occurrence order of their corresponding rPPG signal periods to form a periodic time-series feature sequence. ; in Indicates the first The feature vector corresponding to each rPPG signal period; Subsequently, this periodic-level temporal feature sequence is further processed through depthwise separable convolution to obtain temporal features; Then, the encoder of the lightweight gated loop unit is used for... The temporal features of each periodic segment are encoded to output a set of hidden state vectors. , Indicates the first The hidden state vector of each periodic segment; The set of hidden state vectors is integrated and mapped in terms of features through a fully connected layer, and the final output is... A highly discriminative rPPG morphological feature vector ,in Indicates the first The morphological feature vector of each periodic segment.

6. The lightweight identity authentication method based on remote photoplethysmography (PPG) pulse wave signal according to claim 5, characterized in that, In step 2, the specific process of shrinkage peak detection and period segmentation is as follows: First, the cutoff frequency is adopted as... Zero-phase low-pass filter for rPPG signal Filtering is performed to obtain a smooth signal. ; and on Through a window length of The baseline drift is removed by a moving average, yielding the drift-free signal value, calculated using the following formula: ； in, Indicates the smoothed signal at the 1st The signal value after removing drift at each sampling point Indicates the smoothed signal at the 1st The instantaneous amplitude at each sampling point This indicates the smoothed signal in the summation index variable. The amplitude at the corresponding sampling point, The length of the smooth window; Then calculate First-order difference The calculation formula is as follows: ； And detect The set of candidate contraction peak positions is obtained from the zero-crossing point where the positive peak transitions to the negative peak. ; Subsequently, multiple constraints were used for screening based on amplitude thresholds and physiological intervals; Specifically, the mean amplitude corresponding to each candidate contraction peak is calculated. with standard deviation Retain those that meet the conditions The candidate contraction peaks are identified, and the set of candidate contraction peak positions is obtained after screening. ;in An adjustable threshold coefficient; Next, interval filtering is performed, and calculations are performed. The time interval between adjacent candidate contraction peaks is retained only if it falls within the physiologically reasonable range. The candidate contraction peaks within the range are used to obtain the final set of effective contraction peak locations. ; in The signal sampling rate, Indicates the first A total of 1 effective contraction peak positions were detected. One effective contraction peak, corresponding to One complete rPPG signal cycle; Then, periodic segmentation is performed, using adjacent contraction peaks as boundaries to divide the continuous rPPG signal into segments. The first complete periodic segment, the... One original periodic segment The expression is: ； Each original periodic segment is resampled to a fixed length using linear interpolation. To obtain the set of periodic segments ,in Indicates the first Each periodic segment.

7. The lightweight identity authentication method based on remote photoplethysmography (PPG) signals according to claim 5, characterized in that, In step 2, the specific process of encoding the timing features by the lightweight gated cyclic encoder is as follows: First, the current period segment rPPG feature vector The hidden state vector of the previous period segment Input to encoder; Next, calculate the update gate output. To control the proportion of historical rPPG signal rhythm information transmission and maintain the consistency of rPPG signal rhythm across cycles, the calculation formula is as follows: ; in , These are the input and hidden state weight matrices corresponding to the update gate, respectively. To update the gate's bias vector, ( ) is the Sigmoid activation function; Simultaneously, calculate the reset gate output. This controls whether to reset the historical state, thereby extracting the local waveform changes from the systolic peak to the diastolic end within the current periodic segment. The calculation formula is as follows: ; in , These are the input and hidden state weight matrices corresponding to the reset gate, respectively. To reset the gate's bias vector; Subsequently, the candidate hidden states are calculated. The calculation formula is as follows: ; in , These are the input and hidden state weight matrices corresponding to the candidate hidden states, respectively. Let be the candidate hidden state bias vector. ( ) is the hyperbolic tangent activation function; Then, update the hidden state of the gate pair for the previous period segment. Candidate hidden states of the current periodic segment Adaptive fusion is performed to obtain the hidden state of the current periodic segment, and its calculation formula is as follows: ; in, For the first The hidden state of the output of the i-th period segment, and also as the i-th period segment output. The input state of each periodic segment; After processing all K periodic segments sequentially, the lightweight gated cyclic encoder outputs a set of hidden state vectors. .

8. The lightweight identity authentication method based on remote photoplethysmography (PPG) signals according to claim 1, characterized in that, In step 3, a composite loss function is constructed during model training, which includes signal estimation loss and identity authentication loss. The calculation formula is as follows: ; in Indicates the signal estimation loss. This indicates a loss related to identity verification. , This represents the loss balance coefficient; The signal estimation loss is calculated using the negative Pearson correlation coefficient, and the formula is as follows: ； in To predict the total number of sampling points of the rPPG signal and the real rPPG signal, and The first The model-predicted rPPG signal value and the actual rPPG reference signal value at each sampling point. This represents the mean of the predicted rPPG signal. This represents the mean of the actual rPPG signal; The authentication loss is calculated using cross-entropy loss, and the formula is as follows: ； in Indicates the total number of periodic segments. This indicates that the identity authentication model is for the first... Each periodic segment belongs to the real identity tag The predicted probability, , This represents the total number of identities in the training dataset.

9. The lightweight identity authentication method based on remote photoplethysmography (PPG) signals according to claim 1, characterized in that, In step 2, The biometric authentication module receives the morphological feature vector output by the rPPG morphological feature enhancement module. This module is a classifier composed of a multi-layer fully connected network, which uses the ReLU activation function to introduce non-linearity. After processing the input morphological feature vector, the classifier outputs the matching probability of each preset identity and completes the identity determination based on the highest probability.

10. A lightweight identity authentication system based on remote photoplethysmography (PPG) signals, characterized in that, Includes the following modules: The preprocessing module is used to acquire raw video data and perform preprocessing to build a training dataset; During the preprocessing process, for each frame of the face image in the original video, a spatiotemporal mapping method that fuses RGB and YUV dual color spaces is used to generate a dual-color fused spatiotemporal map corresponding to the video data. And an identity authentication module, used to build a lightweight identity authentication model based on remote photoplethysmography (rPPG) pulse wave signals. This model includes an rPPG signal estimation module, an rPPG morphological feature enhancement module, and a biometric authentication module. The rPPG signal estimation module is used to extract features from a three-stage cascaded heterogeneous convolutional network (HGB) for a dual-color fusion spatiotemporal map, thereby obtaining the remote photoplethysmography (rPPG) signal. Each stage of HGB consists of a downsampled heterogeneous convolutional block HGB_down and a basic heterogeneous convolutional block HGB_base concatenated together; both HGB_down and HGB_base adopt a dual-branch processing path that includes a large kernel perception path and a small kernel aggregation path. The large kernel sensing path captures the global rhythmic features of rPPG signals through depthwise separable convolution; The small kernel aggregation path extracts local morphological features of the rPPG signal through grouped convolution; The output features of the large kernel perception path and the small kernel aggregation path are fused through an adaptive gating mechanism; The rPPG morphological feature enhancement module explicitly models the detailed features in the rPPG signal through periodic segmentation and a lightweight gated cyclic encoder to obtain the morphological feature vector corresponding to the rPPG signal. The biometric authentication module uses a classifier to determine identity based on the morphological feature vector. The model is trained based on the training dataset, and the trained model is used for identity feature matching and authentication.

Images