Newborn identity authentication system based on footprints recognition

Abstract

The application discloses a new-born baby identity authentication system based on foot sole line recognition and belongs to the technical field of new-born baby identity authentication. The system comprises an image acquisition module which acquires a new-born baby foot sole original pressure image sequence and foot sole three-dimensional point cloud data; a primitive decomposition module which decomposes the original pressure image sequence into static foot sole texture primitives and dynamic pressure propagation primitives; a line extraction module which extracts a foot sole coarse line distribution skeleton according to the static foot sole texture primitives and extracts a foot sole fine line line offset trajectory in a pressure deformation process according to the dynamic pressure propagation primitives; a deformation correction module which determines foot sole region curvature change gradients according to the foot sole three-dimensional point cloud data, corrects the line offset trajectory by using the gradients and generates a foot sole line topology structure; and a hierarchical matching module which performs hierarchical matching on the coarse line distribution skeleton and the line topology structure to obtain a foot sole line identity code of the new-born baby. The system can completely characterize and accurately recognize the foot sole lines of the new-born baby.

Description

Technical Field

[0001] This invention relates to the field of newborn identification technology, specifically a newborn identification system based on footprint recognition. Background Technology

[0002] Newborn identification is a necessary step in maternity ward management, infant theft prevention, and paternity testing. Current technologies typically rely on wristbands, footprint rubbings, or conventional fingerprint collection to confirm newborn identity. Wristbands are susceptible to detachment or tampering; footprint rubbings are difficult to control due to the newborn's feet often curling or twisting during collection, and rubbings only record the rough outline and a few deep lines of the sole, failing to reflect the complete morphology of the skin under pressure. Existing plantar crease analysis techniques mostly extract crease distribution from static two-dimensional images, treating the sole as a rigid plane and ignoring the non-uniform elastic deformation of the soft tissue during pressure collection, which disturbs the position and direction of the creases. This leads to distortion of crease features, making it difficult to stably reproduce the same plantar crease from different collections, resulting in low identification accuracy in the newborn population.

[0003] Newborns have thin plantar skin and an incompletely developed stratum corneum, resulting in significant differences in the mechanical response characteristics of coarse and fine plantar creases. Under pressure, coarse creases shift less, while the shift and trajectory of fine creases are strongly modulated by changes in the curvature and tissue hardness of different areas of the sole, forming a dynamic crease shift field. Simply removing the dynamic deformation information under pressure and relying solely on the static crease distribution loses a crucial distinguishing factor reflecting the plantar mechanical response characteristics. Directly using deformed crease images for matching leads to unreliable feature position and angle information due to the uncertainty of deformation. Solving the problem of dynamic deformation compensation of plantar creases under pressure and jointly encoding the overall distribution framework of coarse creases with the deformation-corrected topology of fine creases are the technical challenges that need to be addressed in constructing reliable newborn plantar crease identification. Summary of the Invention

[0004] The purpose of this invention is to provide a newborn identification system based on plantar crease recognition, addressing the problem that existing technologies cannot correct for the non-uniform dynamic deformation of a newborn's foot under pressure, leading to unstable plantar crease feature extraction and unreliable identification. It also solves the problem that coarse and fine creases are treated uniformly during pressure, failing to utilize the offset trajectory information of fine creases during dynamic deformation, thus limiting the distinguishability of plantar creases. This system performs primitive decomposition on the pressure image sequence to separate static texture and dynamic pressure propagation information. It then combines the gradient of the curvature change of the three-dimensional surface of the foot to perform inverse deformation compensation for the pressure offset trajectory of fine creases. Finally, it performs hierarchical matching between the coarse crease skeleton and the compensated crease topology to generate a stable plantar crease identification code.

[0005] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a newborn identification system based on plantar crease recognition, including an image acquisition module, a primitive decomposition module, a crease extraction module, a deformation correction module, and a hierarchical matching module. The image acquisition module acquires the original pressure image sequence and three-dimensional point cloud data of the newborn's sole. The primitive decomposition module decomposes the original pressure image sequence into static plantar texture primitives and dynamic pressure propagation primitives. The crease extraction module extracts the coarse crease distribution skeleton of the sole based on the static plantar texture primitives, and extracts the crease offset trajectory of the fine creases during pressure deformation based on the dynamic pressure propagation primitives. The deformation correction module determines the curvature change gradient of each region of the sole based on the three-dimensional point cloud data, and uses the curvature change gradient to correct the crease offset trajectory, generating a dynamically deformed corrected plantar crease topology. The hierarchical matching module performs hierarchical matching between the coarse crease distribution skeleton and the plantar crease topology to obtain the newborn's plantar crease identification code.

[0006] As a technical solution of the present invention, the primitive decomposition module selects the first frame pressure image from the original pressure image sequence as the reference pressure image, performs spatial domain filtering on the reference pressure image, and extracts the texture components whose pressure values ​​change with spatial position but whose change with inter-frame time is less than a preset threshold as the static plantar texture primitives; and subtracts the mapping components of the static plantar texture primitives in each frame from the original pressure image sequence to obtain a high-frequency residual sequence of pressure values ​​changing with inter-frame time, and performs temporal envelope extraction on the high-frequency residual sequence to obtain the dynamic pressure propagation primitives. Through the above primitive decomposition, the inherent texture structure of the newborn's sole and the dynamic pressure response caused by pressure deformation can be effectively distinguished, providing a basis for the subsequent separate extraction of coarse and fine textures.

[0007] Preferably, the texture extraction module performs grayscale gradient calculation on the static plantar texture primitives to obtain a gradient magnitude map and a gradient direction map of the plantar texture; it traces local maxima points in the gradient magnitude map along the main direction of the gradient direction map, connects adjacent local maxima points to form texture ridges, and refines the texture ridges to obtain the coarse texture distribution skeleton of the plantar texture. Simultaneously, the texture extraction module identifies the texture edge points of the fine textures of the plantar texture from the dynamic pressure propagation primitives, records the pixel coordinate change sequence of each texture edge point in consecutive frames of the original pressure image sequence; it performs trajectory fitting on the pixel coordinate change sequence to generate a displacement vector field of each texture edge point during the pressure deformation process, and uses the displacement vector field as the texture offset trajectory. Thus, the system can obtain both a coarse texture skeleton reflecting the overall texture direction of the plantar texture and capture the dynamic offset characteristics of fine textures under pressure.

[0008] As a further improvement of the present invention, the deformation correction module divides the three-dimensional point cloud data of the sole into heel region, arch region, and forefoot region according to the anatomical structure of the sole. It then performs surface fitting on the point cloud data within each region to obtain the local surface equations for each region. The module calculates the Gaussian curvature and mean curvature of each local surface equation at the corresponding region boundary, and performs a difference operation on the Gaussian curvature and mean curvature along the region boundary to obtain the curvature change gradient. When correcting the texture offset trajectory using the curvature change gradient, the deformation correction module projects the texture offset trajectory onto the three-dimensional spatial coordinates corresponding to the three-dimensional point cloud data of the sole to obtain a three-dimensional offset trajectory. Based on the curvature change gradient, it determines the strain tensor field of each region of the sole during the compression deformation process, and uses the strain tensor field to perform inverse deformation compensation on the three-dimensional offset trajectory to obtain a corrected texture spatial coordinate sequence. Finally, it topologically connects the corrected texture spatial coordinate sequence according to the connectivity of the sole texture to generate the sole texture topology structure. This approach incorporates information on the shape and curvature changes of the sole to accurately compensate for the positional shift of fine lines caused by skin elastic deformation, ensuring that the extracted line topology reflects the true line distribution of a newborn's sole under natural, pressure-free conditions.

[0009] Preferably, the hierarchical matching module performs first-level feature encoding on the coarse foot texture distribution skeleton, extracting the coordinates of the main bifurcation points and main endpoints of the coarse textures to generate a first-level feature vector; it performs second-level feature encoding on the topological structure of the foot texture, extracting the relative spacing and relative direction angles of the corrected fine textures to generate a second-level feature vector; and using the coordinates of the main bifurcation points in the first-level feature vector as anchor points, it registers the relative spacing and relative direction angles of the fine textures in the second-level feature vectors to a local coordinate system with the anchor points as the origin; it concatenates the registered first-level feature vectors and the second-level feature vectors into a joint feature sequence, and performs quantization encoding on the joint feature sequence to generate a fixed-length binary sequence as the foot texture identification code. Through hierarchical matching and joint encoding of coarse and fine texture features, the system can integrate the global texture architecture and local fine texture information of the foot to generate a unique and stable identification code, improving the accuracy and reliability of newborn identification.

[0010] In this invention, the image acquisition module controls the pressure acquisition device to apply increasing pressure to the sole of the newborn's foot at a preset sampling frequency and simultaneously acquires multiple frames of sole pressure distribution images. These multiple frames are arranged chronologically to form the original pressure image sequence. Simultaneously, the structured light scanning device projects a structured light pattern while applying the increasing pressure, acquiring structured light deformation images of the newborn's sole. Phase calculation is performed on these structured light deformation images to generate the three-dimensional point cloud data of the sole. Preferably, the acquisition time interval between two adjacent frames in the original pressure image sequence is 0.3 to 0.5 times the time required for the newborn's sole skin to complete one elastic recovery. This ensures a complete record of the continuous changes in pressure distribution and texture displacement during skin deformation and rebound, providing a sufficient data foundation for dynamic feature extraction and deformation correction. Through the collaborative work of these multiple modules, the system of this invention can simultaneously acquire static texture, dynamic pressure propagation, and three-dimensional morphological information from the newborn's sole, thereby eliminating the influence of pressure deformation on the texture and ultimately generating a robust and accurate sole texture identification code.

[0011] The technical effects and advantages provided by the present invention in the above technical solution are as follows: The original pressure image sequence is decomposed into static plantar texture primitives and dynamic pressure propagation primitives. The coarse texture distribution skeleton is extracted from the static primitives, and the fine texture offset trajectories between consecutive frames are extracted from the dynamic primitives. This decomposition method separates the coarse textures, which exhibit almost no positional shift during pressure, from the dynamic responses of the fine textures, which are significantly modulated by soft tissue elastic deformation. The coarse texture skeleton provides a relatively constant spatial reference framework for the neonatal plantar texture, while the fine texture offset trajectories capture the continuous positional changes of each fine texture edge point during pressure application. Compared to directly performing single texture analysis on the entire pressure image, this separation process avoids the interference of fine texture deformation information on the positioning accuracy of coarse textures. Simultaneously, it transforms the skin mechanical response behavior of neonates during pressure loading into a coded feature sequence, preserving the personalized deformation patterns of the plantar soft tissue under dynamic load. This ensures that even when the static coarse texture distribution is similar, the dynamic texture offset trajectories of different neonates remain distinguishable.

[0012] Based on 3D point cloud data of the foot, the curvature gradients of the heel, arch, and forefoot regions are determined. These curvature gradients are then used to correct the tread pattern offset trajectory, generating a dynamically deformed topological structure of the foot tread pattern. The 3D surface of the foot exhibits abrupt curvature changes at the boundaries of different regions. Under pressure, the direction and amplitude of tissue strain vary significantly across regions with different curvatures. The offset trajectory of the fine tread pattern incorporates non-uniform deformation components determined by the surface geometry. By projecting the offset trajectory into 3D space and calculating the strain tensor field of each region, and using this tensor field to perform inverse deformation compensation on the offset trajectory, the drastic trajectory changes introduced by individual surface differences such as arch curvature and forefoot curvature can be eliminated. This restores the spatial position of the corrected fine tread pattern to its relative geometric relationship before pressure application. This process removes the modulation effect of the macroscopic surface morphology of the foot on the fine tread pattern offset from the original trajectory, achieving quantitative correction of the tread pattern position without losing dynamic response information. The spatial consistency of the corrected texture topology was improved across multiple acquisitions. The relative spacing and orientation angle of fine lines obtained from different pressure acquisitions of the same newborn tended to be stable, and the repeatability of texture feature acquisition was improved. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of a newborn identification system based on footprint recognition; Figure 2 This is a flowchart of the generation process for the topology of the sole pattern after dynamic deformation correction. Figure 3 This is a flowchart of the synchronous acquisition process of neonatal plantar pressure image sequence and 3D point cloud; Figure 4 It is the time-domain variation curve of static plantar texture primitives and dynamic pressure propagation primitives. Detailed Implementation

[0014] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this invention.

[0015] See Figure 1This invention provides a newborn identification system based on plantar crease recognition, comprising an image acquisition module, a primitive decomposition module, a crease extraction module, a deformation correction module, and a hierarchical matching module. The image acquisition module acquires the original pressure image sequence and three-dimensional point cloud data of the newborn's sole. The primitive decomposition module decomposes the original pressure image sequence into static plantar texture primitives and dynamic pressure propagation primitives. The crease extraction module extracts the coarse crease distribution skeleton based on the static plantar texture primitives and extracts the crease offset trajectory of the fine creases during pressure deformation based on the dynamic pressure propagation primitives. The deformation correction module determines the curvature change gradient of each region of the sole based on the three-dimensional point cloud data, and uses the curvature change gradient to correct the crease offset trajectory, generating a dynamically deformed corrected plantar crease topology. The hierarchical matching module performs hierarchical matching between the coarse crease distribution skeleton and the plantar crease topology to obtain the newborn's plantar crease identification code.

[0016] Example 1: In the specific implementation process, the original pressure image sequence is decomposed into static plantar texture primitives and dynamic pressure propagation primitives in the following way.

[0017] The first frame of pressure image is selected from the original pressure image sequence as the reference pressure image. The first frame of pressure image acquired by the image acquisition module at the initial moment of applying incremental pressure is determined as the reference pressure image. At this time, the pressure on the sole of the foot is minimal, the sole shape is close to a natural relaxed state, and the pressure value distribution mainly reflects the inherent texture characteristics of the sole.

[0018] Spatial domain filtering is performed on the reference pressure image. Spatial domain filtering involves convolving a Gaussian kernel with the reference pressure image. The Gaussian kernel window size is set to 5 pixels × 5 pixels, and its standard deviation parameter is set to 1.0 pixel. This standard deviation parameter is selected based on the spatial resolution of the original pressure image sequence and the minimum feature scale of the plantar texture, thus smoothing random noise while preserving the pressure value's variation with spatial location. After spatial domain filtering, high-frequency random fluctuations in the reference pressure image are suppressed.

[0019] Components whose pressure values ​​change less than a preset threshold over time with respect to the spatially filtered reference pressure image are extracted and used as the static plantar texture primitives. For each pixel in the spatially filtered reference pressure image, the preceding consecutive pixels in the original pressure image sequence are... Extract the pressure value of the pixel frame by frame within the frame. The preset number of frames to analyze. The value is 5. For each pixel, calculate the value in consecutive... The pressure change amplitude within a frame is defined as continuous. The difference between the maximum and minimum pressure values ​​in a frame. Preset threshold. Calculated using the following formula:

[0020] in, This indicates that all pixels in the spatial domain filtered reference pressure image are in continuous... The mean of the intra-frame pressure variation amplitude, This indicates that all pixels in the spatial domain filtered reference pressure image are in continuous... The standard deviation of the intra-frame pressure variation amplitude This is the threshold adjustment coefficient. The value is 3, based on the normal distribution. In principle, the range of change is within Pixels within the specified range are considered to have pressure value fluctuations that are noise or minor elastic recovery, not exceeding the normal range of static texture fluctuations. The pressure variation amplitude in the spatially filtered baseline pressure image is less than a preset threshold. The pressure value of each pixel is preserved, and pressure changes greater than or equal to a preset threshold are allowed. The pressure value of each pixel is set to zero to obtain the static foot texture primitive. The static foot texture primitive reflects the basically stable texture pressure distribution of the foot in the time dimension during the static pressure stage.

[0021] The high-frequency residual sequence is obtained by subtracting the mapping component of the static plantar texture primitive in the corresponding frame from each frame of the original pressure image sequence. The mapping component of the static plantar texture primitive in any frame of the pressure image is the static plantar texture primitive image itself, because the static plantar texture primitive represents a pressure pattern that does not change with inter-frame time, and its spatial distribution and amplitude in each frame are the same as the static plantar texture primitive. The subtraction operation is performed frame by frame in the original pressure image sequence. The pressure value of each pixel in each frame of the pressure image is subtracted from the pressure value at the corresponding pixel coordinate in the static plantar texture primitive. If the pressure value at a certain pixel coordinate in the static plantar texture primitive is zero, the subtraction operation at that pixel coordinate is not performed. After subtraction, a high-frequency residual image is generated for each frame. All high-frequency residual images are arranged in chronological order to form a high-frequency residual sequence. The high-frequency residual sequence contains the high-frequency fluctuation component of the pressure value changing with inter-frame time.

[0022] The high-frequency residual sequence is subjected to pixel-by-pixel temporal envelope extraction. For each fixed pixel coordinate position in the high-frequency residual sequence, the high-frequency residual values ​​in all frames are extracted along the time axis to form a one-dimensional high-frequency residual signal. A Hilbert transform is performed on each high-frequency residual signal to construct an analytical signal. The Hilbert transform is performed through frequency domain processing: first, a fast Fourier transform is performed on the high-frequency residual signal to obtain the spectrum; the amplitude values ​​of the negative frequency components in the spectrum are set to zero, and the amplitude values ​​of the positive frequency components in the spectrum are doubled; then, an inverse fast Fourier transform is performed on the processed spectrum. The inverse transform result is a complex sequence, where the real part of the complex sequence is equal to the original high-frequency residual signal, and the imaginary part of the complex sequence is the Hilbert transform result of the original high-frequency residual signal. The magnitude value is calculated point-by-point for the complex sequence, and the magnitude value sequence is the temporal envelope of the pressure fluctuation at that pixel coordinate position. All pixel coordinate positions are traversed to obtain the amplitude profile of each pixel during the pressure change process; the amplitude profiles of all pixel positions constitute the dynamic pressure propagation primitive.

[0023] In some implementations, the continuous Number of frames Based on the preset sampling frequency of the original pressure image sequence and the elastic recoil time of the newborn's plantar skin, the continuous... The frame time span covers 1.2 to 1.8 times the time required for the plantar skin to complete one elastic recoil cycle, in order to fully capture the statistical characteristics of pressure fluctuations during static compression. When the time interval between adjacent frames is 0.4 times the time required for the plantar skin to complete one elastic recoil cycle, A value of 5 can cover the observation period of approximately twice the elastic recovery cycle.

[0024] Optionally, the Gaussian kernel standard deviation of the spatial domain filter can be adjusted between 0.8 and 1.5 pixels based on the actual pressure image sensor noise level, increasing the standard deviation when the noise is high and decreasing the standard deviation when the texture details are fine.

[0025] It is understood that when the mapping component of the static foot texture primitive is subtracted from each frame of the original pressure image sequence, the foot position between each frame in the original pressure image sequence is spatially registered by the fixed clamp of the acquisition device, without the need for additional image registration transformation.

[0026] See Figure 4In the figure, the horizontal axis represents the frame number (time index), ranging from 0 to 100, representing the temporal order of 100 consecutively acquired pressure images in the original pressure image sequence. The vertical axis represents the pressure value, in kilopascals, ranging from 0 to 130 kilopascals. The figure plots four curves: the static plantar texture primitive (blue solid line with dots) obtained by spatial domain filtering of the baseline pressure image; the pressure change of two different pixels (pixel A and pixel B) in the dynamic pressure propagation primitive (orange dashed line and green dotted line with square and triangular dots); and the temporal envelope of the dynamic pressure fluctuation corresponding to pixel A (red dotted line).

[0027] The blue static plantar texture primitive curve shows a basically stable pressure value, which fluctuates slightly around 115 kPa. This indicates that the primitive reflects the static texture pressure distribution during the plantar compression process. The pressure value fluctuates little over time and has no obvious trend change, which is consistent with the definition and extraction principle of static plantar texture primitive in Example 1.

[0028] The orange dynamic pressure propagation primitive (pixel A) and the green dynamic pressure propagation primitive (pixel B) curves demonstrate high-frequency pressure fluctuation characteristics. Both curves exhibit periodic fluctuations with a wide range of pressure values, fluctuating between approximately 80 and 120 kPa, indicating that the dynamic pressure propagation primitive captures the rapid changes caused by the compressive deformation of fine lines during pressure propagation. A certain phase difference exists between the peaks and troughs of the pixel A and pixel B curves on the time axis, reflecting the spatial-temporal differences in the pressure response of the dynamic pressure propagation primitive at different pixel locations.

[0029] The red time-domain envelope curve corresponds to the envelope of the dynamic pressure propagation primitive curve of pixel A. It exhibits a periodic positive half-wave pattern, with a maximum value of approximately 20 kPa and a minimum value close to 5 kPa. This clearly shows that the time-domain envelope extracts the amplitude envelope of the high-frequency fluctuations of the dynamic pressure propagation primitive, capturing the amplitude profile of the pressure fluctuations. The period of the time-domain envelope matches the fluctuation period of the dynamic pressure propagation primitive, verifying the technical solution in Example 1 of constructing an analytic signal and calculating the modulus to obtain the time-domain envelope.

[0030] Example 2: In the specific implementation process, the method for extracting the distribution skeleton of coarse foot texture based on the static foot texture primitive is as follows.

[0031] The grayscale gradient of the static foot texture primitive is calculated. The Sobel operator is used to calculate the gradient magnitude along the horizontal direction and the gradient magnitude along the vertical direction for each pixel in the static foot texture primitive. The Sobel horizontal convolution kernel is a 3-pixel × 3-pixel matrix, with the first row [-1, 0, 1], the second row [-2, 0, 2], and the third row [-1, 0, 1]. The Sobel vertical convolution kernel is also a 3-pixel × 3-pixel matrix, with the first row [-1, -2, -1], the second row [0, 0, 0], and the third row [1, 2, 1]. The Sobel horizontal convolution kernel is convolved with the static foot texture primitive to obtain the horizontal gradient component at each pixel location. The Sobel vertical convolution kernel is also convolved with the static foot texture primitive to obtain the vertical gradient component at each pixel location. At each pixel location, the gradient magnitude is calculated using the horizontal and vertical gradient components. The gradient magnitude is the square root of the sum of the squares of the horizontal and vertical gradient components. The gradient direction angle is calculated using the horizontal and vertical gradient components. The gradient direction angle is the arctangent of the ratio of the vertical to the horizontal gradient components. The gradient magnitudes at all pixel locations form a gradient magnitude map, and the gradient direction angles at all pixel locations form a gradient direction map.

[0032] The local maxima in the gradient magnitude map are traced along the main direction of the gradient direction map. The main direction of the gradient direction map is the quantized direction of the gradient direction angle at each pixel location. The gradient direction angle ranges from 0 degrees to 180 degrees, and is uniformly quantized into four direction intervals: direction interval one corresponds to 0 degrees to 22.5 degrees and 157.5 degrees to 180 degrees, direction interval two corresponds to 22.5 degrees to 67.5 degrees, direction interval three corresponds to 67.5 degrees to 112.5 degrees, and direction interval four corresponds to 112.5 degrees to 157.5 degrees. The main direction of direction interval one is horizontal, the main direction of direction interval two is a 45-degree diagonal direction, the main direction of direction interval three is vertical, and the main direction of direction interval four is a 135-degree diagonal direction. For any pixel position in the gradient direction map, the main tracking direction is determined based on the directional interval into which the gradient direction angle falls. In the gradient magnitude map, one adjacent pixel position is selected along both the positive and negative directions of the main tracking direction. The gradient magnitude of the center pixel position is compared with the gradient magnitudes of the two adjacent pixel positions along the main tracking direction. If the gradient magnitude of the center pixel position is greater than the gradient magnitudes of the two adjacent pixel positions along the main tracking direction, the center pixel position is marked as a local maximum. If the gradient magnitude of the center pixel position is not greater than the gradient magnitude of any adjacent pixel position along the main tracking direction, the center pixel position is discarded.

[0033] Connecting adjacent local maxima forms a ridge line. Starting from any unprocessed local maximum, search for other local maxima within the 8-connected neighborhood of the starting pixel. The 8-connected neighborhood is the eight pixels remaining after excluding the center pixel within a 3-pixel × 3-pixel window centered on the starting pixel. If a local maximum exists within the 8-connected neighborhood, connect the starting pixel to the local maximum with the largest gradient magnitude within the 8-connected neighborhood, and use the local maximum with the largest gradient magnitude as the new starting point. Repeat the aforementioned search and connection process until there are no unconnected local maxima within the 8-connected neighborhood of the current starting point, forming a continuous ridge line segment. Replace the unprocessed local maximum as the new starting point, repeat the search and connection process, and after traversing all local maxima, a set of ridge lines composed of multiple ridge line segments is obtained.

[0034] The texture ridges are then refined. For each texture ridge segment in the set, mathematical morphology hit-or-miss transformation is used for iterative refinement. In each iteration, all pixels on the current texture ridge segment are traversed, and an 8-neighborhood pixel template is created for each pixel, centered on that pixel. The 8-neighborhood pixel template contains nine pixel positions (3 pixels × 3 pixels). The center pixel position is numbered as follows. The eight neighboring pixels, starting from the pixel directly above the center pixel and arranged clockwise, are numbered sequentially as follows: , , , , , , , Define a deletion marker condition that simultaneously satisfies the following four sub-conditions: The first sub-condition is... The pixel value is 255, meaning the center pixel is a foreground pixel; the second sub-condition is... arrive The number of pixels with a value of 255 is greater than or equal to 2 and less than or equal to 6, meaning the center pixel is neither an endpoint nor an interior point. This condition is set based on the distribution characteristics of the eight neighboring pixels. When the number of foreground pixels is 1, the center pixel is considered an endpoint; when the number of foreground pixels is 7, the center pixel is considered an interior point. Endpoints are preserved to prevent ridge line contraction, while interior points are preserved to maintain ridge line connectivity. The third sub-condition is... , , , , , , , In a clockwise sequence, the number of times a pixel transitions from background to foreground is 1, thus preserving the connectivity of its eight neighbors; the fourth sub-condition is... Multiply Multiply The product of is equal to 0, and Multiply Multiply The product of is equal to 0, or Multiply Multiply The product of is equal to 0 and Multiply Multiply The product of and is equal to 0. Pixels that meet the deletion marking condition and whose center pixel value is 255 are marked as deletion pixels. In one iteration, the pixel values ​​of all pixels that meet the deletion marking condition are updated, and the pixel values ​​of pixels marked as deletion pixels are updated from 255 to 0. The above marking and updating process is executed iteratively until no pixels are marked as deletion pixels in two consecutive iterations. The iteration ends, and the refined single-pixel width foot texture distribution skeleton is obtained.

[0035] The method for extracting the texture offset trajectory of the foot sole fine texture during the compression deformation process based on the dynamic pressure propagation primitive is as follows.

[0036] The edge points of the fine texture of the sole are identified from the dynamic pressure propagation primitive. A Canny edge detection operator is applied to each frame of the dynamic pressure propagation primitive to extract the edge positions on both sides of the fine texture. The processing flow of the Canny edge detection operator is as follows: the current frame image of the dynamic pressure propagation primitive is smoothed using a Gaussian filter with a standard deviation parameter of 0.8 pixels; the horizontal and vertical gradient components of each pixel in the smoothed image are calculated, where the horizontal gradient component is obtained by convolving the smoothed image with the aforementioned Sobel horizontal convolution kernel, and the vertical gradient component is obtained by convolving the smoothed image with the aforementioned Sobel vertical convolution kernel; the gradient magnitude and gradient direction angle are calculated at each pixel position based on the horizontal and vertical gradient components; non-maximum suppression is applied to the gradient magnitude, and the pixel position is determined based on the gradient direction angle at each pixel position. In the principal direction, compare the gradient magnitude of the pixel position with its two adjacent pixels in the gradient magnitude map, retaining the maximum value and setting the gradient magnitude of non-maximum values ​​to zero. Set the high threshold parameter to twice the average of all non-zero gradient magnitudes in the gradient magnitude map, and the low threshold parameter to 0.5 times the high threshold parameter. Mark pixels with gradient magnitudes greater than the high threshold parameter as strong edge pixels, and mark pixels with gradient magnitudes greater than the low threshold parameter and less than the high threshold parameter as weak edge pixels. Traverse all strong and weak edge pixels, retaining weak edge pixels connected to strong edge pixels within their 8-connected neighborhoods as edge pixels, and discarding the rest, to obtain the edge binary image of the current frame. Retain the pixel coordinates of all pixels with a value of 255 in the edge binary image, and use the retained pixel coordinates as the texture edge points of the foot texture in the current frame.

[0037] Record the pixel coordinate change sequence of each texture edge point in consecutive frames of the original pressure image sequence. Assign a time index to each frame of the original pressure image sequence, incrementing frame by frame from zero with a step size of 1. For any texture edge point identified in the first frame, use its pixel coordinates in the first frame as the initial coordinates. In the edge binary image of the second frame, search for the nearest texture edge point of the same category within a circular search area centered on the initial coordinates and with a radius of 5 pixels. "Same category" means that the texture edge point and the texture edge point corresponding to the initial coordinates belong to the same side edge of the same fine texture. Same-side edge points are identified by comparing the difference in gradient direction angles; gradient direction angle differences within 15 degrees are considered same-side edge points. If a same-side edge point exists within the search area, record the coordinates of the nearest same-side edge point within the search area as the coordinates of the texture edge point in the second frame. If no same-side edge point exists within the search area, use the coordinates of the first two frames to linearly extrapolate the predicted coordinates of the texture edge point in the second frame as replacement coordinates. The aforementioned search and recording operations are performed frame by frame to obtain the sequence of pixel coordinate changes of texture edge points in consecutive frames.

[0038] Trajectory fitting is performed on the pixel coordinate change sequence. For any texture edge point's pixel coordinate change sequence, its horizontal coordinate components are arranged by time index to form a horizontal coordinate sequence, and its vertical coordinate components are arranged by time index to form a vertical coordinate sequence. A cubic polynomial is then used to perform least-squares fitting on both the horizontal and vertical coordinate sequences. The cubic polynomial model is as follows:

[0039] in, This represents the time index, and its value is the frame number of the original pressure image sequence; Indicates time index The fitted value of the horizontal coordinate at that location; The zeroth-order polynomial coefficients of the horizontal axis represent the intercept of the fitted curve. The coefficients of the first-order polynomial on the horizontal axis represent the linear rate of change of the fitted curve at the initial moment. The coefficients of the second-order polynomial on the horizontal axis represent the acceleration term contribution of the fitted curve. The coefficients of the third-order polynomial for the horizontal coordinates represent the contribution of the accelerometer term to the fitted curve. The vertical coordinate sequence is fitted using a cubic polynomial of the same form, and the parameters of the fitted curve are the polynomial coefficients corresponding to the vertical coordinates. The polynomial coefficients that minimize the sum of squared residuals between the fitted curve and the actual pixel coordinate sequence are solved using the least squares method. , , , The obtained polynomial coefficients are substituted into the cubic polynomial model to calculate the fitted horizontal and vertical coordinates at each time index. These coordinates are then arranged according to the time index to form the fitted texture edge point trajectory. The aforementioned trajectory fitting operation is performed sequentially on all texture edge points. The fitted texture edge point trajectories constitute a displacement vector field. The coordinate difference between adjacent frames for each texture edge point in the displacement vector field constitutes a displacement vector. All displacement vectors constitute the texture offset trajectory.

[0040] In some implementations, the high threshold parameter of the Canny edge detection operator is set between 1.5 and 2.5 times the mean of all non-zero gradient magnitudes in the gradient magnitude map after non-maximum suppression, while the low threshold parameter is set between 0.4 and 0.6 times the high threshold parameter, thus balancing the recall and precision of the foot crease edges. The radius parameter of the circular search region is set between 3 and 7 pixels, covering the maximum inter-frame displacement caused by compressive deformation of the newborn's foot creases at the sampling frame rate.

[0041] Optionally, the fitting of the horizontal and vertical coordinate sequences can also employ quadratic or quartic polynomials. A quadratic polynomial is used when the inter-frame displacement change is approximately linear; the fitting order is increased when the inter-frame displacement change exhibits more complex nonlinearity. The fitting order is determined based on residual analysis. A cubic polynomial model can simultaneously capture both uniform and accelerated deformation characteristics, making it suitable for non-uniform pressure conditions on the soft tissues of a newborn's foot.

[0042] Example 3: In the specific implementation process, please refer to Figure 2 The curvature change gradient of each region of the sole is determined based on the three-dimensional point cloud data of the sole, and the method of using the curvature change gradient to correct the texture offset trajectory and generate the dynamic deformation corrected sole texture topology is as follows.

[0043] In specific implementation, the three-dimensional point cloud data of the foot is divided into the heel region, arch region, and forefoot region according to the anatomical structure of the foot. The division is based on the spatial position of the point cloud and the curvature characteristics of the foot. A bounding box is extracted from the three-dimensional point cloud data, and the bounding box is divided into three segments along the length of the foot, with the length direction being the principal axis of the foot from heel to toes. Within each segment, the local surface normal vector of each point in the point cloud is calculated, and the maximum curvature value in the principal curvature is obtained. The heel region corresponds to the set of points in the point cloud whose principal curvature value is lower than the first curvature threshold and whose spatial position is in the posterior third of the bounding box. The surface curvature of the point set in the heel region is gentle and located at the posterior end of the foot. The forefoot region corresponds to the set of points in the point cloud whose principal curvature value is higher than the second curvature threshold and whose spatial position is in the anterior third of the bounding box. The point set in the forefoot region includes the metatarsal head prominence. The arch region corresponds to the set of points in the point cloud whose spatial position is in the middle third of the bounding box, and the continuous change in the principal curvature value of the arch region exhibits the concave surface characteristics of the arch. The first curvature threshold was set at 1.5 times the average maximum curvature value of the heel region, and the second curvature threshold was set at 0.8 times the average maximum curvature value of the forefoot region. These thresholds were determined by statistical analysis of the curvature of a large number of newborns' plantar point clouds to stably separate the three functional areas of the plantar anatomical structure.

[0044] Surface fitting was performed on the point cloud data within each region, using a quadratic polynomial surface model for least-squares fitting. In the quadratic polynomial surface model, the coordinates... From coordinates and coordinates The quadratic polynomial expression, where , , Let be the spatial coordinates of the point cloud in the reference coordinate system. For the point set within each region, construct an overdetermined system of linear equations with polynomial coefficients. Solve for the polynomial coefficient vector using the least squares method to minimize the sum of squared Euclidean distances between the fitted surface and all points within the region. Solve for the local surface equations of each region. The local surface equation corresponding to the heel region is denoted as the first local surface equation, the local surface equation corresponding to the arch region as the second local surface equation, and the local surface equation corresponding to the forefoot region as the third local surface equation.

[0045] Calculate the Gaussian curvature and mean curvature of each local surface equation at the corresponding region boundary. For each local surface equation, a series of boundary sampling points are obtained by sampling at equal arc length intervals along the region boundary. At each boundary sampling point, the Gaussian curvature and mean curvature are calculated using the first and second fundamental form coefficients in differential geometry, based on the first and second partial derivatives of the local surface equation. The Gaussian curvature is the ratio of the determinant of the second fundamental form coefficient matrix to the determinant of the first fundamental form coefficient matrix, and the mean curvature is a combination of the first and second fundamental form coefficients. After performing the above calculations, each sampling point on the heel region boundary yields a set of heel Gaussian curvature values ​​and heel mean curvature values; each sampling point on the arch region boundary yields a set of arch Gaussian curvature values ​​and arch mean curvature values; and each sampling point on the forefoot region boundary yields a set of forefoot Gaussian curvature values ​​and forefoot mean curvature values.

[0046] The Gaussian curvature and the average curvature are differentially calculated along the region boundary to determine the curvature change between adjacent boundary points, thus obtaining the curvature change gradient. Taking the average curvature change gradient as an example, the boundary sampling points are arranged into an ordered sequence along the region boundary in an increasing arc length direction. Let the nth point in the sequence be... Each sampling point is recorded as , The value ranges from 1 to integers, This represents the total number of boundary sampling points. The average curvature change gradient is calculated using the central difference scheme:

[0047] in, Indicates sampling point The magnitude of the average curvature change gradient vector at a given location along the boundary direction; Indicates sampling point The average curvature value at that point, For the direction of advance along the boundary and Adjacent sampling points; Indicates sampling point The average curvature value at that point, For the retreat direction along the boundary and Adjacent sampling points; This represents the arc length step size parameter between adjacent sampling points. It is set as the statistical mean of the Euclidean distances between all adjacent pairs of sampling points on the boundary, and the calculation process ensures that... It is a constant value, and it will be used in this implementation. The value is set to 0.2 mm, determined based on the actual sampling density of the 3D point cloud data of the foot, ensuring that at least 50 sampling points are obtained on the boundary of each anatomical region to fully characterize the curvature variation trend. For the sequence endpoints, unidirectional differencing is used instead of central differencing. The Gaussian curvature is differentially calculated along the boundary direction using the same steps to obtain the Gaussian curvature variation gradient. The average curvature variation gradient and the Gaussian curvature variation gradient calculated at each boundary sampling point together constitute the curvature variation gradient, which covers all boundary positions of the heel region, arch region, and forefoot region.

[0048] In specific implementation, the curvature change gradient is used to correct the texture offset trajectory. The texture offset trajectory is projected onto the three-dimensional spatial coordinates corresponding to the three-dimensional point cloud data of the foot. The projection method is to use the pre-calibrated camera internal parameter matrix and the registration relationship between the pressure image and the three-dimensional point cloud to convert the pixel coordinates of each texture edge point in the texture offset trajectory into coordinates in three-dimensional space. For each texture edge point, the point cloud point with the closest distance to its projection ray is searched in the three-dimensional point cloud data of the foot, and the three-dimensional coordinates of the searched point cloud point are assigned to the texture edge point to form a three-dimensional offset trajectory. The three-dimensional offset trajectory is composed of a series of three-dimensional coordinate nodes connected in time frame order, and each node represents the spatial position of the fine texture edge point on the foot surface at the corresponding frame time.

[0049] The strain tensor field of each region of the sole during compressive deformation is determined based on the curvature change gradient. Using a pre-established curvature-strain mapping relationship, the average curvature change gradient and Gaussian curvature change gradient of each boundary sampling point in the curvature change gradient are used as inputs. The mapping relationship table is consulted to obtain the membrane strain component and bending strain component at that boundary sampling point, and then a second-order strain tensor is synthesized. The curvature-strain mapping relationship table was obtained through a combination of in vitro indentation experiments and finite element simulations using neonatal foot soft tissue as the object. The mapping relationship table records the values ​​of the six independent components of the strain tensor corresponding to different curvature change modes. For all non-boundary regions in the three-dimensional point cloud data of the sole, the radial basis function interpolation method is used, with the strain tensor of all boundary sampling points as known quantities, to interpolate the points inside each region of the sole, generating a strain tensor field covering the entire sole. The strain tensor field gradually changes with spatial position in the heel region, arch region, and forefoot region.

[0050] The strain tensor field is used to perform inverse deformation compensation on the three-dimensional offset trajectory. For each node on the three-dimensional offset trajectory, the strain tensor provided by the strain tensor field at the node's spatial coordinates is read, the inverse matrix of the strain tensor is calculated, and the node's three-dimensional coordinate vector is multiplied by the inverse matrix to obtain the corrected node three-dimensional coordinates. The corrected node three-dimensional coordinates represent the spatial position of the foot's fine texture when no compressive deformation occurs. All nodes on the three-dimensional offset trajectory are processed sequentially to obtain the corrected texture spatial coordinate sequence.

[0051] The corrected spatial coordinate sequence of the plantar ridges is topologically connected according to the connectivity of the plantar ridges to generate the plantar ridge topology. The topological connection is based on the spatial proximity criterion and the orientation consistency criterion. The spatial proximity criterion stipulates that two nodes are considered potentially connected when the Euclidean distance between them is less than a first connection threshold, set at 0.4 mm. This threshold is based on the average distribution density of fine plantar ridges in newborns and corresponds to one-third of the shortest distance between adjacent fine ridges. The orientation consistency criterion stipulates that two nodes are considered directionally compatible when the angle between the direction vectors of the respective ridge segments is less than a second connection threshold, set at 15 degrees. Two nodes that simultaneously satisfy both the spatial proximity and orientation consistency criteria are grouped into the same connected group. Nodes within the connected group are connected using a minimum Euclidean distance spanning tree method to form ridge topology edges. After traversing all nodes in the corrected spatial coordinate sequence and completing the connected group partitioning and connection, a graph structure consisting of nodes and edges is obtained; this graph structure is the plantar ridge topology. The plantar texture topology retains the overall spatial distribution and branch connection relationship of the fine texture after pressure deformation correction, directly expressing the inherent plantar texture pattern of newborns.

[0052] It is understood that in the calculation of the curvature change gradient, the number of sampling points at the region boundary... Arc length step size The value can be adaptively adjusted based on the point density of the actual three-dimensional point cloud data of the foot sole; when the point cloud density increases, the value decreases. To obtain higher resolution curvature variation information, the increase is achieved when the point cloud density decreases. To avoid undersampling. The first and second connection thresholds can also be adjusted according to the statistical characteristics of the plantar morphology of different newborns. For newborns with particularly dense plantar morphology, the first connection threshold is reduced to 0.3 mm and the second connection threshold is correspondingly narrowed.

[0053] Example 4: In the specific implementation process, the foot pattern distribution skeleton and the foot pattern topology are hierarchically matched to obtain the newborn's foot pattern identification code as follows.

[0054] First-level feature encoding is performed on the coarse texture distribution skeleton of the foot. The coordinates of the main bifurcation points and main endpoints of the coarse textures are extracted. The main bifurcation point is the position of the three-way branch on the ridge line, and the main endpoint is the position of the ridge line's end point. After refinement, the coarse texture distribution skeleton of the foot is a set of connected curves with a single pixel width. For each foreground pixel on a single-pixel-width connected curve, the number of foreground pixels in its eight-neighborhood is counted. Foreground pixels with 1 foreground pixel in their eight-neighborhood are marked as main endpoints. Foreground pixels with 3 foreground pixels in their eight-neighborhood are marked as main bifurcation points. Foreground pixels with 4 or more foreground pixels in their eight-neighborhood are also marked as main bifurcation points. However, for the case where there are 4 foreground pixels in their eight-neighborhood, if the distribution pattern of the four foreground pixels is a diagonal distribution with two opposite sides, they are marked as intersection points rather than main bifurcation points. Intersection points are considered as the overlap of two independent main bifurcation points, and during unwrapping, the intersection points are split into two independent main bifurcation points. The splitting method involves breaking the connecting ridges at the intersection points, creating three branches at each intersection. All foreground pixels are traversed to obtain all main bifurcation points and main endpoints on the foot's coarse texture distribution skeleton. Using the last endpoint in the heel region of the foot's coarse texture distribution skeleton as a reference point, the polar coordinate angle and Euclidean distance of each main bifurcation point and main endpoint relative to the reference point are calculated. All main bifurcation points and main endpoints are sorted in ascending order of polar coordinate angle and ascending order of Euclidean distance. The pixel x and y coordinates of the sorted main bifurcation points and main endpoints are extracted. All coordinate values ​​are sequentially placed into a vector, with the first half containing the x and y coordinates of all main bifurcation points and the second half containing the x and y coordinates of all main endpoints, generating a first-level feature vector.

[0055] The plantar texture topology is encoded using a two-level feature encoding method to extract the relative spacing and relative orientation angle of the corrected fine lines. The plantar texture topology consists of nodes and topological edges, with each topological edge corresponding to a corrected fine line segment. For each topological edge, the spatial three-dimensional coordinates of all nodes on that edge are extracted, and the first principal direction vector of the spatial distribution of these nodes is calculated using principal component analysis, serving as the fine line orientation vector for that topological edge. All topological edges are traversed, and a fine line orientation vector is obtained for each edge. The relative spacing and relative orientation angle between any two topological edges are calculated. The relative spacing is defined as the shortest Euclidean distance between the node sets of two topological edges; if two topological edges share the same node, the shortest Euclidean distance is zero. The Euclidean distance between any node in the node set of one topological edge and any node in the node set of the other topological edge is calculated, and the minimum value among all distances is taken as the relative spacing between the two topological edges. The relative orientation angle is defined as the angle between the fine line orientation vectors of two topological edges, obtained using inverse cosine. For all unique topological edge pairs, calculate the relative spacing and relative orientation angle to obtain a set of spacing values ​​and a set of angle values. Arrange the spacing and angle values ​​according to spatial proximity, that is, take each topological edge as a reference and record the relative spacing and relative orientation angle between the reference topological edge and the other topological edge with the closest spatial distance. If there are ties for the closest, sort them in ascending order of angle value. Finally, combine the spacing and angle values ​​corresponding to all reference topological edges in sequence to generate a secondary feature vector.

[0056] Align the primary and secondary feature vectors according to their spatial positions on the sole of the foot. Use the coordinates of the main bifurcation point in the primary feature vector as the anchor point. Select the main bifurcation point with the smallest polar coordinate angle and the smallest Euclidean distance from the primary feature vector as the anchor point. This anchor point corresponds to the feature point closest to the heel bifurcation position in the sole crease. Translate the three-dimensional spatial coordinates of all nodes in the sole crease topology. The translation vector is the three-dimensional spatial coordinate vector of the anchor point. After translation, the anchor point becomes the new origin. Calculate the principal axis direction of the sole coarse crease distribution skeleton. The principal axis direction is obtained by calculating the eigenvector corresponding to the largest eigenvalue of the distribution covariance matrix of all foreground pixels on the sole coarse crease distribution skeleton. The principal axis direction points from the heel to the forefoot. Rotate the node coordinates of the translated sole crease topology. The rotation axis is vertical, and the rotation angle is the angle between the projection of the principal axis direction onto the horizontal plane and a preset standard direction. The preset standard direction is taken as the direction from left to right in the horizontal plane. The rotation transformation unifies the node coordinates of the foot ridge topology into a local coordinate system with the anchor point as the origin and the principal axis as the positive horizontal direction. In this local coordinate system, the fine ridge direction vector of each topological edge is recalculated, and the relative spacing and relative direction angle between all topological edge pairs are also recalculated to generate the registered secondary feature vector. Simultaneously, the coordinates of the main bifurcation point and the main endpoint in the primary feature vector are also transformed to the local coordinate system through the same translation and rotation transformations to generate the registered primary feature vector.

[0057] The registered primary and secondary feature vectors are concatenated into a joint feature sequence. The coordinates of all major bifurcation points and major endpoints in the registered primary feature vector are extracted element-by-element. Then, the relative spacing and relative orientation angle values ​​in the registered secondary feature vector are extracted element-by-element. These are concatenated together in the order of primary feature vectors first, then secondary feature vectors, resulting in a one-dimensional array, i.e., the joint feature sequence. The length of the joint feature sequence is fixed. When the number of major bifurcation points, major endpoints, and topological edges on the sole of a newborn's foot is less than a preset upper limit, the insufficient portion is padded with zeros. The preset upper limit is set as a conservative maximum value after statistically analyzing the number of sole crease features of a large number of newborns.

[0058] The joint feature sequence is quantized and encoded to generate a fixed-length binary sequence as the plantar morphology identification code. For each feature component in the joint feature sequence, a lower and upper bound for quantization are determined. These lower and upper bounds are set after statistical analysis of the joint feature sequences from multiple pre-collected neonatal plantar morphology samples. The lower bound for each feature component is the minimum value of that component in the sample set minus a protective offset, and the upper bound is the maximum value in the sample set plus a protective offset, where the protective offset is 0.2 times the standard deviation of the feature component. The value range of each feature component is evenly divided into... One quantization interval, The value is 16. The rationale for setting it to 16 is that using 4 bits to represent a single feature component provides sufficient codeword space to accommodate inter-individual differences in newborn plantar crease patterns, while limiting the impact of quantization errors on matching performance. In the ratio analysis of the standard deviation of intra-class fluctuation to the standard deviation of inter-class difference in plantar crease features, the signal-to-noise ratio margin corresponding to 4-bit quantization meets the requirements for identity differentiation. Each feature component value is mapped to a quantization level number using the following quantization formula:

[0059] in, Indicates the first element in the joint feature sequence The quantization level number corresponding to the dimensional feature component. For feature component index, Indicates the first The original values ​​of the dimensional feature components. Indicates the first The lower bound of quantization for dimensional feature components. Indicates the first The upper bound of quantization of dimensional feature components. To quantize the total number of intervals. The calculated... Convert to a four-bit binary number. The bit width of a four-bit binary number is determined by... The quantization level determines the number of quantization levels, with 16 quantization levels corresponding to binary numbers from 0000 to 1111. After performing the aforementioned quantization encoding operation on all feature components, the four-bit binary numbers of each dimension are concatenated end-to-end according to the feature component index order to generate a fixed-length binary sequence, which serves as the identification code for the sole pattern.

[0060] In some implementations, when identifying the main bifurcation point, if the number of foreground pixels within the eight-neighborhood is four, and the foreground pixels are distributed in four consecutive adjacent positions, they are determined to be two bifurcation points close together. These are then separated into two independent main bifurcation points using a local ridge decoupling algorithm. The protection offset coefficient of 0.2 can be adjusted according to the distribution trail of the foot tread features; for dispersed feature components, the coefficient can be appropriately increased.

[0061] It is understood that the selection of anchor points is not limited to the main bifurcation point with the smallest polar coordinate angle, but can also be based on prior knowledge of the stability of the plantar texture distribution to select the bifurcation point with the highest probability of occurrence. After the anchor point is changed, the origin of the local coordinate system is updated accordingly, and the encoding rule of the plantar texture identification code remains consistent with the anchor point selection method. The calculation of the main axis direction can also be replaced by the least second moment axis direction of the plantar coarse texture distribution skeleton, and the two have equivalent effects on the calibration results.

[0062] Example 5: In the specific implementation process, please refer to Figure 3 The method for obtaining the original pressure image sequence and three-dimensional point cloud data of the newborn's foot is as follows.

[0063] The pressure acquisition device applies increasing pressure to the soles of the newborn's feet at a preset sampling frequency and simultaneously acquires multiple frames of sole pressure distribution images. The pressure acquisition device includes a pressure platform, a pressure sensor array, and an image acquisition unit. The pressure platform is pneumatically driven; by adjusting the air pressure inside the airbag, the foot-holding bracket placed above the platform slowly rises vertically, pressing the newborn's soles against the surface of the pressure sensor array. The pressure sensor array consists of high-resolution pressure-sensitive thin-film sensors with a sensing unit density of at least four sensing points per square centimeter, capable of sensing and outputting the pressure amplitude at various locations in space. The pressure sensor array outputs the pressure value measured at each sensing point as a two-dimensional grayscale image, with the grayscale value proportional to the pressure amplitude, forming a single-frame sole pressure distribution image. The air pressure regulating valve of the pressure application platform starts from an initial zero-pressure state and increases the air pressure linearly at a rate of 0.5 kPa per second. This rate is selected based on the elastic response characteristics of the soft tissue of a newborn's foot under pressure, ensuring a quasi-static loading state during the pressure application process. This allows for continuous and trackable foot deformation between adjacent acquisition frames. During the pressure increase, the image acquisition unit continuously and synchronously acquires foot pressure distribution images at a preset sampling frequency of 60 Hz. This value ensures that the pixel displacement corresponding to the foot deformation under pressure within the time interval between two adjacent frames does not exceed 5 pixels, maintaining the stability of inter-frame tracking of fine texture feature points. The acquired multiple frames of foot pressure distribution images are arranged in chronological order of acquisition to form the original pressure image sequence.

[0064] The acquisition time interval between two adjacent frames in the original pressure image sequence is 0.3 to 0.5 times the time required for the newborn's plantar skin to complete one elastic recoil. The time required for the newborn's plantar skin to complete one elastic recoil was determined through a preliminary experiment. The method involved applying instantaneous local pressure to the plantar skin under non-pressure conditions and then releasing it. A high-speed camera recorded the time it took for the indentation to disappear from the skin surface. The time required for the indentation to decay from its maximum depth to within 5% of its initial depth was defined as the elastic recoil completion time. After completing the preliminary experiment on multiple newborn samples, the sample mean of the elastic recoil completion time was taken as the standard elastic recoil time. The sample mean was set to an integer value in milliseconds as the basis for setting. Here, the acquisition time interval was set to 0.4 times the elastic recoil completion time. Specifically, the value was the elastic recoil completion time multiplied by 0.4 and rounded to an approximate integer millisecond. One value for the elastic recoil completion time was 80 milliseconds, corresponding to an acquisition time interval of 32 milliseconds. The reason for setting the acquisition time interval to 0.4 times the elastic recoil completion time is that the 0.4 times interval can ensure that at least two frames of pressure distribution images are acquired in each skin elastic recoil cycle, which meets the sampling density requirements of temporal envelope extraction, and avoids excessive deformation amplitude between adjacent frames, which would cause the inter-frame movement of texture edge points to exceed the search radius limit.

[0065] While the pressure acquisition device applies the increasing pressure, the structured light scanning device projects a structured light pattern onto the sole of the newborn's foot. The structured light scanning device includes a structured light projection unit and a structured light imaging unit. The structured light projection unit uses a digital light processing projector to output a blue light stripe pattern with a wavelength of 450 nanometers. The structured light pattern uses a phase-shifting fringe projection mode to generate a set of structured light fringe images with a sinusoidal intensity distribution. The fringe spatial frequency is set to 0.5 line pairs per millimeter, corresponding to a fringe period of 2 millimeters, which is sufficient to resolve the surface height variation details required for the three-dimensional point cloud data of the sole. Within one structured light scanning cycle, the structured light projection unit sequentially projects four sinusoidal fringe patterns, each with a phase difference of π / 2, onto the sole surface. The phase offsets of the four sinusoidal fringe patterns are 0, π / 2, π, and 3π / 2, respectively. The structured light imaging unit consists of an industrial camera mounted at a fixed baseline angle to the structured light projection unit. The camera's acquisition rate is synchronized with the image acquisition unit of the pressure acquisition device, both operating at 60 Hz. The structured light imaging unit is simultaneously exposed during the projection of each set of sinusoidal fringe patterns, acquiring four structured light deformable images of the newborn's sole corresponding to the four sinusoidal fringe patterns. In the structured light deformable images, the sinusoidal fringes are modulated by the three-dimensional morphology of the sole surface, and the fringe phase distribution carries the height information of the sole surface.

[0066] Phase calculation is performed on the structured light deformation images to generate the three-dimensional point cloud data of the foot. The phase calculation employs a four-step phase-shifting method. Let the four acquired structured light deformation images be the first deformation images. Second deformed image Third deformed image and the fourth deformed image ,in Represents the pixel coordinates of the image sensor plane of the structured light imaging unit. Wrap phase. The intensity values ​​of the four structured light deformation images are calculated using the following formula:

[0067] in, Indicates pixel coordinates The package phase value at the location has a range of values. This can be expanded by determining the positive and negative combinations of the numerator and denominator. interval; This represents the first deformed image with a phase offset of 0 at pixel coordinates. Pixel intensity value at; This represents the second deformed image with a phase offset of π / 2 at pixel coordinates. Pixel intensity value at; This represents the third deformed image with a phase offset of π at pixel coordinates. Pixel intensity value at; This represents the fourth deformed image with a phase offset of 3π / 2 in pixel coordinates. The pixel intensity value at that location. The calculated wrap-around phase. Spatial phase unwrapping is performed using a reliability-based quality-guided path unwrapping algorithm. Starting from the pixel with the highest quality value in the central region of the structured light deformation image, the unwrapping proceeds outwards along the direction of minimum phase gradient, eliminating phase jumps and obtaining a continuously distributed absolute phase map. This absolute phase map is then combined with the system calibration parameters of the structured light scanning device, and the three-dimensional spatial coordinates of the foot surface points are calculated pixel-by-pixel based on the principle of optical triangulation. The system calibration parameters include the intrinsic parameter matrices of the structured light projection unit and the structured light imaging unit, the relative rotation and translation matrices, and the phase-to-height mapping coefficients. These parameters are obtained in advance through an offline calibration process using a planar calibration plate. The three-dimensional spatial coordinates of the foot surface at all pixel positions constitute the three-dimensional point cloud data of the foot.

[0068] In some embodiments, the pressure increment rate of the pressure acquisition device is adjustable between 0.3 kPa / s and 0.8 kPa / s. The adjustment principle is that the original pressure image sequence contains no less than 40 frames of pressure distribution images within the time period of applying the maximum value, to ensure sufficient frame count support for the temporal analysis of the subsequent primitive decomposition module. The fringe spatial frequency used by the structured light scanning device can be selected in the range of 0.3 line pairs per millimeter to 0.8 line pairs per millimeter. A lower fringe spatial frequency is suitable for the plantar morphology of the foot arch region with deeper depressions, reducing phase loss in the shadow area.

[0069] Optionally, the ratio of the acquisition time interval to the elastic recoil completion time of 0.4 can be fine-tuned based on the actual measured differences in elastic recoil time among different newborns. When the elastic recoil time distribution measured in the preliminary experiment is wide, 0.3 times the elastic recoil completion time is taken as the acquisition time interval to cover faster elastic recoil rates. When the overall elastic recoil time is relatively long, 0.5 times the elastic recoil completion time is taken to increase the amount of deformation information in adjacent frames.

[0070] It is understandable that in the four-step phase-shifting method used for phase calculation, the foot deformation under pressure has not stopped during the capture of the four structured light deformation images. Therefore, a rapid switching method of sinusoidal fringe sequence with an extremely short time is adopted. The total duration of a complete structured light scanning cycle is no more than 10 milliseconds. When the acquisition time interval between adjacent frames of the original pressure image sequence is 32 milliseconds, the structured light scanning cycle is completely embedded in the time window of each pressure image acquisition frame. The three-dimensional point cloud data corresponding to each pressure image acquisition frame is taken from the output of the most recently completed structured light scanning cycle at that frame time point.

Claims

1. A newborn identification system based on footprint recognition, characterized in that, include: The image acquisition module acquires the original pressure image sequence and three-dimensional point cloud data of the newborn's sole. The primitive decomposition module decomposes the original pressure image sequence into static plantar texture primitives and dynamic pressure propagation primitives; The texture extraction module extracts the distribution skeleton of coarse textures on the sole based on the static sole texture primitives, and extracts the texture offset trajectory of fine textures on the sole during the compression deformation process based on the dynamic pressure propagation primitives. The deformation correction module determines the curvature change gradient of each region of the sole based on the three-dimensional point cloud data of the sole, and uses the curvature change gradient to correct the texture offset trajectory, thereby generating a dynamically deformed and corrected sole texture topology. The hierarchical matching module performs hierarchical matching between the distribution skeleton of the coarse foot lines and the topological structure of the foot lines to obtain the foot line identification code of the newborn.

2. The newborn identification system based on footprint recognition according to claim 1, characterized in that, The original pressure image sequence is decomposed into static plantar texture primitives and dynamic pressure propagation primitives, including: The first frame pressure image is selected as the reference pressure image from the original pressure image sequence. Spatial domain filtering is performed on the reference pressure image to extract the texture components whose pressure values ​​change with spatial position but whose change with inter-frame time is less than a preset threshold as the static foot texture primitives. Subtracting the mapping components of the static foot texture primitive in each frame from the original pressure image sequence yields a high-frequency residual sequence of pressure values ​​changing with inter-frame time. Temporal envelope extraction is then performed on the high-frequency residual sequence to obtain the dynamic pressure propagation primitive.

3. The newborn identification system based on footprint recognition according to claim 2, characterized in that, Extracting the skeleton of coarse foot texture distribution based on the static foot texture primitives includes: The grayscale gradient of the static plantar texture primitives is calculated to obtain the gradient magnitude map and gradient direction map of the plantar texture. Tracing local maxima in the gradient magnitude map along the main direction of the gradient direction map, connecting adjacent local maxima to form ridge lines, and refining the ridge lines to obtain the coarse texture distribution skeleton of the sole.

4. The newborn identification system based on footprint recognition according to claim 2, characterized in that, The dynamic pressure propagation primitives are used to extract the texture displacement trajectory of the fine texture on the sole of the foot during the compression deformation process, including: Identify the edge points of the fine lines on the sole of the foot from the dynamic pressure propagation primitives, and record the sequence of pixel coordinate changes of each edge point in consecutive frames of the original pressure image sequence; The pixel coordinate change sequence is fitted with a trajectory to generate a displacement vector field of each texture edge point during the compression deformation process, and the displacement vector field is used as the texture offset trajectory.

5. The newborn identification system based on footprint recognition according to claim 1, characterized in that, Based on the aforementioned three-dimensional point cloud data of the foot, the curvature change gradient of each region of the foot is determined, including: The three-dimensional point cloud data of the foot is divided into the heel region, the arch region and the forefoot region according to the anatomical structure of the foot. The point cloud data in each region is fitted with a surface to obtain the local surface equation of each region. Calculate the Gaussian curvature and mean curvature of each of the local surface equations at the corresponding region boundary, and perform a difference operation on the Gaussian curvature and the mean curvature along the direction of the region boundary to obtain the curvature change gradient.

6. The newborn identification system based on footprint recognition according to claim 5, characterized in that, The curvature change gradient is used to correct the texture offset trajectory, generating a dynamically deformed plantar texture topology, including: The texture offset trajectory is projected onto the three-dimensional spatial coordinates corresponding to the three-dimensional point cloud data of the sole of the foot to obtain the three-dimensional offset trajectory. Based on the curvature change gradient, the strain tensor field of each region of the sole during the compression deformation process is determined, and the strain tensor field is used to perform reverse deformation compensation on the three-dimensional offset trajectory to obtain the corrected texture space coordinate sequence. The corrected spatial coordinate sequence of the texture is topologically connected according to the connectivity of the sole texture to generate the sole texture topology structure.

7. The newborn identification system based on footprint recognition according to claim 1, characterized in that, The foot pattern distribution skeleton is hierarchically matched with the foot pattern topology to obtain the newborn's foot pattern identification code, including: The foot sole coarse texture distribution skeleton is subjected to first-level feature encoding, and the coordinates of the main bifurcation point and the main endpoint of the coarse texture are extracted to generate a first-level feature vector; The topological structure of the foot crease is encoded using a two-level feature encoding method. The relative spacing and relative orientation angle of the corrected fine creases are extracted to generate a two-level feature vector. The primary feature vector and the secondary feature vector are aligned according to the spatial position of the foot and then hashed to generate the foot pattern identification code.

8. The newborn identification system based on footprint recognition according to claim 1, characterized in that, Obtain the raw pressure image sequence and 3D point cloud data of the newborn's foot, including: The pressure acquisition device applies increasing pressure to the soles of the newborn's feet at a preset sampling frequency and simultaneously acquires multiple frames of sole pressure distribution images. The multiple frames of sole pressure distribution images are arranged in chronological order to form the original pressure image sequence. The structured light scanning device projects a structured light pattern while the pressure acquisition device applies the increasing pressure, acquires a structured light deformation image of the newborn's sole, performs phase calculation on the structured light deformation image, and generates the three-dimensional point cloud data of the sole.

9. The newborn identification system based on footprint recognition according to claim 1, characterized in that, The acquisition time interval between two adjacent frames in the original pressure image sequence is 0.3 to 0.5 times the time required for the newborn's sole skin to complete one elastic recoil.

10. The newborn identification system based on footprint recognition according to claim 7, characterized in that, Aligning the primary feature vector and the secondary feature vector according to the spatial position of the foot sole, and then performing a hash mapping, the foot sole pattern identification code is generated, including: Using the coordinates of the main bifurcation point in the first-level feature vector as the anchor point, the relative spacing and relative orientation angle of the fine lines in the second-level feature vector are registered to a local coordinate system with the anchor point as the origin. The registered first-level feature vector and the second-level feature vector are concatenated into a joint feature sequence. The joint feature sequence is then quantized and encoded to generate a fixed-length binary sequence as the identification code for the sole pattern.

Images