River terrain three-dimensional point cloud dynamic reconstruction method
By using manifold triangular meshes and Laplace-Beltrami operator matrix eigenvalue decomposition, adaptive low-pass filters, and weighted Laplace shrinkage algorithms, combined with the Alpha-Shape algorithm, an accurate three-dimensional surface model of the river channel topography was generated. This solved the geometric distortion problem caused by noise removal in existing technologies and improved the accuracy of hydrodynamic analysis.
Patent Information
- Application Number
- CN202610125846.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-01-29
- Publication Date
- 2026-05-15
AI Technical Summary
Existing technologies struggle to effectively remove noise without sacrificing topographic features when processing 3D point cloud data of river channels, leading to distortion of river channel geometry and affecting the accuracy of calculating river cross-sectional area and assessing flood control capacity.
Eigenvalue decomposition was performed using manifold triangular mesh and Laplace-Beltramian operator matrix. An adaptive low-pass filter was used for noise reduction. The weighted Laplace shrinkage algorithm was combined to extract the river channel central axis skeleton line. Finally, the Alpha-Shape algorithm was used to generate a continuous three-dimensional surface model of the river channel topography.
It effectively removes noise from river point clouds, preserves the macroscopic topographic features of the river, improves the accuracy of hydrodynamic analysis and engineering calculations, and solves the technical problem that unstructured scattered point clouds are difficult to use directly for hydrological feature extraction.
Smart Images

Figure CN122049265A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of point cloud modeling technology, and in particular to a method for dynamic reconstruction of three-dimensional point clouds of river topography. Background Technology
[0002] Point cloud modeling technology involves using a large number of discrete spatial coordinate points (X, Y, Z) to digitally represent the geometric shape of a three-dimensional object or scene surface.
[0003] Existing technologies, when processing massive discrete spatial coordinate points to represent complex terrain, employ Gaussian filtering based on Euclidean neighborhood averaging or moving least squares methods for denoising. While these spatial domain smoothing operations suppress noise, they indiscriminately smooth out sharp terrain features and edges, leading to irreversible shrinkage and deformation of the object's geometric volume. In river mapping applications, this volume shrinkage effect manifests as blurred and inward-shrinking riverbank boundaries and an erroneous elevation of the riverbed bottom, causing the generated digital elevation model to deviate significantly from the actual terrain. This results in an underestimation of the river's cross-sectional area and inaccurate flood control capacity assessments. Traditional methods, when directly triangulating or reconstructing scattered point clouds, often struggle to distinguish between high-frequency noise caused by water flow fluctuations and actual terrain undulations due to a lack of prior knowledge of the object's topological framework. This leads to the generated model surface being riddled with artifacts, holes, and non-manifold structures. Therefore, improvements are needed. Summary of the Invention
[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing a dynamic reconstruction method for three-dimensional point clouds of river topography.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: a method for dynamic reconstruction of three-dimensional point clouds of river topography, comprising the following steps: Obtain the original 3D point cloud data of the target river channel, and construct a manifold triangular mesh based on the original 3D point cloud data; Calculate the Laplace-Beltramian operator matrix of the manifold triangular mesh, perform eigenvalue decomposition on the Laplace-Beltramian operator matrix to obtain the manifold harmonic basis functions, and transform the manifold triangular mesh from Euclidean space to the spectral domain based on the manifold harmonic basis functions; The high-frequency noise components in the spectral domain are truncated using the constructed adaptive low-pass filter. The processed spectral coefficients are then inversely transformed back to Euclidean space to obtain the denoised river point cloud model. The river point cloud model is iteratively shrunk using a weighted Laplace shrinkage algorithm, which collapses the three-dimensional river point cloud model into a one-dimensional set of points. The river central axis skeleton line is then extracted based on the set of points. An orthogonal plane is generated along the tangent direction of the central axis skeleton line of the river channel. The point cloud model of the river channel is projected onto the orthogonal plane, and the cross-sectional contour on each orthogonal plane is extracted using the Alpha-Shape algorithm. Using the central axis of the river channel as the scanning path, spline interpolation algorithm is used to connect the contours of each section to generate a continuous three-dimensional surface model of the river channel topography.
[0006] Preferably, the step of calculating the Laplace-Beltrami operator matrix of the manifold triangular mesh specifically includes: Traverse each vertex in the manifold triangular mesh, identify the set of neighboring vertices adjacent to the vertex, and calculate the edge vector connecting the vertex to each neighboring vertex in the set of neighboring vertices; Calculate the average curvature of each vertex, and based on the topology of the manifold triangular mesh, calculate the cotangent of the diagonal angle associated with the edge vector; A terrain curvature sensitivity factor based on the average curvature is introduced to correct the cotangent value, an edge weight coefficient is constructed, and the area of the mixed region of each vertex is calculated to construct a quality matrix. The discretized Laplace-Beltrami operator matrix is generated based on the edge weight coefficients and the mass matrix; The elements in the Laplace-Beltrami operator matrix are calculated based on the following formula for edge weight coefficients: ; Based on the edge weight coefficient The off-diagonal elements of the Laplace-Beltrami operator matrix and diagonal elements They are defined as follows: , ; in, Indicates the first The area of the mixed region of the vertices, Indicates the first The set of neighboring vertices of the given vertex. and This represents the index in the set of neighboring vertices. and Indicates connection to vertices and vertex The two opposite diagonal angles of the common side, Indicates the first The average curvature of the vertices. Indicates the first The average curvature of the neighboring vertices This represents the curvature-sensitive control parameter. This represents the absolute value of the curvature difference. Represented by natural constant An exponential function with base 0.
[0007] Preferably, the step of using the constructed adaptive low-pass filter to truncate high-frequency noise components in the spectral domain specifically includes: The Laplace-Beltrami operator matrix is subjected to eigenvalue decomposition to obtain an eigenvalue sequence sorted by size and a corresponding eigenvector sequence. The eigenvector sequence is the manifold harmonic basis function. The vertex coordinates of the manifold triangular mesh are projected onto the feature vector sequence to calculate the corresponding spectral coefficient vector; Design a cutoff frequency threshold based on energy percentage, and construct a transfer function to retain low-frequency components below the cutoff frequency threshold and suppress high-frequency components above the cutoff frequency threshold; The transfer function formula of the adaptive low-pass filter is: ; in, Indicates that for the first eigenvalues The filter gain coefficient, Represents the th in the sequence of eigenvalues 1 eigenvalue, This represents the cutoff frequency threshold. This indicates the filter's order and is used to control the filter's roll-off rate. This represents the high-frequency suppression compensation coefficient. This represents the largest eigenvalue in the eigenvalue sequence; The spectral coefficient vector is updated using the transfer function, and the updated spectral coefficient vector is linearly combined with the feature vector sequence to complete the inverse transformation from the spectral domain to the spatial domain, generating the denoised river point cloud model.
[0008] Preferably, the step of iteratively shrinking the river point cloud model based on the weighted Laplace shrinkage algorithm specifically includes: Initialize the shrinkage iteration parameters, and use the denoised river point cloud model as the initial point cloud set; In each iteration, the Laplacian weight coefficients between points in the point cloud set are updated and the local neighborhood geometric center of each point is calculated. A contraction vector pointing to the geometric center is constructed, and a volume-preserving constraint term is introduced to restrict excessive tangential movement while contracting along the normal direction, thereby preventing topological breaks in the river point cloud model during the contraction process. Perform iterative update operations until the point cloud set no longer shifts, and obtain the converged zero-volume point set; The shrinkage formula for the iterative update operation is: ; in, Indicates the first After the nth iteration The three-dimensional coordinate vector of a point Indicates the first During the nth iteration The three-dimensional coordinate vector of a point Indicates the number of iterations Decreasing contraction rate factor Indicates the first A set of neighborhood indexes for each point Indicates based on current coordinates Calculated Laplace weighting coefficients, Represents the three-dimensional coordinate vector of neighborhood points. This represents the volume retention constraint strength coefficient. Indicates the first Normal vector at each point This represents the local volume change rate function. Indicates the first The local neighborhood volume at each point, and the local volume change rate function are used to provide feedback on the current volume compression state.
[0009] Preferably, the step of extracting the river channel central axis skeleton line based on the set of points specifically includes: Connectivity analysis is performed on the converged set of zero-volume points to construct a minimum spanning tree graph structure; Identify the main branch path and spurious branch path in the minimum spanning tree graph structure, and remove the spurious branch path whose ratio is less than a preset threshold by calculating the ratio of the branch length to the main path length. The retained main branch path is fitted with a B-spline curve, and the fitted curve is resampled at equal intervals to generate uniformly distributed skeleton nodes. Calculate the tangent vector, normal vector, and binormal vector at the skeleton nodes to construct a local Fleischer frame that varies with the river channel central axis skeleton line; The smoothness of the river channel central axis skeleton line is corrected based on the local Flyner frame to ensure that the river channel central axis skeleton line conforms to the dynamic continuity characteristics of river flow.
[0010] Preferably, the step of generating an orthogonal plane along the tangent direction of the river channel's central axis specifically includes: The sampling nodes on the central axis of the river channel are traversed according to a preset step size, and the tangent vector at each sampling node is calculated using the central difference method. Based on the tangent vector, construct the plane equation passing through the sampling node, and define the plane equation as the orthogonal plane; Determine the subset of local point clouds in the denoised river point cloud model that is closest to the orthogonal plane; The local point cloud subset is projected onto the orthogonal plane along the direction of the tangent vector to obtain a two-dimensional planar point set.
[0011] Preferably, the step of extracting the cross-sectional profile on each of the orthogonal planes using the Alpha-Shape algorithm specifically includes: The two-dimensional planar point set is subjected to Delaunay triangulation to generate a two-dimensional triangulation network set; Calculate the circumcircle radius of each triangle in the two-dimensional triangulation set, and compare the circumcircle radius with the Alpha threshold parameter; If the radius of the circumcircle is smaller than the Alpha threshold parameter, the corresponding triangle is retained; otherwise, the corresponding triangle is deleted. Extract the boundary edges of all the retained triangles, and connect the boundary edges in sequence to form a closed polygon chain, which is the cross-sectional profile; The cross-sectional contour is simplified by polygon processing. While keeping the contour shape feature error less than a preset value, the number of vertices constituting the cross-sectional contour is reduced, thereby optimizing the amount of data for modeling.
[0012] Preferably, the step of using a spline interpolation algorithm to connect the profiles of each section to generate a continuous three-dimensional surface model of the river channel topography specifically includes: All extracted cross-sectional profiles are indexed and numbered according to their arrangement order along the central axis of the river channel; Identify corresponding feature points on two adjacent cross-sectional contours and establish a mapping relationship between feature point pairs; The Catmull-Rom spline function is used to generate transition surface meshes between adjacent cross-sectional profiles.
[0013] Preferably, the method further includes a step of optimizing the generated three-dimensional surface model of the river channel topography based on fluid dynamic constraints, specifically including: Calculate the gradient along the course and the rate of change of cross-sectional area of the water passage in the three-dimensional surface model of the river channel topography; Based on the principles of open channel fluid mechanics, detect whether there are abnormal abrupt changes that violate the equation of continuity of water flow; If the abnormal abrupt change region is detected, the scale parameters of the cross-sectional profile at the corresponding location are adjusted back, and the step of connecting each cross-sectional profile using the spline interpolation algorithm is re-executed until the three-dimensional surface model of the river channel topography meets the hydraulic smoothness requirements. The final three-dimensional surface model of the river channel topography is output and converted into a digital elevation model format for storage.
[0014] Preferably, the step of designing the cutoff frequency threshold based on energy percentage specifically includes: Calculate the spectral energy corresponding to each feature value in the feature value sequence, and calculate the cumulative spectral energy distribution; Set an energy retention ratio threshold, and search for critical feature value indices that satisfy the energy retention ratio threshold in the cumulative spectral energy distribution; The feature value corresponding to the critical feature value index is determined as the cutoff frequency threshold; The energy retention ratio threshold ranges from 85% to 95%.
[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the original three-dimensional point cloud of the target river channel is constructed into a manifold triangular mesh and the Laplace-Beltramian operator matrix is calculated. The manifold harmonic basis functions are obtained through eigenvalue decomposition. The geometric spatial information is transformed to the spectral domain for analysis. The constructed adaptive low-pass filter is used to accurately truncate the high-frequency components representing noise interference in the spectral domain. While preserving the macroscopic topographic features of the river channel, measurement errors and water surface disturbances are eliminated. After inverse transformation, a pure model that retains the original geometric volume and morphological features is obtained, effectively avoiding geometric distortion problems such as riverbed shallowing or bank shrinkage caused by spatial domain filtering. The weighted Laplace shrinkage algorithm is used to iteratively shrink the model, collapsing the complex three-dimensional river channel morphology into a one-dimensional set of points and extracting the central skeleton line. This can capture the topological direction and central trend of the river. Orthogonal planes are generated along the tangent direction of the central skeleton, and the cross-sectional contours are extracted using the Alpha-Shape algorithm. This achieves a parameterized dimensionality reduction expression of the river channel geometry. Using the central skeleton as the scanning path, spline interpolation algorithm is used to connect the contours of each cross-section, generating a three-dimensional surface model of the river channel topography with a continuous topological structure. This improves the accuracy and usability of the model in hydrodynamic analysis and engineering calculations, and solves the technical problem that unstructured scattered point clouds are difficult to use directly for hydrological feature extraction. Attached Figure Description
[0016] Figure 1 This is a schematic diagram of the steps of the present invention. Detailed Implementation
[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.
[0018] Please see Figure 1 This invention provides a technical solution: a method for dynamic reconstruction of three-dimensional point clouds of river topography, comprising the following steps: Obtain the original 3D point cloud data of the target river channel, and construct a manifold triangular mesh based on the original 3D point cloud data; Calculate the Laplace-Beltramian operator matrix of the manifold triangular mesh, perform eigenvalue decomposition on the Laplace-Beltramian operator matrix to obtain the manifold harmonic basis functions, and transform the manifold triangular mesh from Euclidean space to the spectral domain based on the manifold harmonic basis functions. The high-frequency noise components in the spectral domain are truncated using the constructed adaptive low-pass filter. The processed spectral coefficients are then inversely transformed back to Euclidean space to obtain the denoised river point cloud model. The weighted Laplace shrinkage algorithm is used to perform iterative shrinkage calculations on the river point cloud model, which collapses the three-dimensional river point cloud model into a one-dimensional set of points, and the river central axis skeleton line is extracted based on the set of points. An orthogonal plane is generated along the tangent direction of the river channel's central axis skeleton line. The river channel point cloud model is then projected onto the orthogonal plane, and the Alpha-Shape algorithm is used to extract the cross-sectional contours on each orthogonal plane. Using the central axis of the river channel as the scanning path, spline interpolation algorithm is used to connect the contours of each section to generate a continuous three-dimensional surface model of the river channel terrain.
[0019] In this embodiment, the steps for calculating the Laplace-Beltrami operator matrix of the manifold triangular mesh specifically include: traversing each vertex in the manifold triangular mesh, identifying the set of neighboring vertices adjacent to the vertex, calculating the edge vector connecting the vertex to each neighboring vertex in the set of neighboring vertices, calculating the average curvature of each vertex, and based on the topology of the manifold triangular mesh, calculating the cotangent value of the diagonal angle associated with the edge vector, introducing a terrain curvature sensitivity factor based on the average curvature to correct the cotangent value, constructing edge weight coefficients, and simultaneously calculating the area of the mixed region of each vertex to construct the quality matrix, generating a discretized Laplace-Beltrami operator matrix based on the edge weight coefficients and the quality matrix. The calculation of elements in the Laplace-Beltrami operator matrix is based on the following edge weight coefficient formula: Based on edge weight coefficients Off-diagonal elements of the Laplace-Beltrami operator matrix (When i ≠ j and j ∈ N(i)) and diagonal elements They are defined as follows: , ,in, Indicates the first The area of the mixed region at each vertex, Indicates the first The set of neighboring vertices of a vertex. and This represents an index in the set of neighboring vertices. and Indicates connection to vertices and vertex The two opposite diagonal angles of the common side, Indicates the first The average curvature of each vertex Indicates the first The average curvature of the neighboring vertices, This represents the curvature-sensitive control parameter. This represents the absolute value of the curvature difference. Represented by natural constant An exponential function with base 0.
[0020] Specifically, the process iterates through each vertex in the manifold triangular mesh, identifying the set of neighboring vertices adjacent to the vertex. This set is determined by setting a search radius (e.g., 0.05 meters) or directly querying the vertex index of a shared edge, ensuring that directly connected neighboring points are obtained. The edge vector connecting the vertex to each neighboring vertex in the set is calculated. A local quadratic surface is fitted to each vertex and its neighbors. The average curvature of each vertex is obtained by calculating the principal curvatures of this quadratic surface. Based on the topology of the manifold triangular mesh, a third vertex not adjacent to the edge vector is found in the triangular facet. The angle between the two edges at this third vertex is calculated using the cosine theorem or the inverse cosine formula of the dot product, thus calculating the diagonal angle associated with the edge vector. The cotangent value is modified by introducing a terrain curvature sensitivity factor based on the average curvature. The curvature sensitivity control parameter is set between 1.5 and 2 times the standard deviation of the average curvature difference of all vertices, for example, a value of 0.5. This is used to construct the edge weight coefficient. At the same time, the mixed region area of each vertex is calculated to construct the quality matrix. Specifically, for each vertex, the triangles around it are traversed. If the triangle is non-obtuse, its Voronoi diagram area is calculated. If it is obtuse, the area of its midpoint region is calculated. These regions are accumulated to obtain the mixed region area of the vertex. Based on the edge weight coefficient and the quality matrix, a discretized Laplace-Beltrami operator matrix is generated. The calculation of the elements in the Laplace-Beltrami operator matrix is based on the following edge weight coefficient formula: ,in Indicates connection to vertices and vertex The weight coefficients of the edges, and Indicates connection to vertices and vertex The two opposite diagonal angles of the common side, Indicates the first The average curvature of each vertex Indicates the first The average curvature of the neighboring vertices, This represents the curvature-sensitive control parameter. This represents the absolute value of the curvature difference. Represented by natural constant An exponential function with a base of 0, based on edge weight coefficients. Off-diagonal elements of the Laplace-Beltrami operator matrix (When i ≠ j and j ∈ N(i)) and diagonal elements They are defined as follows: , ,in, Indicates the first The area of the mixed region at each vertex, Indicates the first The set of neighboring vertices of a vertex. and This represents the index in the set of neighboring vertices, which is used to locate the position of an element in the matrix.
[0021] In this embodiment, the step of using the constructed adaptive low-pass filter to truncate high-frequency noise components in the spectral domain specifically includes: performing eigenvalue decomposition on the Laplace-Beltrami operator matrix to obtain an eigenvalue sequence sorted by size and a corresponding eigenvector sequence, where the eigenvector sequence is the manifold harmonic basis function; projecting the vertex coordinates of the manifold triangular mesh onto the eigenvector sequence to calculate the corresponding spectral coefficient vector; designing a cutoff frequency threshold based on energy proportion; and constructing a transfer function to retain low-frequency components below the cutoff frequency threshold and suppress high-frequency components above the cutoff frequency threshold. The transfer function formula for the adaptive low-pass filter is as follows: ,in, Indicates that for the first eigenvalues The filter gain coefficient, Represents the eigenvalue sequence of the eigenvalue sequence. 1 eigenvalue, Indicates the cutoff frequency threshold. This indicates the filter's order and is used to control the filter's roll-off rate. This represents the high-frequency suppression compensation coefficient. The maximum eigenvalue in the eigenvalue sequence is represented by the transfer function. The spectral coefficient vector is updated using the transfer function, and the updated spectral coefficient vector is linearly combined with the eigenvector sequence to complete the inverse transformation from the spectral domain to the spatial domain, generating a denoised river point cloud model.
[0022] Specifically, eigenvalue decomposition is performed on the Laplace-Beltrami operator matrix using the Arnoldi iterative algorithm or the Lanczos algorithm for sparse matrices. The number of eigenvalues obtained is set to 10% to 20% of the total number of vertices, resulting in a sequence of eigenvalues sorted by size and the corresponding eigenvector sequence. The eigenvector sequence is the manifold harmonic basis function. The vertex coordinates of the manifold triangular mesh are projected onto the eigenvector sequence. By calculating the dot product of the vertex coordinate vectors and the eigenvectors, the spatial coordinate information is converted into frequency domain coefficients, and the corresponding spectral coefficient vector is calculated. A cutoff frequency threshold based on energy proportion is designed, and a transfer function is constructed to retain low-frequency components below the cutoff frequency threshold and suppress high-frequency components above the cutoff frequency threshold. When constructing the transfer function, the filter order is set to, for example, 4, to control the steepness of the frequency response, and the high-frequency suppression compensation coefficient is set to 0.1 to prevent loss of detail due to excessive smoothing. The transfer function formula for the adaptive low-pass filter is: ,in, Indicates that for the first eigenvalues The filter gain coefficient, Represents the eigenvalue sequence of the eigenvalue sequence. 1 eigenvalue, Indicates the cutoff frequency threshold. This indicates the filter's order and is used to control the filter's roll-off rate. This represents the high-frequency suppression compensation coefficient. The maximum eigenvalue in the eigenvalue sequence is represented by the transfer function. The spectral coefficient vector is updated by multiplying the original spectral coefficients by the corresponding filter gain coefficients. The updated spectral coefficient vector is then linearly combined with the eigenvector sequence. The frequency domain signals are summed to perform a complete inverse transformation from the spectral domain to the spatial domain, generating a denoised river point cloud model.
[0023] In this embodiment, the steps for iteratively shrinking the river point cloud model based on the weighted Laplace shrinkage algorithm specifically include: initializing the shrinkage iteration parameters, using the denoised river point cloud model as the initial point cloud set, updating and calculating the Laplace weight coefficients between points in the point cloud set in each iteration, calculating the local neighborhood geometric center of each point, constructing a shrinkage vector pointing to the geometric center, and introducing a volume preservation constraint term to restrict excessive tangential movement while shrinking along the normal direction, preventing topological breaks in the river point cloud model during the shrinkage process, performing iterative update operations until the point cloud set no longer shifts, obtaining a converged zero-volume point set. The shrinkage formula for the iterative update operation is: ,in, Indicates the first After the nth iteration The three-dimensional coordinate vector of a point Indicates the first During the nth iteration The three-dimensional coordinate vector of a point Indicates the number of iterations Decreasing contraction rate factor Indicates the first A set of neighborhood indexes for each point Indicates based on current coordinates Calculated Laplace weighting coefficients, Represents the three-dimensional coordinate vector of neighborhood points. This represents the volume retention constraint strength coefficient. Indicates the first Normal vector at each point This represents the local volume change rate function. Indicates the first The local neighborhood volume at each point, and the local volume change rate function are used to provide feedback on the current volume compression state.
[0024] Specifically, the shrinkage iteration parameters are initialized, the initial iteration count counter is set to 0, and the maximum iteration count is set to, for example, 50. The denoised river point cloud model is used as the initial point cloud set. In each iteration, the Laplace weight coefficients between points in the point cloud set are updated and calculated, using cotangent weights or inverse distance weights as the basis. The local neighborhood geometric center of each point is calculated, which is the weighted average of the coordinates of all points in the neighborhood. A shrinkage vector pointing to the geometric center is constructed, i.e., the vector direction from the current point to the geometric center. A volume preservation constraint is introduced, which limits excessive tangential movement while shrinking along the normal direction by calculating the local tetrahedral volume or the bounding box volume formed by local neighborhood points, to prevent topological breaks in the river point cloud model during the shrinkage process. The iterative update operation is performed. After each step of coordinate update, the maximum displacement of all points is checked. If the maximum displacement is less than the preset convergence threshold, for example, 0.001 meters, the iteration stops until the point cloud set no longer shifts, resulting in a converged zero-volume point set. The shrinkage formula for the iterative update operation is: ,in, Indicates the first After the nth iteration The three-dimensional coordinate vector of a point Indicates the first During the nth iteration The three-dimensional coordinate vector of a point Indicates the number of iterations The shrinkage rate factor decreases, initially set to 0.5 and decreasing linearly with iteration. Indicates the first A set of neighborhood indexes for each point Indicates based on current coordinates Calculated Laplace weighting coefficients, Represents the three-dimensional coordinate vector of neighborhood points. This represents the volume retention constraint strength coefficient, for example, a value of 1.0. Indicates the first Normal vector at each point This represents the local volume change rate function. Indicates the first The local neighborhood volume at each point, and the local volume change rate function are used to provide feedback on the current volume compression state.
[0025] In this embodiment, the step of extracting the river channel central axis skeleton line based on the point set specifically includes: performing connectivity analysis on the converged zero-volume point set, constructing a minimum spanning tree structure, identifying the main branch path and spur branch path in the minimum spanning tree structure, removing spur branch paths with a ratio less than a preset threshold by calculating the ratio of the branch length to the main path length, fitting B-spline curves to the retained main branch paths, resampling the fitted curves at equal intervals to generate uniformly distributed skeleton nodes, calculating the tangent vector, normal vector, and binormal vector at the skeleton nodes, constructing a local Flyner frame that varies with the river channel central axis skeleton line, and correcting the smoothness of the river channel central axis skeleton line based on the local Flyner frame to ensure that the river channel central axis skeleton line conforms to the dynamic continuity characteristics of river flow.
[0026] Specifically, connectivity analysis is performed on the converged set of zero-volume points. A connection distance threshold is set, for example, 1.2 times the average spacing of the point cloud. Points with distances less than this threshold are connected to form a graph, and a minimum spanning tree (MPT) structure is constructed. Starting from any point in the graph, the Prim algorithm or Kruskal algorithm is used to generate a minimum spanning tree containing all points. The main branch paths and spurious branch paths in the MPT structure are identified, and the degree of each node is calculated. Nodes with a degree of 1 are used as endpoints, and the geodesic length of each branch path is calculated. By calculating the ratio of the branch length to the main path length, a pruning threshold of 0.15 is set, and spurious branch paths with a ratio less than the preset threshold are removed. B-spline curve fitting is performed on the retained main branch paths to select the path. Key nodes along the path are used as control points. Cubic B-spline basis functions are used for approximation. The fitted curve is resampled at equal intervals, with a sampling interval of 0.5 meters. Uniformly distributed skeleton nodes are generated by interpolation along the curve arc length. The tangent vector, normal vector, and binormal vector at the skeleton nodes are calculated. The first derivative is calculated using the curve differential geometry formula to obtain the tangent vector. The second derivative is calculated and combined with the tangent vector to obtain the normal vector. The binormal vector is obtained by using the cross product of the tangent vector and the normal vector. A local Flyner frame that varies with the river channel mid-axis skeleton line is constructed. The smoothness of the river channel mid-axis skeleton line is corrected based on the local Flyner frame. Smoothing interpolation is performed in areas where the tangent vector changes drastically to ensure that the river channel mid-axis skeleton line conforms to the dynamic continuity characteristics of the river flow.
[0027] In this embodiment, the step of generating an orthogonal plane along the tangent direction of the river channel's central axis skeleton line specifically includes: traversing the sampling nodes along the river channel's central axis skeleton line at a preset step size, calculating the tangent vector at each sampling node using the central difference method, constructing a plane equation passing through the sampling nodes based on the tangent vector, defining the plane equation as an orthogonal plane, determining the local point cloud subset in the denoised river channel point cloud model that is closest to the orthogonal plane, and projecting the local point cloud subset onto the orthogonal plane along the tangent vector direction to obtain a two-dimensional plane point set. The formula for constructing and projecting the orthogonal plane is as follows: ,in, Representing a spatial point In the Projected coordinates on an orthogonal plane Let represent the coordinate vector of any point in the local point cloud subset. The first line on the central axis of the river channel The coordinate vector of each sampling node, Indicates the first Tangent vectors at each sampling node This represents the vector dot product operation. It represents the magnitude of the vector.
[0028] Specifically, the sampling nodes along the central axis of the river channel are traversed at a preset step size. This preset step size is set based on the rate of change of the river channel width: 2 meters in areas with a width change rate less than 5%, and 0.5 meters in areas with a change rate greater than 5%. The tangent vector at each sampling node is calculated using the central difference method, i.e., the coordinate difference between the previous and next nodes of the current node is taken as the approximate tangent direction. Based on the tangent vector, the plane equation passing through the sampling nodes is constructed. The tangent vector is used as the normal vector of the plane, and the coordinates of the sampling node are taken as a point on the plane. The plane equation is defined as an orthogonal plane. The local point cloud subset closest to the orthogonal plane in the denoised river channel point cloud model is determined. The KD-Tree spatial indexing algorithm is used to find point cloud data whose vertical distance from the orthogonal plane is less than a set bandwidth, such as 0.2 meters. The local point cloud subset is projected onto the orthogonal plane along the tangent vector direction, and the coordinates of the projected points on the plane are calculated to obtain the two-dimensional plane point set. The construction and projection formula of the orthogonal plane is as follows: ,in, Representing a spatial point In the Projected coordinates on an orthogonal plane Let represent the coordinate vector of any point in the local point cloud subset. The first line on the central axis of the river channel The coordinate vector of each sampling node, Indicates the first Tangent vectors at each sampling node This represents the vector dot product operation. It represents the magnitude of the vector.
[0029] In this embodiment, the step of extracting the cross-sectional contour on each orthogonal plane using the Alpha-Shape algorithm specifically includes: performing Delaunay triangulation on the two-dimensional plane point set to generate a two-dimensional triangular mesh set; calculating the circumcircle radius of each triangle in the two-dimensional triangular mesh set; comparing the circumcircle radius with the Alpha threshold parameter; if the circumcircle radius is less than the Alpha threshold parameter, the corresponding triangle is retained; otherwise, the corresponding triangle is deleted; extracting the boundary edges of all retained triangles; connecting the boundary edges in sequence to form a closed polygon chain, which is the cross-sectional contour; and performing polygon simplification processing on the cross-sectional contour to reduce the number of vertices constituting the cross-sectional contour while keeping the contour shape feature error less than a preset value, thereby optimizing the amount of data for modeling.
[0030] Specifically, the Delaunay triangulation is performed on the two-dimensional planar point set. Connecting the points on the plane generates a series of non-overlapping triangles, creating a two-dimensional triangulation set. The circumcircle radius of each triangle in the triangulation set is calculated. All generated triangles are traversed, and the circumcircle radius is calculated using the three-point coordinate formula. The circumcircle radius is compared with an Alpha threshold parameter, which is set between 1.5 and 2.5 times the average side length of all triangles. If the circumcircle radius is less than the Alpha threshold parameter, the triangle is considered to be inside the contour and the corresponding triangle is retained; otherwise, the triangle is considered to be outside the contour. Empty or external regions are identified and the corresponding triangles are deleted. The boundary edges of all remaining triangles are extracted, and the frequency of each edge in the triangle is counted. Only the edge with a frequency of 1 is retained. The boundary edges are connected in sequence, and the first and last connected edge sequences are found to form a closed polygon chain, which is the cross-sectional profile. The cross-sectional profile is then simplified into polygons using the Douglas-Peucker algorithm. The vertical distance threshold is set to 0.05 meters. While keeping the profile shape feature error less than the preset value, vertices with small contribution to the shape are recursively removed to reduce the number of vertices constituting the cross-sectional profile and optimize the amount of data for modeling.
[0031] In this embodiment, the step of connecting the cross-sectional contours using a spline interpolation algorithm to generate a continuous three-dimensional surface model of the river channel terrain specifically includes: indexing and numbering all extracted cross-sectional contours according to their arrangement order on the river channel's central axis skeleton line; identifying corresponding feature points on two adjacent cross-sectional contours; establishing a mapping relationship between feature point pairs; and generating a transition surface mesh between adjacent cross-sectional contours using the Catmull-Rom spline function. The interpolation formula for the Catmull-Rom spline function is as follows: ,in, This represents the coordinates of a point on the surface generated by interpolation. This represents the interpolation parameter between 0 and 1. This represents the coordinates of the corresponding control points on four cross-sectional profiles continuously distributed along the central axis of the river channel. This can be achieved by changing the parameters. The value at the control point and Smooth, continuous surface vertices are generated between them.
[0032] Specifically, all extracted cross-sectional profiles are indexed and numbered according to their arrangement on the river channel's central axis, and are sequentially labeled from upstream to downstream as [1] to [2]. The process involves identifying corresponding feature points on two adjacent cross-sectional profiles, calculating the Euclidean distance or polar angle between these points to find pairs of feature points with similar shapes, establishing a mapping relationship between these pairs, and generating a transition surface mesh between adjacent cross-sectional profiles using the Catmull-Rom spline function. Corresponding points on four consecutive cross-sections are selected as control points, and interpolation is performed between the second and third cross-sections, with interpolation parameters set. The step size is 0.1, which means that 9 transition point sequences are generated between the two sections. The interpolation formula for the Catmull-Rom spline function is: ,in, This represents the coordinates of a point on the surface generated by interpolation. This represents the interpolation parameter between 0 and 1. This represents the coordinates of the corresponding control points on four cross-sectional profiles continuously distributed along the central axis of the river channel. This can be achieved by changing the parameters. The value at the control point and Smooth, continuous surface vertices are generated between them.
[0033] In this embodiment, the method further includes a step of optimizing the generated three-dimensional surface model of the river channel topography based on fluid dynamics constraints. Specifically, this includes: calculating the gradient along the river channel topography and the rate of change of the cross-sectional area of the cross-section; detecting, according to the principles of open channel fluid dynamics, whether there are abnormal abrupt change regions that violate the flow continuity equation; if abnormal abrupt change regions are detected, backtracking and adjusting the scale parameters of the cross-sectional profile at the corresponding location, and re-executing the step of connecting each cross-sectional profile using the spline interpolation algorithm until the three-dimensional surface model of the river channel topography meets the hydraulic smoothness requirements; outputting the final three-dimensional surface model of the river channel topography; and converting it into a digital elevation model format for storage.
[0034] Specifically, the gradient and cross-sectional area change rate of the three-dimensional surface model of the river channel are calculated. The elevation difference of the riverbed is calculated every 1 meter along the central axis as the gradient, and the polygon area of the corresponding cross-section and its derivative with respect to distance are calculated. Based on the principles of open channel fluid dynamics, and referring to Manning's formula and the continuity equation, a reasonable range of flow velocity and cross-sectional change is set. Abnormal abrupt changes that violate the flow continuity equation are detected. For example, if a cross-sectional area shrinks by more than 30% in a short distance without a corresponding drastic change in gradient, the scale parameters of the cross-sectional profile at the corresponding location are adjusted back, and the cross-sectional profile points at that location are scaled or smoothed. The steps of connecting the cross-sectional profiles using the spline interpolation algorithm are re-executed, the surface is regenerated and checked, until the three-dimensional surface model of the river channel meets the hydraulic smoothness requirements, that is, the change rate of all detected cross-sections is within the preset allowable range. The final three-dimensional surface model of the river channel is then output and converted into a digital elevation model format for storage.
[0035] In this embodiment, the step of designing a cutoff frequency threshold based on energy percentage specifically includes: calculating the spectral energy corresponding to each feature value in the feature value sequence, calculating the cumulative spectral energy distribution, setting an energy retention ratio threshold, searching for a critical feature value index that satisfies the energy retention ratio threshold in the cumulative spectral energy distribution, and determining the feature value corresponding to the critical feature value index as the cutoff frequency threshold. The value range of the energy retention ratio threshold is 85% to 95%.
[0036] Specifically, the spectral energy corresponding to each feature value in the feature value sequence is calculated. The square of each spectral coefficient is taken as the energy value of that frequency component. The cumulative spectral energy distribution is calculated. The sorted energy values are summed and divided by the total energy value to obtain the normalized cumulative energy curve. An energy retention ratio threshold is set. Based on the prediction of the point cloud noise level, it is set to 85% if the noise level is high and 95% if it is low. The critical feature value index that meets the energy retention ratio threshold is found in the cumulative spectral energy distribution. The cumulative energy array is traversed to find the index position of the first element that is greater than or equal to the set threshold. The feature value corresponding to the critical feature value index is determined as the cutoff frequency threshold. The value range of the energy retention ratio threshold is 85% to 95%.
[0037] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.
Claims
1. A method for dynamic reconstruction of three-dimensional point clouds of river channel topography, characterized in that, Includes the following steps: Obtain the original 3D point cloud data of the target river channel, and construct a manifold triangular mesh based on the original 3D point cloud data; Calculate the Laplace-Beltramian operator matrix of the manifold triangular mesh, perform eigenvalue decomposition on the Laplace-Beltramian operator matrix to obtain the manifold harmonic basis functions, and transform the manifold triangular mesh from Euclidean space to the spectral domain based on the manifold harmonic basis functions; The high-frequency noise components in the spectral domain are truncated using the constructed adaptive low-pass filter. The processed spectral coefficients are then inversely transformed back to Euclidean space to obtain the denoised river point cloud model. The river point cloud model is iteratively shrunk using a weighted Laplace shrinkage algorithm, which collapses the three-dimensional river point cloud model into a one-dimensional set of points. The river central axis skeleton line is then extracted based on the set of points. An orthogonal plane is generated along the tangent direction of the central axis skeleton line of the river channel. The point cloud model of the river channel is projected onto the orthogonal plane, and the cross-sectional contour on each orthogonal plane is extracted using the Alpha-Shape algorithm. Using the central axis skeleton line of the river channel as the scanning path, spline interpolation algorithm is used to connect the contours of each section to generate a continuous three-dimensional surface model of the river channel terrain.
2. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 1, characterized in that, The steps for calculating the Laplace-Beltramian operator matrix of the manifold triangular mesh specifically include: Traverse each vertex in the manifold triangular mesh, identify the set of neighboring vertices adjacent to the vertex, and calculate the edge vector connecting the vertex to each neighboring vertex in the set of neighboring vertices; Calculate the average curvature of each vertex, and based on the topology of the manifold triangular mesh, calculate the cotangent of the diagonal angle associated with the edge vector; A terrain curvature sensitivity factor based on the average curvature is introduced to correct the cotangent value, an edge weight coefficient is constructed, and the area of the mixed region of each vertex is calculated to construct a quality matrix. The discretized Laplace-Beltrami operator matrix is generated based on the edge weight coefficients and the mass matrix; The elements in the Laplace-Beltrami operator matrix are calculated based on the following formula for edge weight coefficients: ; Based on the edge weight coefficient The off-diagonal elements of the Laplace-Beltrami operator matrix and diagonal elements They are defined as follows: , ; in, Indicates the first The area of the mixed region of each of the vertices, Indicates the first The set of neighboring vertices of the given vertex. and This represents the index in the set of neighboring vertices. and Indicates connection to vertices and vertex The two opposite angles of the common side, Indicates the first The average curvature of the vertices. Indicates the first The average curvature of the neighboring vertices This represents the curvature-sensitive control parameter. This represents the absolute value of the curvature difference. Represented by natural constant An exponential function with base 0.
3. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 2, characterized in that, The steps for truncating high-frequency noise components in the spectral domain using a constructed adaptive low-pass filter specifically include: The Laplace-Beltrami operator matrix is subjected to eigenvalue decomposition to obtain an eigenvalue sequence sorted by size and a corresponding eigenvector sequence. The eigenvector sequence is the manifold harmonic basis function. The vertex coordinates of the manifold triangular mesh are projected onto the feature vector sequence to calculate the corresponding spectral coefficient vector; Design a cutoff frequency threshold based on energy percentage, and construct a transfer function to retain low-frequency components below the cutoff frequency threshold and suppress high-frequency components above the cutoff frequency threshold; The transfer function formula of the adaptive low-pass filter is: ; in, Indicates that for the first eigenvalues The filter gain coefficient, Represents the th in the sequence of eigenvalues 1 eigenvalue, This represents the cutoff frequency threshold. This indicates the filter's order and is used to control the filter's roll-off rate. This represents the high-frequency suppression compensation coefficient. This represents the largest eigenvalue in the eigenvalue sequence; The spectral coefficient vector is updated using the transfer function, and the updated spectral coefficient vector is linearly combined with the feature vector sequence to complete the inverse transformation from the spectral domain to the spatial domain, generating the denoised river point cloud model.
4. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 1, characterized in that, The steps for iteratively shrinking the river point cloud model based on the weighted Laplace shrinkage algorithm specifically include: Initialize the shrinkage iteration parameters, and use the denoised river point cloud model as the initial point cloud set; In each iteration, the Laplacian weight coefficients between points in the point cloud set are updated and the local neighborhood geometric center of each point is calculated. A contraction vector pointing to the geometric center is constructed, and a volume-preserving constraint term is introduced to restrict excessive tangential movement while contracting along the normal direction, thereby preventing topological breaks in the river point cloud model during the contraction process. Perform iterative update operations until the point cloud set no longer shifts, and obtain the converged zero-volume point set; The shrinkage formula for the iterative update operation is: ; in, Indicates the first After the nth iteration The three-dimensional coordinate vector of a point Indicates the first During the nth iteration The three-dimensional coordinate vector of a point Indicates the number of iterations Decreasing contraction rate factor Indicates the first The neighborhood index set of each point Indicates based on current coordinates Calculated Laplace weighting coefficients, Represents the three-dimensional coordinate vector of the neighborhood points. This represents the volume retention constraint strength coefficient. Indicates the first Normal vector at each point This represents the local volume change rate function. Indicates the first The local neighborhood volume at each point, and the local volume change rate function are used to provide feedback on the current volume compression state.
5. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 4, characterized in that, The steps for extracting the river channel central axis skeleton line based on the aforementioned point set specifically include: Connectivity analysis is performed on the converged set of zero-volume points to construct a minimum spanning tree graph structure; Identify the main branch path and spurious branch path in the minimum spanning tree graph structure, and remove the spurious branch path whose ratio is less than a preset threshold by calculating the ratio of the branch length to the main path length. The retained main branch path is fitted with a B-spline curve, and the fitted curve is resampled at equal intervals to generate uniformly distributed skeleton nodes. Calculate the tangent vector, normal vector, and binormal vector at the skeleton nodes to construct a local Fleischer frame that varies with the river channel central axis skeleton line; The smoothness of the river channel central axis skeleton line is corrected based on the local Flyner frame to ensure that the river channel central axis skeleton line conforms to the dynamic continuity characteristics of river flow.
6. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 1, characterized in that, The step of generating an orthogonal plane along the tangent direction of the central axis of the river channel specifically includes: The sampling nodes on the central axis of the river channel are traversed according to a preset step size, and the tangent vector at each sampling node is calculated using the central difference method. Based on the tangent vector, construct the plane equation passing through the sampling node, and define the plane equation as the orthogonal plane; Determine the subset of local point clouds in the denoised river point cloud model that is closest to the orthogonal plane; The local point cloud subset is projected onto the orthogonal plane along the direction of the tangent vector to obtain a two-dimensional planar point set.
7. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 6, characterized in that, The steps for extracting the cross-sectional profile on each of the orthogonal planes using the Alpha-Shape algorithm specifically include: The two-dimensional planar point set is subjected to Delaunay triangulation to generate a two-dimensional triangulation network set; Calculate the circumcircle radius of each triangle in the two-dimensional triangulation set, and compare the circumcircle radius with the Alpha threshold parameter; If the radius of the circumcircle is smaller than the Alpha threshold parameter, the corresponding triangle is retained; otherwise, the corresponding triangle is deleted. Extract the boundary edges of all the retained triangles, and connect the boundary edges in sequence to form a closed polygon chain, which is the cross-sectional profile; The cross-sectional contour is simplified by polygon processing. While keeping the contour shape feature error less than a preset value, the number of vertices constituting the cross-sectional contour is reduced, thereby optimizing the amount of data for modeling.
8. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 1, characterized in that, The steps of connecting the cross-sectional contours using a spline interpolation algorithm to generate a continuous three-dimensional surface model of the river channel topography specifically include: All extracted cross-sectional profiles are indexed and numbered according to their arrangement order along the central axis of the river channel; Identify corresponding feature points on two adjacent cross-sectional contours and establish a mapping relationship between feature point pairs; The Catmull-Rom spline function is used to generate transition surface meshes between adjacent cross-sectional profiles.
9. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 1, characterized in that, The method further includes a step of optimizing the generated three-dimensional surface model of the river channel topography based on fluid dynamic constraints, specifically including: Calculate the gradient along the course and the rate of change of cross-sectional area of the water passage in the three-dimensional surface model of the river channel topography; Based on the principles of open channel fluid mechanics, detect whether there are abnormal abrupt changes that violate the equation of continuity of water flow; If the abnormal abrupt change region is detected, the scale parameters of the cross-sectional profile at the corresponding location are adjusted back, and the step of connecting each cross-sectional profile using the spline interpolation algorithm is re-executed until the three-dimensional surface model of the river channel topography meets the hydraulic smoothness requirements. The final three-dimensional surface model of the river channel topography is output and converted into a digital elevation model format for storage.
10. The method for dynamic reconstruction of three-dimensional point clouds of river topography according to claim 3, characterized in that, The steps for designing a cutoff frequency threshold based on energy percentage include: Calculate the spectral energy corresponding to each feature value in the feature value sequence, and calculate the cumulative spectral energy distribution; Set an energy retention ratio threshold, and search for critical feature value indices that satisfy the energy retention ratio threshold in the cumulative spectral energy distribution; The feature value corresponding to the critical feature value index is determined as the cutoff frequency threshold; The energy retention ratio threshold ranges from 85% to 95%.
