RPPG image recognition and detection method based on vascular network topology features

Abstract

The invention relates to the technical field of image recognition and detection, in particular to an rPPG image recognition and detection method based on topological features of a blood vessel network. On this basis, non-rigid motion correction based on a vascular topological structure, remote photoplethysmography signal weighted fusion combining topological importance and a time sequence quality index, and acquisition quality confidence evaluation and closed-loop control taking vascular network topological integrity, global time sequence signal periodicity and a signal-to-noise ratio as cores are introduced; then the blood vessel structure features and the dynamic pulse features are used for analyzing vital signs and hemodynamic states together, so that the signal extraction process more closely corresponds to a real blood flow path, and the robustness under the conditions of head raising, speaking, expression change and illumination fluctuation is improved; and the risk of outputting a misjudgment result under the condition that the vascular structure is invisible or the signal quality is poor is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition detection, and particularly relates to a rPPG image recognition detection method based on blood vessel network topological features. BACKGROUND

[0002] At present, remote photoplethysmography technology generally uses a camera to shoot a video of a skin area such as a face, analyzes the subtle periodic changes in color or brightness in the picture, and calculates heart rate, respiration and other vital signs, which are used in remote health monitoring, sub-health assessment, rehabilitation follow-up and other scenarios. Correspondingly, vein imaging related technologies usually use near-infrared and other methods to shoot static vein images of fingers, palms and other parts, and use "what does the blood vessel pattern look like" to represent the blood vessel structure characteristics of individuals, which are mostly used to improve measurement stability or reduce discomfort caused by direct contact between traditional sensors and the skin.

[0003] However, the prior art still has some deficiencies: remote photoplethysmography treats the skin as a whole area, rarely truly recovers and utilizes the direction and bifurcation structure of subcutaneous blood vessels, and when encountering slight head lifting, speaking, expression changes or light changes, the signal is easily disturbed, and the heart rate, respiration and other results are not stable enough, in addition, the existing image quality and signal quality evaluation often only focuses on picture brightness, definition or single-channel signal signal-to-noise ratio, lacks evaluation and closed-loop control of the collection quality from the angle of "whether the blood vessel network is complete and clear, and whether the blood flow pulsation is normally propagated along the blood vessel", and is easy to give vital sign results when the blood vessel structure is invisible or the signal is unreliable, which affects the accuracy and safety of remote monitoring.

[0004] Therefore, a rPPG image recognition detection method based on blood vessel network topological features is proposed. SUMMARY

[0005] In order to overcome the deficiencies of the prior art, the present application provides a rPPG image recognition detection method based on blood vessel network topological features.

[0006] To solve the above technical problems, the present application provides the following technical solutions: A rPPG image recognition detection method based on vascular network topological features, comprising the following steps: S1, collecting a plurality of video images containing the skin area of a target object, performing face or skin area positioning, illumination correction and spatial scale normalization on the plurality of video images, and obtaining a target area for remote photoplethysmography analysis of a monitoring image sequence; S2, performing vascular enhancement processing, binarization and thinning processing on each frame of video image in the monitoring image sequence to obtain a skeleton graph of a subcutaneous vascular network, constructing a vascular network topological graph by taking the skeleton pixels as nodes and the adjacent connected relationship as edges, and determining the main trunk blood vessel branches and bifurcation points in the vascular network topological graph; S3, based on the skeleton graph of the subcutaneous vascular network, performing optical flow tracking or feature point tracking on the skeleton points corresponding to the main trunk blood vessel branches and bifurcation points in adjacent frames of video images, combining the topological constraints of continuous smooth displacement of the skeleton points on the same blood vessel branch and coordinated displacement of the multi-branch displacement of the bifurcation points, calculating the deformation field of the target area, performing non-rigid motion compensation and registration on the monitoring image sequence, and obtaining a topologically constrained aligned monitoring image sequence; S4, in the registered monitoring image sequence, selecting a blood vessel branch that meets the preset topological feature condition as a signal extraction area for remote photoplethysmography analysis according to the vascular network topological graph, calculating a monitoring time sequence signal for remote photoplethysmography analysis of each signal extraction area, and its periodicity, signal-to-noise ratio quality index, and combining at least one of the node degree, main trunk path distance, and topological centrality to determine the weight of each signal extraction area, and performing weighted fusion and topological smooth constraint on each monitoring time sequence signal to obtain a global monitoring time sequence signal; S5, based on at least one of the total length of the extracted blood vessel network branches, the number of bifurcation points, the number of connected components, and the periodicity, signal-to-noise ratio of the global monitoring time sequence signal within a preset time window, calculating the vascular network topological confidence, when the vascular network topological confidence is lower than the acquisition quality threshold, adjusting at least one of the exposure, gain, frame rate and field of view range of the image acquisition device and outputting the acquisition quality prompt information to the user, otherwise when it is higher than the acquisition quality threshold in continuous multiple time windows, confirming that it meets the standard and entering the recognition detection stage; S6, performing frequency domain analysis or time domain analysis on the global monitoring time sequence signal when the acquisition quality meets the standard, obtaining a series of physiological parameters including heart rate and respiration rate, and inputting the physiological parameters and the vascular network topological features into a pre-trained recognition detection model to output an image recognition detection result for the target object.

[0007] As a preferred technical solution of the present application, in step S1, when collecting a plurality of video images containing the skin area of a target object, the image acquisition device at least includes a visible light channel and a near-infrared channel, the frame rate of acquisition is not less than twenty frames per second, and the exposure time and gain parameters are adjusted according to the environmental brightness during the acquisition process, so that the average gray scale of the target area falls within the preset dynamic range.

[0008] As a preferred technical solution of the present invention, step S2, performing vascular enhancement processing on the monitoring image sequence includes: applying multi-scale linear structure enhancement filtering to the target region to obtain an enhanced image, performing threshold segmentation on the enhanced image to obtain vascular candidate regions, using morphological opening and closing operations to remove isolated noise points and fill in small breaks, and then performing thinning operations on the processed binary image to obtain a skeleton map of the subcutaneous vascular network.

[0009] As a preferred embodiment of the present invention, in step S3, when performing optical flow tracking on the skeleton points corresponding to the branches and bifurcation points of the main blood vessel, the displacement vector of the skeleton points and the deformation field of the target region are estimated by minimizing the following objective function: ;in, Represents skeleton points The optical flow observation displacement vector, Represents skeleton points The estimated displacement vector, This represents the set of adjacent skeletal point pairs within the same vascular branch. This represents the set of pairs of skeletal points connecting branches at the bifurcation point. , These are the preset topology constraint coefficients.

[0010] As a preferred embodiment of the present invention, in step S4, when performing weighted fusion of the time-series signals of each remote photoplethysmography (TPM) signal extraction region, for each signal extraction region... Calculate the periodicity consistency index and signal-to-noise ratio index, and calculate the topological importance index of the corresponding branches based on the vascular network topology diagram. After normalizing the above indices, calculate the weights according to the following formula. : ;in, This is a normalized periodic consistency indicator. This is the normalized signal-to-noise ratio metric. As a topological importance indicator, , , To preset non-negative weighting coefficients, This is to sum over all signal extraction regions.

[0011] As a preferred embodiment of the present invention, in step S5, the confidence level of the vascular network topology is calculated as follows: the total length of effective vascular branches is statistically analyzed within a preset time window. Number of branching points Number of vascular network connectivity components Periodic indicators of global monitoring timing signals and signal-to-noise ratio (SNR) Normalizing the above indicators yields , , and , blood vessel network topology confidence satisfy: ; wherein, , , , is a non-negative weight coefficient, is a non-negative penalty coefficient.

[0012] As a preferred technical solution of the present application, when analyzing the global monitoring time sequence signal in step S6, the time sequence signal is divided into multiple time windows, the heart rate and the respiratory rate are extracted by performing frequency domain analysis on each time window respectively, and the rising edge time, the falling edge time and the waveform form of the time domain waveform are analyzed, and the time window with periodic significant abnormality is eliminated or marked, when the physiological parameters and the blood vessel network topology features are input into the recognition detection model in step S6, the constructed feature vector at least includes: branch length distribution statistical value reflecting the blood vessel network structure, bifurcation mode statistical value, connected component statistical value, and heart rate, respiratory rate, periodicity index, signal-to-noise ratio index and waveform form index reflecting dynamic behavior, the feature vector is input into the recognition detection model after normalization processing and dimension mapping, the recognition detection model includes a living body detection sub-model and an identity recognition sub-model, the living body detection sub-model judges whether the target object is a living body based on the blood vessel network topology features and the phase difference distribution of the global monitoring time sequence signal on multiple blood vessel paths, the periodicity index and the signal-to-noise ratio index, and the identity recognition sub-model judges the identity of the target object based on the blood vessel network topology features and the physiological parameters on the premise that the living body detection passes.

[0013] Compared with the prior art, the present application has the following beneficial effects: by explicitly extracting the approximate network shape of subcutaneous blood vessels in the video, and on this basis, introducing non-rigid motion correction based on blood vessel topology structure, combining topology importance and time sequence quality index for remote photoplethysmogram signal weighted fusion, and taking blood vessel network topology integrity and global time sequence signal periodicity and signal-to-noise ratio as the core of acquisition quality confidence evaluation and closed-loop control, and then using blood vessel structure features and dynamic pulse features together to analyze vital signs and hemodynamic state, so that the signal extraction process more closely corresponds to the real blood flow path, improves the robustness under the conditions of looking up, speaking, expression change and light fluctuation, reduces the risk of outputting false results in the case of invisible blood vessel structure or poor signal quality, significantly improves the stability, safety and actual usability of remote photoplethysmogram image detection without increasing additional wearable devices, and is suitable for remote vital sign monitoring and related medical auxiliary evaluation scenes. BRIEF DESCRIPTION OF DRAWINGS

[0014] Figure 1A flow chart of the method of the present application is shown. DETAILED DESCRIPTION

[0015] In order to make the technical means, creative features, purposes and effects of the present application easy to understand, the present application is further described below in combination with specific embodiments, but the following embodiments are only preferred embodiments of the present application, not all. Based on the embodiments in the embodiments, other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the present application.

[0016] Embodiment: As shown in the drawing, a rPPG image recognition detection method based on vascular network topological features comprises the following steps: Figure 1 S1, a plurality of video images containing the skin region of a target object are collected, and face or skin region positioning, illumination correction and spatial scale normalization are performed on the plurality of video images to obtain a target region for remote photoplethysmography analysis monitoring image sequence.

[0017] S2, each frame of video image in the monitoring image sequence is subjected to blood vessel enhancement processing, binarization and thinning processing to obtain a skeleton graph of subcutaneous vascular network, the skeleton pixels are taken as nodes, the adjacent connected relationship is taken as edges to construct a vascular network topological graph, and the main trunk blood vessel branches and bifurcation points are determined in the graph.

[0018] S3, based on the skeleton graph of the subcutaneous vascular network, the skeleton points corresponding to the main trunk blood vessel branches and bifurcation points in adjacent frame video images are subjected to optical flow tracking or feature point tracking, and the topological constraints of continuous smooth displacement of skeleton points on the same blood vessel branch and coordinated displacement of multi-branch of bifurcation points are combined to calculate the deformation field of the target region, to perform non-rigid motion compensation and registration on the monitoring image sequence to obtain a topologically constrained aligned monitoring image sequence.

[0019] S4, in the registered monitoring image sequence, the blood vessel branches meeting the preset topological feature conditions are selected as signal extraction regions for remote photoplethysmography according to the vascular network topological graph, the monitoring time sequence signals of each signal extraction region for remote photoplethysmography analysis are calculated, and the periodicity, signal-to-noise ratio quality indicators thereof are calculated, and at least one of the node degree, the main trunk path distance and the topological centrality is combined to determine the weight of each signal extraction region, to perform weighted fusion and topological smooth constraint on each monitoring time sequence signal to obtain a global monitoring time sequence signal.

[0020] ​S5, based on at least one of the total length of the blood vessel network branch, the number of branch points, the number of connected components, and the periodicity, signal-to-noise ratio of the global monitoring time sequence signal extracted within the preset time window, calculate the blood vessel network topology confidence, when the blood vessel network topology confidence is lower than the acquisition quality threshold, adjust at least one of the exposure, gain, frame frequency and field of view range of the image acquisition device, and output the acquisition quality prompt information to the user, otherwise, if it is higher than the acquisition quality threshold in the continuous multiple time windows, it is confirmed to meet the standard and enters the recognition detection stage.

[0021] S6, frequency domain analysis or time domain analysis is performed on the global monitoring time sequence signal when the acquisition quality meets the standard, a series of physiological parameters including heart rate and respiration rate are obtained, and the physiological parameters and the blood vessel network topology features are input into the pre-trained recognition detection model to output the image recognition detection result for the target object.

[0022] It should be noted that, in the image preprocessing and blood vessel network construction link, by performing blood vessel enhancement, binarization and thinning on the monitoring image sequence, the blood vessel structure originally implied in the subtle changes of skin reflection brightness is made explicit, the skeleton graph of subcutaneous blood vessel network is obtained, and the blood vessel network topology graph is further constructed. Through skeletonization processing, the blood vessel region can be compressed to a single-pixel-wide center line, so that the blood vessel trend, branch, branch point and connectivity relationship are clearly presented in the topology graph. In this way, the subsequent motion estimation, signal extraction and quality evaluation are no longer based on rough skin color region or regular grid, but on image points located in the blood vessel trunk or important branch for processing. The image area used for remote photoplethysmography signal analysis is highly coincided with the real blood flow path, reducing the interference of background skin, uniform texture area and noise area on the signal.

[0023] In addition, the skeleton points in the blood vessel network skeleton graph are taken as the main tracking objects, and the motion of the skeleton points is constrained by the blood vessel network topology structure. The displacement of the skeleton points on the same blood vessel branch in adjacent frames is constrained to be continuous and smooth, and the multiple branches at the branch point are constrained to move coordinately. Based on this topological constraint, the deformation field estimated is more consistent with the actual deformation characteristics of the blood vessels and their surrounding tissues under natural motion, expression change and small posture change. Compared with the overall rigid body compensation based on only the facial contour points or a small number of feature points, this non-rigid motion compensation based on the blood vessel skeleton can more finely align the spatial positions of the same blood vessel at different times, significantly weaken the artifacts caused by head micro-motion and local stretching on the remote photoplethysmography signal, and make the brightness changes at the same position in the time sequence more accurately correspond to the pulsation of the same blood flow path.

[0024] Next, the main branch, bifurcation point position, node degree, main path distance, and topological centrality in the vascular network topology graph are used to identify the more critical vascular part in blood flow transport, and the periodicity and signal-to-noise ratio of the quality indicators of the monitoring time sequence signal of each candidate signal extraction region are calculated to measure the reliability of the region for subsequent physiological parameter estimation. By using the topological importance and signal quality together for weight calculation, the vascular branch with high quality and located in the topological key position plays a dominant role in global signal fusion, while the contribution of short, isolated or easily affected by noise branches is naturally weakened. At the same time, by applying a smoothing constraint between topologically adjacent branches, local abnormal fluctuations can be suppressed, and the pulsation along the entire vascular path remains consistent in time. Such a fusion method has obvious advantages in signal-to-noise ratio, waveform stability, and the ability to suppress local noise compared to traditional regional averaging or simple block-level weighting, providing a more reliable global monitoring time sequence signal for subsequent frequency domain analysis and identification modeling.

[0025] By statistically analyzing the total length of the vascular branches, the number of bifurcation points, the number of connected components, and other topological structure indicators within a certain time window, and combining the periodicity and signal-to-noise ratio of the global monitoring time sequence signal, the vascular network topology confidence is obtained. This confidence reflects whether the vascular structure is complete and clear under the current acquisition conditions, and whether the time sequence signal supported by the structure has a stable physiological rhythm. When the vascular network topology confidence is lower than the acquisition quality threshold, the system does not directly enter the identification stage, but adjusts the exposure, gain, frame rate, field of view, and other acquisition parameters, or outputs posture, distance, and lighting adjustment suggestions to the user to improve the degree of vascular exposure and signal conditions. Only when the vascular network topology confidence is stable and exceeds the threshold in multiple consecutive time windows, the acquisition quality is determined to be up to standard and the identification and detection process is started. This closed-loop mechanism has stronger pertinence and discrimination ability compared to the method of judging whether the acquisition is successful only by overall brightness, contrast, or single signal quality indicator, and can effectively avoid outputting identification results in cases of severe vascular structure loss or highly unstable signal, thereby reducing the false detection rate and improving the effective acquisition efficiency.

[0026] Finally, in the final identification detection stage, the physiological parameters such as heart rate and respiratory rate obtained based on the global monitoring time sequence signal are taken as the composite input features together with the vascular network topological features to the identification detection model, the vascular network topological features reflect the geometric difference and connected mode of the individual subcutaneous vascular distribution, have relatively stable individual characteristics, and the monitoring time sequence signal reflects the actual hemodynamic behavior on the vascular network, has real-time and living body properties, by jointly modeling the structural features and dynamic features, the existence of real blood flow pulsation conforming to the physiological law can be verified, and the individual difference of the vascular topology and waveform form is used for identity distinction, compared with the scheme of simply relying on static vein texture or single physiological parameter, the living body judgment ability and identification accuracy are improved, and the image identification detection is more safe and reliable in the remote photoplethysmography scene.

[0027] In step S1, when collecting multiple frames of video images containing the skin region of the target object, the image acquisition device at least includes a visible light channel and a near-infrared channel, the frame frequency is not less than 20 frames per second, and the exposure time and gain parameters are adjusted according to the ambient brightness during the acquisition process, so that the average gray value of the target region falls within the preset dynamic range.

[0028] It should be noted that the image acquisition channel, frame frequency and gray dynamic range are limited and constrained in the acquisition link, and the acquisition of the remote photoplethysmography signal essentially depends on the time sampling of the periodic change of subcutaneous blood flow. If the sampling frame frequency is too low, the number of sampling points obtained in one cardiac cycle will be too small, and the accuracy of frequency domain analysis and periodic analysis will be significantly reduced. At the same time, it is not conducive to distinguish the components in the similar frequency band such as heart rate and respiratory rate. The visibility of subcutaneous vascular structure in the image is highly related to the imaging gray range. If the exposure time and gain are too small, the overall image will be dark and the vascular details will be submerged in noise. If they are too large, the high-light area will be saturated, resulting in a decrease in vascular contrast. Therefore, by selecting the visible light and near-infrared channels, ensuring that the frame frequency is not less than 20 frames per second, and dynamically adjusting the exposure and gain according to the ambient brightness, the average gray value of the target region can be stabilized within the preset range. This can provide a quality controllable and condition stable input in the time and gray dimensions for the subsequent vascular enhancement, skeleton extraction and time sequence signal analysis, thereby reducing the influence of the fluctuation at the acquisition end on the overall scheme performance.

[0029] In step S2, the vascular enhancement processing of the monitoring image sequence includes: applying a multi-scale linear structure enhancement filter to the target region to obtain an enhanced image, performing threshold segmentation on the enhanced image to obtain a vascular candidate region, removing isolated noise points and filling small cracks by using morphological opening and closing operations, and performing thinning operation on the processed binary image to obtain a skeleton map of the subcutaneous vascular network.

[0030] It should be noted that in the image preprocessing and vascular network construction stages, a clear processing flow is formed by cascading multi-scale linear structure enhancement filtering, threshold segmentation, morphological operations, and thinning operations. Blood vessels in remote photoplethysmography (LPG) images typically appear as slender, weak-contrast structures extending along a certain direction. Direct segmentation on the original image is easily affected by skin texture, noise, and uneven illumination, leading to vascular breakage or adhesion. Multi-scale linear structure enhancement filtering first enhances the slender structures resembling vascular morphology at different scales, suppressing blocky noise and background. Then, threshold segmentation transforms the enhancement results. Based on the binary blood vessel candidate regions, opening and closing operations are used to remove isolated small noise points and fill in small breaks, making the blood vessel regions more consistent with the real blood vessel network in terms of connectivity. Finally, through thinning operations, blood vessel regions with varying widths are shrunk to a single-pixel-wide centerline skeleton, resulting in a clear skeleton map. This provides a clear and well-connected foundation for the subsequent construction of a blood vessel network topology composed of nodes and edges. The technical effect of this process is to significantly improve the stability and accuracy of subcutaneous blood vessel network extraction, so that subsequent motion compensation and signal fusion based on blood vessel topology can be built on a reliable structural representation.

[0031] In step S3, when performing optical flow tracking on the skeleton points corresponding to the branches and bifurcation points of the main blood vessel, the displacement vector of the skeleton points and the deformation field of the target region are estimated by minimizing the following objective function: .

[0032] in, Represents skeleton points The optical flow observation displacement vector, Represents skeleton points The estimated displacement vector, This represents the set of adjacent skeletal point pairs within the same vascular branch. This represents the set of pairs of skeletal points connecting branches at the bifurcation point. , These are the preset topology constraint coefficients.

[0033] It should be noted that in the motion compensation and image registration stages, an objective function is constructed that includes optical flow observation terms and topological constraint terms. To estimate the skeleton point displacement vector and the deformation field of the target region, the skeleton point observation displacement vector obtained solely by the local window optical flow algorithm is not sufficient. Larger errors are likely to occur in areas with indistinct textures or high noise levels. Since a blood vessel is a continuous physical structure, the actual displacements of adjacent positions should be spatially continuous, and the multiple branches at bifurcation points should be coordinated and consistent. The objective function... The first term in the equation is used to constrain the estimation of displacement. The second term is to punish the displacement difference between adjacent skeleton points in the same blood vessel branch, so as to realize the smoothing along the branch direction, and the third term is to constrain the displacement difference between the connected branches at the branch point, so as to avoid the tearing or interlacing in the branch area which does not conform to the physiological anatomy structure. Through the minimization solving of the objective function, the prior of the blood vessel topology structure can be introduced while the image observation information is used, so that the obtained deformation field is more consistent with the actual motion state of the blood vessel and the surrounding tissue. In the registered remote photoplethysmogram image, the spatial positions of the same blood vessel at different time frames are more accurately aligned, the interference of the slight head movement, expression change and local non-rigid deformation on the time series signal is significantly inhibited, and a more stable basis is provided for subsequent signal extraction and analysis.

[0034] In step S4, when the time series signals of each remote photoplethysmogram signal extraction area are weighted and fused, the periodic consistency index and the signal-to-noise ratio index of each signal extraction area are calculated The periodic consistency index and the signal-to-noise ratio index are calculated, and the topology importance index of the corresponding branch is calculated based on the blood vessel network topology graph. After the above indexes are normalized, the weight is calculated according to the following formula .

[0035] Wherein, is the normalized periodic consistency index, is the normalized signal-to-noise ratio index, is the topology importance index, , , is a preset non-negative weight coefficient, is the sum of all signal extraction areas.

[0036] It should be noted that the weighted mechanism combining the quality and the topology importance of the monitoring time series signal of each blood vessel branch is introduced in the signal extraction and fusion link. Different blood vessel branches belong to different positions, local imaging quality, motion residues and other factors, which will lead to the differences in periodicity and signal-to-noise ratio of the monitoring time series signal. At the same time, the roles of each branch in the whole blood vessel network also have the difference between the main stem and the distal end. If all branch signals are simply averaged with equal weight, it is easy to be restricted by the branch with poor quality or unimportant topology.

[0037] The periodic consistency and the signal-to-noise ratio of each signal extraction area are calculated and normalized, and then the topology importance index given by the blood vessel network topology graph is combined to construct a linear combination and obtain the weight after normalization ​The branches that meet the high quality of timing signal and are located in the key position of topology can be highlighted at the time of fusion, and the influence of isolated short branches, structural edge branches or general quality branches on the global result is naturally reduced, so that the global monitoring timing signal obtained is obviously better than the unweighted or single quality index weighted case in terms of signal-to-noise ratio, waveform stability and robustness to local abnormalities, and provides more reliable input for subsequent physiological parameter estimation and identification modeling.

[0038] In step S5, the blood vessel network topology confidence is calculated as follows: the total length of effective blood vessel branches, the number of bifurcation points, the number of connected components of the blood vessel network, the periodicity index of the global monitoring timing signal, and the signal-to-noise ratio index in a preset time window are counted The above indexes are normalized to obtain 、 、 and The blood vessel network topology confidence satisfies: .

[0039] wherein 、 、 、 are non-negative weight coefficients, is a non-negative penalty coefficient.

[0040] It should be noted that the calculation model of the blood vessel network topology confidence is introduced in the acquisition quality evaluation and closed-loop control link, and the data quality in the remote photoplethysmography scene depends not only on the overall brightness of the image or a single signal index, but also on whether the structure is complete and coherent, and whether the periodicity is clear and the signal-to-noise ratio is high.

[0041] The total length of effective branches, the number of bifurcation points, the number of connected components, and the periodicity and signal-to-noise ratio of the global timing signal in a preset time window are counted, and these indexes are normalized and linearly combined according to the preset weight to construct the blood vessel network topology confidence ​​​​Wherein the total length of the branches and the number of bifurcation points reflect the richness of the blood vessels, the number of connected components is used to punish the fragmentation of the network, and the periodicity and signal-to-noise ratio reflect the quality of the signal extracted based on the structure; The confidence score can comprehensively depict the availability of the blood vessel structure and blood flow signal under the current acquisition condition. When it is lower than the threshold, the system can automatically adjust the exposure, gain, frame rate or field of view, or prompt the user to adjust the posture and distance. Only in the condition that the confidence score is continuously up to standard, the recognition stage is entered, thereby forming an acquisition closed-loop control mechanism with the quality of the blood vessel topology as the core, effectively reducing the probability of poor quality data entering the recognition process, and improving the overall reliability of the system.

[0042] In step S6, when analyzing the global monitoring time sequence signal, the time sequence signal is divided into multiple time windows, and the heart rate and respiratory rate are extracted by performing frequency domain analysis on each time window respectively, and the rising edge time, falling edge time and waveform morphology of the time domain waveform are analyzed, and the time window with periodic significant abnormality is eliminated or marked.

[0043] It should be noted that the time windowing frequency domain and time domain joint analysis strategy is adopted when analyzing the global monitoring time sequence signal. The physiological parameters such as heart rate and respiratory rate change slowly over time in actual scenarios. If the frequency domain analysis of the signal in the entire observation period is performed at one time, it is difficult to accurately describe this dynamic change process, and the single global analysis is easily affected by local strong interference or short-time motion artifacts.

[0044] By dividing the global time sequence signal into multiple time windows, the frequency domain analysis is performed on each window to locate the main frequency peak of the heart rate and respiratory rate, and the quality of the pulse wave in the window is judged in combination with the rising edge time, falling edge time and overall waveform morphology of the time domain waveform. The change of physiological parameters can be tracked on the time axis, and the time period with periodic significant abnormality or severely distorted waveform can be identified in time. The abnormal windows are eliminated or marked, which helps to avoid the influence of local interference on the overall recognition result, makes the physiological parameter sequence input into the recognition model more smooth and stable, and enhances the robustness of the scheme in long-time monitoring and complex motion scenarios.

[0045] In step S6, when the physiological parameters and the blood vessel network topology features are input into the recognition detection model, the constructed feature vector at least includes: branch length distribution statistical value reflecting the blood vessel network structure, bifurcation mode statistical value, connected component statistical value and reflecting dynamic behavior heart rate, respiratory rate, periodicity index, signal-to-noise ratio index and waveform morphology index. The feature vector is normalized and dimensionally mapped before being input into the recognition detection model.

[0046] It should be noted that the blood vessel network topological features and the physiological parameters extracted from the monitoring time sequence signals are uniformly constructed as a feature vector input into the recognition detection model in the recognition detection stage. The branch length distribution, bifurcation mode, and connected component statistics of the subcutaneous blood vessel network are strongly different between individuals and relatively constant in a short time, which is an individual feature at the geometric level. The heart rate, respiratory rate, periodic index, signal-to-noise ratio, and waveform morphology obtained based on the global time sequence signal reflect the hemodynamic behavior occurring on the blood vessel network, which has both living body properties and certain individual differences.

[0047] By uniformly encoding the above structural features and dynamic features into a normalized and dimensionally mapped feature vector, both the “whose blood vessel network” and the “current how the blood vessel network beats” information can be used for discrimination in the feature space. Compared with using only static vein texture or a single physiological parameter, the recognition model can capture more rich difference dimensions, thereby significantly improving the identity discrimination degree and sensitivity to counterfeit attacks.

[0048] The recognition detection model includes a living body detection sub-model and an identity recognition sub-model. The living body detection sub-model judges whether the target object is a living body based on the blood vessel network topological features and the phase difference distribution, periodic index, and signal-to-noise ratio index of the global monitoring time sequence signal on multiple blood vessel paths. The identity recognition sub-model judges the identity of the target object based on the blood vessel network topological features and physiological parameters under the premise that the living body detection passes.

[0049] It should be noted that the living body detection sub-model and the identity recognition sub-model are separated in the recognition detection model. The living body detection sub-model takes the blood vessel network topological features, the phase difference distribution, the periodic index, and the signal-to-noise ratio on multiple blood vessel paths as inputs, and focuses on checking whether the extracted pulsatile signal conforms to the physiological rules of real blood flow propagation in space and time, such as whether the phase difference between adjacent positions along the same blood vessel path is within a reasonable range, whether the overall periodicity is stable, and whether the signal-to-noise ratio meets the living body determination threshold, thereby eliminating counterfeit signals such as video replay and screen flipping that lack real blood flow propagation characteristics.

[0050] On this basis, the identity recognition sub-model only judges the identity of the target object that has passed the living body detection using the feature vector composed of the blood vessel network topological features and the physiological parameters, avoiding direct identity recognition in the case of doubtful living body. The technical effect of this phased structure is to form a secure link of “first verifying as a real person, then judging who it is”, which significantly reduces the success rate of counterfeit attacks and improves the identity recognition accuracy and stability on real living body data.

[0051] The embodiments of the present application are described in detail above with reference to the accompanying drawings, but the present application is not limited to the embodiments, and various changes can be made by those skilled in the art within the scope of knowledge acquired from the present disclosure, without departing from the spirit of the present application.

Claims

1. A method for rPPG image recognition detection based on vascular network topological features, characterized in that, The method comprises the following steps: S1, collecting a plurality of video images containing a skin region of a target object, performing face or skin region positioning, illumination correction and spatial scale normalization on the plurality of video images to obtain a monitoring image sequence for remote photoplethysmography analysis of the target region; S2, performing blood vessel enhancement processing, binarization and thinning processing on each frame of video image in the monitoring image sequence to obtain a skeleton map of a subcutaneous blood vessel network, constructing a blood vessel network topology graph by taking the skeleton pixels as nodes and the adjacent connected relationship as edges, and determining the main blood vessel branches and bifurcation points in the topology graph; S3, based on the skeleton map of the subcutaneous blood vessel network, performing optical flow tracking or feature point tracking on the skeleton points corresponding to the main blood vessel branches and bifurcation points in adjacent frame video images, combining the topological constraints of continuous and smooth displacement of the skeleton points on the same blood vessel branch and coordinated displacement of the multi-branch of the bifurcation points, calculating the deformation field of the target region, performing non-rigid motion compensation and registration on the monitoring image sequence to obtain a topology-constrained aligned monitoring image sequence; S4, in the registered monitoring image sequence, selecting a blood vessel branch that meets a preset topological feature condition as a signal extraction region for remote photoplethysmography according to the blood vessel network topology graph, calculating a monitoring time sequence signal for remote photoplethysmography analysis of each signal extraction region, its periodicity, signal-to-noise ratio quality index, and combining at least one of the node degree, main path distance, and topological centrality to determine the weight of each signal extraction region, and performing weighted fusion and topological smoothing constraint on each monitoring time sequence signal to obtain a global monitoring time sequence signal; S5, based on at least one of the total length of the blood vessel network branches, the number of bifurcation points, the number of connected components, and the periodicity, signal-to-noise ratio of the global monitoring time sequence signal extracted within a preset time window, calculating the blood vessel network topology confidence, when the blood vessel network topology confidence is lower than the acquisition quality threshold, adjusting at least one of the exposure, gain, frame rate and field of view range of the image acquisition device and outputting acquisition quality prompt information to the user, otherwise when it is higher than the acquisition quality threshold in continuous multiple time windows, confirming that it meets the standard and entering the identification and detection stage; S6, performing frequency domain analysis or time domain analysis on the global monitoring time sequence signal when the acquisition quality meets the standard to obtain a series of physiological parameters including heart rate and respiration rate, and inputting the physiological parameters and the blood vessel network topological features into a pre-trained identification and detection model to output an image identification and detection result for the target object.

2. The rPPG image recognition detection method based on vascular network topological features according to claim 1, characterized in that, In the step S1, when collecting a plurality of video images containing a skin region of a target object, the image acquisition device at least includes a visible light channel and a near-infrared channel, the frame rate is not less than twenty frames per second, and the exposure time and gain parameters are adjusted according to the ambient brightness during the collection process, so that the average gray level of the target region falls within a preset dynamic range.

3. The rPPG image recognition detection method based on vascular network topological features according to claim 2, characterized in that, The step S2 includes: applying a multi-scale line structure enhancement filter to the target region to obtain an enhanced image, performing threshold segmentation on the enhanced image to obtain a blood vessel candidate region, removing isolated noise points and filling small fractures by using morphological opening and closing operations, and performing thinning operation on the processed binary image to obtain a skeleton map of the subcutaneous blood vessel network.

4. The rPPG image recognition detection method based on vascular network topological features according to claim 3, characterized in that, In the step S3, when tracking the skeleton points corresponding to the main blood vessel branches and bifurcation points by using optical flow, the displacement vector of the skeleton points and the deformation field of the target region are estimated by minimizing the following objective function: ； wherein, denotes the optical flow observed displacement vector of a skeleton point denotes the estimated displacement vector of a skeleton point denotes the set of pairs of adjacent skeleton points within the same vessel branch, denotes the set of pairs of connected skeleton points at a bifurcation point, denotes the set of pairs of adjacent skeleton points within the same vessel branch, denotes the set of pairs of connected skeleton points at a bifurcation point, , is a preset topological constraint coefficient.

5. The rPPG image recognition detection method based on vascular network topological features according to claim 4, characterized in that, In the step S4, when the time sequence signals of each remote photoplethysmography signal extraction region are fused by weighting, the time sequence signals of each signal extraction region are fused by weighting The cycle consistency index and the signal-to-noise ratio index are calculated, and the topological importance index of the corresponding branch is calculated based on the blood vessel network topological graph. After the above indexes are normalized, the weight is calculated according to the following formula : ； wherein, is a normalized periodicity consistency index, is a normalized signal-to-noise ratio index, is a topology importance index, , , is a preset non-negative weight coefficient, is a sum of all signal extraction regions.

6. The rPPG image recognition detection method based on vascular network topological features according to claim 5, characterized in that, In the step S5, the blood vessel network topology confidence is calculated as follows: the total length of effective blood vessel branches is counted within a preset time window , the number of bifurcation points , the number of connected components of the blood vessel network , the periodicity index of the global monitoring timing signal , and the signal-to-noise ratio index . The above indexes are normalized to obtain , , and , and the blood vessel network topology confidence satisfies: ； wherein, , , , are non-negative weight coefficients, is a non-negative penalty coefficient.

7. The rPPG image recognition detection method based on vascular network topological features according to claim 6, characterized in that, In the step S6, when analyzing the global monitoring time series signal, the time series signal is divided into multiple time windows, the heart rate and the respiratory rate are extracted by performing frequency domain analysis on each time window, and the rising edge time, the falling edge time and the waveform shape of the time domain waveform are analyzed, and the time windows with periodic significant abnormalities are removed or marked.

8. The rPPG image recognition detection method based on vascular network topological features according to claim 7, characterized in that, In the step S6, when inputting the physiological parameters and the blood vessel network topology features into the recognition detection model, the constructed feature vector at least includes: branch length distribution statistical values, bifurcation mode statistical values, connected component statistical values reflecting the structure of the blood vessel network, and heart rate, respiratory rate, periodicity index, signal-to-noise ratio index and waveform shape index reflecting the dynamic behavior, and the feature vector is input into the recognition detection model after normalization processing and dimensionality mapping.

9. The rPPG image recognition detection method based on vascular network topological features according to claim 8, characterized in that, The recognition detection model includes a living body detection sub-model and an identity recognition sub-model, the living body detection sub-model judges whether the target object is a living body based on the blood vessel network topology features and the phase difference distribution, the periodicity index and the signal-to-noise ratio index of the global monitoring time series signal on multiple blood vessel paths, and the identity recognition sub-model performs identity recognition on the target object based on the blood vessel network topology features and the physiological parameters on the premise that the living body detection passes.

Images