Three-dimensional point cloud data filtering method and system based on adaptive clustering segmentation and gradient compensation

Abstract

The invention discloses a three-dimensional point cloud data filtering method and system based on adaptive clustering segmentation and gradient compensation, and the method comprises the steps: carrying out the data preprocessing of three-dimensional point cloud data through employing a statistical filtering method, obtaining the preprocessed point cloud data, extracting the point cloud features in the preprocessed point cloud data, and obtaining the point cloud feature data; according to the invention, a function of establishing multi-dimensional features including point cloud density, spatial autocorrelation and local curvature features by using an adaptive clustering segmentation algorithm based on density spatial distribution and carrying out dynamic clustering segmentation on a ground feature-ground cluster is realized; and residual mixed point clouds can be further separated through a coarse-fine granularity grid grading processing strategy, so that ground point clouds are separated, and the technical limitations of high parameter dependence, insufficient terrain adaptability and poor real-time performance in the prior art are broken through; and the filtering precision and robustness in a complex scene are remarkably improved through multi-dimensional collaborative optimization, and the method is suitable for being widely popularized and used.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional point cloud data filtering, and in particular to a three-dimensional point cloud data filtering method and system based on adaptive clustering segmentation and gradient compensation. Background Art

[0002] 3D point cloud data accurately recreates the three-dimensional spatial information of the real world by recording a large number of 3D coordinate points on an object's surface and their properties (such as color and reflection intensity). 3D point cloud data is widely used in autonomous driving, 3D reconstruction, industrial manufacturing, geography and agriculture, medical biology, and daily life.

[0003] At present, most of the existing 3D point cloud data filtering algorithms, such as mathematical morphology, surface methods, and cluster segmentation, have significant technical limitations and generally rely on manually preset structural element sizes, window thresholds, or initial DEM parameters. This not only leads to under-filtering or filtering in complex scenes such as steep slopes, dense vegetation, and mixed buildings, but also makes it difficult for single-scale processing strategies to balance accuracy and efficiency. In particular, failure is prone to occur in areas where low objects overlap with ground points. In addition, due to gradient sensitivity defects, existing methods are unable to effectively deal with local slope mutations in non-uniform terrain such as cliffs and gullies. Therefore, it is necessary to design a 3D point cloud data filtering method and system based on adaptive cluster segmentation and gradient compensation. Summary of the Invention

[0004] The purpose of the present invention is to overcome the shortcomings of the existing technology, in order to better and effectively solve the existing 3D point cloud data filtering algorithms such as mathematical morphology, surface method and cluster segmentation. Most of them have significant technical limitations and generally rely on manually preset structural element size, window threshold or initial DEM parameters, which not only leads to under-filtering or filtering wave phenomenon in complex scenes such as steep slopes, dense vegetation and mixed buildings, but also makes it difficult for a single-scale processing strategy to balance accuracy and efficiency, especially in the overlapping area of low objects and ground points. In addition, due to the gradient sensitivity defect, the existing methods are difficult to effectively deal with the problem of sudden local slope changes in non-uniform terrain such as cliffs and gullies. A 3D point cloud data based on adaptive cluster segmentation and gradient compensation is provided. The filtering method and system realize the function of establishing multi-dimensional features including point cloud density, spatial autocorrelation and local curvature characteristics and dynamically clustering and segmenting the ground object-ground cluster by using an adaptive clustering segmentation algorithm based on density spatial distribution. The building cluster points can be separated by introducing linear feature parameters, and the residual mixed point cloud can be further separated by a coarse-fine granularity grid hierarchical processing strategy, thereby separating the ground point cloud. It not only breaks through the technical limitations of traditional parameters with strong dependence, insufficient terrain adaptability and poor real-time performance, but also significantly improves the filtering accuracy and robustness in complex scenarios through multi-dimensional collaborative optimization. This provides the advantages of an efficient solution with both theoretical innovation and engineering practical value for the generation of high-precision digital elevation models.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is:

[0006] A three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation includes the following steps:

[0007] Step A, using a statistical filtering method to preprocess the three-dimensional point cloud data and obtain preprocessed point cloud data;

[0008] Step B, extracting point cloud features from the pre-processed point cloud data and obtaining point cloud feature data;

[0009] Step C, performing adaptive clustering segmentation on the point cloud feature data and obtaining the clustering segmentation result;

[0010] Step D: Delaunay triangulation is used to determine the spatial relationship of each point in the cluster segmentation result, thereby obtaining the violation points in the point cloud scene;

[0011] Step E, performing linear exponential solution to extract the violation points in the point cloud scene and obtaining the building violation points;

[0012] Step F, identifying the building outline based on the cluster segmentation results, the violation points in the point cloud scene, and the violation points of the building to obtain the building point cloud;

[0013] Step G: Remove the building point cloud from the 3D point cloud data to obtain the remaining point cloud data, then use the statistical filtering method to filter the remaining point cloud data and establish a plane interpolation model. Then, interpolate the target scene point cloud according to the plane interpolation model to obtain the grid plane fitting reference point cloud.

[0014] Step H, using the slope angle grid to perform point cloud filtering on the grid plane fitting reference point cloud and obtain filtered three-dimensional point cloud data;

[0015] Step I: Check the edge contour points in the filtered three-dimensional point cloud data and complete the three-dimensional point cloud data filtering operation.

[0016] The aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, step A, uses a statistical filtering method to preprocess the three-dimensional point cloud data and obtain preprocessed point cloud data, wherein the statistical filtering method specifically calculates the average distance μ between each point in the three-dimensional point cloud data and its neighboring points, and then obtains the distance threshold d from the average distance μ threshold , as shown in formula (1),

[0017] d threshold =μ+k·σ(1)

[0018] Where k is a constant factor and σ is the standard deviation of the distance between each point and its neighboring points.

[0019] The aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, step B, extracts point cloud features from the preprocessed point cloud data and obtains point cloud feature data, wherein the point cloud features include point cloud scene average density features, point cloud scene normal vector features, and point cloud scene neighboring point spatial distance features. The specific steps are as follows:

[0020] Step B1: extract the average density feature of the point cloud scene. The average density feature of the point cloud scene is calculated as shown in formula (2):

[0021] AD=SumD / NP(2)

[0022] Among them, AD is the average density of the preprocessed point cloud data, SumD is the sum of the distances between each point and the nearest point in the preprocessed point cloud data, and NP is the number of point clouds in the preprocessed point cloud data;

[0023] Step B2: extract the normal vector features of the point cloud scene, where the normal vector features include the first covariance matrix, eigenvalues, eigenvectors, and normal vector angles. The specific steps are as follows:

[0024] Step B21, calculate the first covariance matrix, specifically, establish a KD-Tree and find the corresponding m adjacent points, then construct an m*3 matrix through the m point coordinates and calculate the first covariance matrix of the m*3 matrix, as shown in formula (3),

[0025]

[0026] Where Cov(X) is the first covariance matrix of matrix X, μ x is the average value of each dimension, and T is the matrix transpose;

[0027] Step B22, calculate the eigenvalue and eigenvector, as shown in formula (4),

[0028]

[0029] Among them, λ is the eigenvalue and x is the eigenvector;

[0030] Step B23, calculate the normal vector angle, as shown in formula (5),

[0031]

[0032] Among them, a1, b1, c1, a2, b2 and c2 are the components of the normal vector of each point, and θ is the angle between the normal vectors of two points;

[0033] Step B3: extract the spatial distance features of neighboring points in the point cloud scene, as shown in formula (6):

[0034]

[0035] Among them, Dis is the distance between two points in space, (x1, y1, z1) and (x2, y2, z2) are the three-dimensional coordinates of the target point.

[0036] In the aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, step C performs adaptive clustering segmentation on the point cloud feature data and obtains the clustering segmentation result, wherein the clustering threshold of the adaptive clustering segmentation can be changed with the spatial characteristics of the point cloud, and the clustering threshold calculation process is shown in formula (7),

[0037]

[0038] Among them, YC is the clustering threshold of two points, and Δh is the height difference between two points.

[0039] In the aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, step D uses the Delaunay triangulation to determine the spatial relationship between each point in the clustering segmentation result, thereby obtaining the violation points in the point cloud scene. The specific steps are as follows:

[0040] Step D1, using the Delaunay triangulation to determine the spatial relationship of each point in the cluster segmentation result, specifically, calculating the normal vectors of adjacent triangles in the Delaunay triangulation, as shown in formula (8),

[0041]

[0042] in, and is the edge vector formed by two points in the triangle, (x a ,y a , z a )、(x b ,y b , z b ) and (x c ,y c , z c ) are the coordinates of the three vertices of the triangle, is the triangle normal vector;

[0043] Step D2: Obtain the violation points in the point cloud scene, specifically using two triangular normal vectors Determine whether it is a violation triangle. If the two triangle normal vectors If the angle between the highest point and the lowest point of the violating triangle exceeds the set angle, it is a violating triangle. If the height difference between the highest point and the lowest point of the violating triangle meets the set threshold, the highest point of the violating triangle is taken as the violating point.

[0044] The aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, step E, extracts the violation points in the point cloud scene by linear exponential solution and obtains the violation points of the building. Specifically, the obtained violation points in the point cloud scene are extracted and organized into a point cloud set and each point in the point cloud set is judged point by point. The specific steps are as follows:

[0045] Step E1: extract the K nearest neighboring points around the target point to construct a local neighborhood point set and translate the coordinate system so that the center of mass of the point cloud set is the origin, and then calculate the second covariance matrix as shown in formula (9):

[0046]

[0047] Among them, H is the second covariance matrix, n is the number of neighborhood points, x j is the three-dimensional coordinate of the j-th point, is the centroid coordinate of the neighborhood point;

[0048] Step E2, decomposing the eigenvalues of the second covariance matrix H, as shown in formula (10),

[0049]

[0050] Where V is an orthogonal matrix, Λ is a diagonal matrix, λ1 and λ2 are the largest eigenvalue and the second largest eigenvalue of the covariance matrix H, respectively, λ3 is the eigenvalue of the second covariance matrix H, and diag is the square decomposition of the second covariance matrix H;

[0051] In step E3, the obtained eigenvalues are subjected to linear index calculation to obtain the building violation points, wherein the linear index calculation is shown in formula (11):

[0052]

[0053] Among them, Linearity is the linear index of the violation point in the point cloud scene.

[0054] The aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, step F, identifies the building outline based on the clustering segmentation result, the violation points in the point cloud scene and the building violation points and obtains the building point cloud, specifically, compares the cluster outer contour points in the clustering segmentation result with the number of violation points in the point cloud scene belonging to the cluster, if the ratio of the number of violation points in the point cloud scene belonging to the cluster meets the set threshold, then compares the number of violation points in the building with the number of violation points in the point cloud scene, if the ratio of the number of violation points in the building meets the set threshold, then calculates the two-dimensional bounding box area of the cluster point in the clustering segmentation result, thereby obtaining the diagonal angle, and then obtaining the distance to the farthest point in the cluster, and if the sine value of the diagonal angle is less than the set value, the cluster belongs to the building point cloud, wherein the two-dimensional bounding box area is calculated using the built-in area calculation function polyarea in MATLAB as shown in formula (12),

[0055]

[0056] Where S is the area of the geometric figure, d is the side length, and β is the angle between the two sides.

[0057] In the aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, in step G, the building point cloud is removed from the three-dimensional point cloud data and the remaining point cloud data is obtained. The remaining point cloud data is then filtered using a statistical filtering method and a plane interpolation model is established. The target scene point cloud is then interpolated according to the plane interpolation model to obtain a grid plane fitting reference point cloud. The coordinates of each point in the plane interpolation model are shown in formula (13).

[0058] ax+by+cz+d=0(13)

[0059] Among them, a, b and c are all fitting plane parameters, and the fitting plane parameter constraints are a 2 +b 2 +c2 =1.

[0060] The aforementioned three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, step H, uses the slope angle grid to perform point cloud filtering on the grid plane fitting reference point cloud and obtain the filtered three-dimensional point cloud data. The specific steps are as follows:

[0061] Step H1: Take the nine divided grids as a whole to solve the slope angle of the adjacent grids of the target grid, as shown in formula (14):

[0062]

[0063] Among them, φ i is the slope angle of adjacent grids in the nine-square grid, (x i 、y i 、z i ) is the three-dimensional coordinate of the lowest point of the adjacent grid, (x ⑤ ,y ⑤ , z ⑤ ) is the three-dimensional coordinate of the lowest point to be determined in the target grid;

[0064] Step H2: Integrate the surrounding grids of the target grid in the nine-square grid, and then obtain the weighted slope angle corresponding to the target grid by distance weighting, as shown in formula (15):

[0065]

[0066] Among them, D i The Euclidean distance between the lowest point of the adjacent grid and the undetermined lowest point of the target grid, D is the weighted Euclidean distance, and γ is the slope angle of the target grid;

[0067] Step H3: Set the slope angle threshold to calculate the first elevation threshold. The specific calculation of the first elevation threshold is shown in formula (16):

[0068]

[0069] Among them, h t is the first elevation threshold, α is the slope threshold corresponding to the slope angle, and L is the grid division length;

[0070] In step H4, the elevation plane equation is established using the elevation threshold and formula (13), and then the elevation plane equation is solved to obtain the second elevation threshold, as shown in formula (17).

[0071]

[0072] Among them, z t is the second elevation threshold;

[0073] Step H5: comparing the first elevation threshold and the second elevation threshold. If the second elevation threshold is lower than the first elevation threshold, the target grid's undetermined lowest point is a ground point. If the second elevation threshold is not lower than the first elevation threshold, the target grid's undetermined lowest point is a non-ground point.

[0074] Step I, checks the edge contour points in the filtered three-dimensional point cloud data and completes the three-dimensional point cloud data filtering operation. Specifically, the violation points in the point cloud scene in step D and the cluster edge points in step C are re-taken as a whole point cloud set and the building point cloud in step F is removed therefrom. The remaining point cloud is then re-assigned to the corresponding cluster in step C. Then, the number of ground points obtained by step H within the set distance range of each point in the remaining point cloud is retrieved point by point. If the number of ground points within the set distance range meets the set threshold, it is further determined whether there is a point within the retrieval distance range whose height difference with the target grid's pending lowest point does not exceed the set threshold. If there is a point whose height difference with the target grid's pending lowest point does not exceed the set threshold, the target grid's pending lowest point is classified as a ground point.

[0075] A three-dimensional point cloud data filtering system based on adaptive clustering segmentation and gradient compensation includes a preprocessing module, a feature extraction module, an adaptive clustering segmentation module, a point cloud scene violation point retrieval module, a building violation point retrieval module, a building outline recognition module, a point cloud interpolation module, a point cloud filtering module and an edge contour point verification module. The preprocessing module is used to use a statistical filtering method to perform data preprocessing on the three-dimensional point cloud data and obtain preprocessed point cloud data; the feature extraction module is used to extract point cloud features in the preprocessed point cloud data and obtain point cloud feature data; the adaptive clustering segmentation module is used to perform adaptive clustering segmentation on the point cloud feature data and obtain clustering segmentation results; the point cloud scene violation point retrieval module is used to use a Delaunay triangulation to determine the spatial relationship between each point in the clustering segmentation result to obtain the point cloud scene violation points; The building violation point retrieval module is used to perform linear exponential solution to extract the violation points in the point cloud scene and obtain the building violation points; the building outline recognition module is used to identify the building outline based on the cluster segmentation results, the violation points in the point cloud scene and the building violation points and obtain the building point cloud; the point cloud interpolation module is used to remove the building point cloud from the three-dimensional point cloud data and obtain the remaining point cloud data, and then use the statistical filtering method to filter the remaining point cloud data and establish a plane interpolation model, and then perform point cloud interpolation on the target scene point cloud according to the plane interpolation model and obtain the grid plane fitting reference point cloud; the point cloud filtering module is used to use the slope angle grid to perform point cloud filtering on the grid plane fitting reference point cloud and obtain the filtered three-dimensional point cloud data; the edge contour point verification module is used to verify the edge contour points in the filtered three-dimensional point cloud data and complete the three-dimensional point cloud data filtering operation.

[0076] The beneficial effects of the present invention are as follows: a three-dimensional point cloud data filtering method and system based on adaptive clustering segmentation and gradient compensation of the present invention first uses a statistical filtering method to preprocess the three-dimensional point cloud data and obtain preprocessed point cloud data, then extracts the point cloud features in the preprocessed point cloud data and obtains point cloud feature data, then performs adaptive clustering segmentation on the point cloud feature data and obtains a clustering segmentation result, then uses the Delaunay triangulation to determine the spatial relationship of each point in the clustering segmentation result to obtain the point cloud scene violation points, then performs linear exponential solution to extract the point cloud scene violation points and obtains the building violation points, then identifies the building outline based on the clustering segmentation result, the point cloud scene violation points and the building violation points and obtains the building point cloud, then removes the building point cloud from the three-dimensional point cloud data and obtains the remaining point cloud data, then uses a statistical filtering method to filter the remaining point cloud data and establishes a plane interpolation model, then performs point cloud interpolation on the target scene point cloud according to the plane interpolation model and obtains a grid Plane fitting benchmark point cloud, finally use the slope angle grid to perform point cloud filtering on the grid plane fitting benchmark point cloud and obtain the filtered three-dimensional point cloud data, then check the edge contour points in the filtered three-dimensional point cloud data and complete the three-dimensional point cloud data filtering operation; effectively realize the three-dimensional point cloud data filtering method and system. Specifically, an adaptive clustering segmentation algorithm based on density spatial distribution is used to establish multi-dimensional features including point cloud density, spatial autocorrelation and local curvature characteristics and dynamically cluster and segment the ground object-ground cluster. By introducing linear feature parameters, the building cluster points can be separated, and the coarse-fine grained grid hierarchical processing strategy can further separate the residual mixed point cloud, thereby separating the ground point cloud. It not only breaks through the traditional technical limitations of strong parameter dependence, insufficient terrain adaptability and poor real-time performance, but also significantly improves the filtering accuracy and robustness in complex scenes through multi-dimensional collaborative optimization. This provides an efficient solution for the generation of high-precision digital elevation models with both theoretical innovation and engineering practical value. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is an overall flow chart of a three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation of the present invention;

[0078] Figure 2 It is a schematic diagram of the adaptive cluster segmentation of the present invention;

[0079] Figure 3 This is a schematic diagram of a violation point search without using the linear index proposed by the present invention;

[0080] Figure 4 Schematic diagram of linear index solution and extraction of violation points of the present invention;

[0081] Figure 5It is a schematic diagram of grid division of the present invention;

[0082] Figure 6 It is a schematic diagram of establishing a slope angle grid of the present invention;

[0083] Figure 7 This is a schematic diagram of the center-weighted slope angle solution of the present invention;

[0084] Figure 8 is a schematic diagram of establishing the first elevation threshold of the present invention;

[0085] Figure 9 This is the edge profile inspection result diagram of the present invention. DETAILED DESCRIPTION

[0086] The present invention will be further described below with reference to the accompanying drawings.

[0087] like Figure 1 As shown, the present invention provides a three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, comprising the following steps:

[0088] Step A: Preprocess the 3D point cloud data using a statistical filtering method to obtain preprocessed point cloud data. Specifically, the statistical filtering method calculates the average distance μ between each point in the 3D point cloud data and its neighboring points, and then obtains the distance threshold d from the average distance μ. threshold , as shown in formula (1),

[0089] d threshold =μ+k·σ(1)

[0090] Where k is a constant factor and σ is the standard deviation of the distance between each point and its neighboring points.

[0091] Step B, extracting point cloud features from the pre-processed point cloud data and obtaining point cloud feature data, wherein the point cloud features include point cloud scene average density features, point cloud scene normal vector features, and point cloud scene neighboring point spatial distance features. The specific steps are as follows:

[0092] Step B1: extract the average density feature of the point cloud scene. The average density feature of the point cloud scene is calculated as shown in formula (2):

[0093] AD=SumD / NP(2)

[0094] Among them, AD is the average density of the preprocessed point cloud data, SumD is the sum of the distances between each point and the nearest point in the preprocessed point cloud data, and NP is the number of point clouds in the preprocessed point cloud data;

[0095] Step B2: extract the normal vector features of the point cloud scene, where the normal vector features include the first covariance matrix, eigenvalues, eigenvectors, and normal vector angles. The specific steps are as follows:

[0096] Step B21, calculate the first covariance matrix, specifically, establish a KD-Tree and find the corresponding m adjacent points, then construct an m*3 matrix through the m point coordinates and calculate the first covariance matrix of the m*3 matrix, as shown in formula (3),

[0097]

[0098] Where Cov(X) is the first covariance matrix of matrix X, μ x is the average value of each dimension, and T is the matrix transpose;

[0099] Step B22, calculate the eigenvalue and eigenvector, as shown in formula (4),

[0100]

[0101] Among them, λ is the eigenvalue and x is the eigenvector;

[0102] Step B23, calculate the normal vector angle, as shown in formula (5),

[0103]

[0104] Among them, a1, b1, c1, a2, b2 and c2 are the components of the normal vector of each point, and θ is the angle between the normal vectors of two points;

[0105] Step B3: extract the spatial distance features of neighboring points in the point cloud scene, as shown in formula (6):

[0106]

[0107] Among them, Dis is the distance between two points in space, (x1, y1, z1) and (x2, y2, z2) are the three-dimensional coordinates of the target point.

[0108] like Figure 2 As shown, step C performs adaptive clustering segmentation on the point cloud feature data and obtains the clustering segmentation result, wherein the clustering threshold of the adaptive clustering segmentation can be changed with the spatial characteristics of the point cloud, and the clustering threshold calculation process is shown in formula (7),

[0109]

[0110] Among them, YC is the clustering threshold of two points, and Δh is the height difference between two points.

[0111] Step D: Use the Delaunay triangulation to determine the spatial relationship of each point in the cluster segmentation result, so as to obtain the violation points in the point cloud scene. The specific steps are as follows:

[0112] Step D1, using the Delaunay triangulation to determine the spatial relationship of each point in the cluster segmentation result, specifically, calculating the normal vectors of adjacent triangles in the Delaunay triangulation, as shown in formula (8),

[0113]

[0114] in, and is the edge vector formed by two points in the triangle, (x a ,y a , z a )、(x b ,y b , z b ) and (x c ,y c , z c ) are the coordinates of the three vertices of the triangle, is the triangle normal vector;

[0115] Step D2: Obtain the violation points in the point cloud scene, specifically using two triangular normal vectors Determine whether it is a violation triangle. If the two triangle normal vectors If the angle between the highest point and the lowest point of the violating triangle exceeds the set angle, it is a violating triangle. If the height difference between the highest point and the lowest point of the violating triangle meets the set threshold, the highest point of the violating triangle is taken as the violating point.

[0116] Such as 3 and Figure 4 As shown, step E is to perform linear exponential solution to extract the violation points in the point cloud scene and obtain the violation points of the building. Specifically, the violation points in the obtained point cloud scene are extracted and sorted into a point cloud set and each point in the point cloud set is judged point by point. The specific steps are as follows:

[0117] Step E1: extract the K nearest neighboring points around the target point to construct a local neighborhood point set and translate the coordinate system so that the center of mass of the point cloud set is the origin, and then calculate the second covariance matrix as shown in formula (9):

[0118]

[0119] Among them, H is the second covariance matrix, n is the number of neighborhood points, x j is the three-dimensional coordinate of the j-th point, is the centroid coordinate of the neighborhood point;

[0120] Step E2, decomposing the eigenvalues of the second covariance matrix H, as shown in formula (10),

[0121]

[0122] Where V is an orthogonal matrix, Λ is a diagonal matrix, λ1 and λ2 are the largest eigenvalue and the second largest eigenvalue of the covariance matrix H, respectively, λ3 is the eigenvalue of the second covariance matrix H, and diag is the square decomposition of the second covariance matrix H;

[0123] In step E3, the obtained eigenvalues are subjected to linear index calculation to obtain the building violation points, wherein the linear index calculation is shown in formula (11):

[0124]

[0125] Among them, Linearity is the linear index of the violation point in the point cloud scene.

[0126] Step F, based on the cluster segmentation results, the violation points in the point cloud scene and the building violation points, the building outline is identified and the building point cloud is obtained. Specifically, the cluster outer contour points in the cluster segmentation results are compared with the number of violation points in the point cloud scene belonging to the cluster. If the ratio of the number of violation points in the point cloud scene belonging to the cluster meets the set threshold, the number of building violation points is compared with the number of violation points in the point cloud scene. If the ratio of the number of building violation points meets the set threshold, the two-dimensional bounding box area of the cluster points in the cluster segmentation results is calculated to obtain the diagonal angle, and then the distance to the farthest point in the cluster is obtained. If the sine value of the diagonal angle is less than the set value, the cluster belongs to the building point cloud. The two-dimensional bounding box area is calculated using the built-in area calculation function polyarea in MATLAB as shown in formula (12).

[0127]

[0128] Where S is the area of the geometric figure, d is the side length, and β is the angle between the two sides.

[0129] In step G, the building point cloud is removed from the three-dimensional point cloud data and the remaining point cloud data is obtained. The remaining point cloud data is filtered using the statistical filtering method and a plane interpolation model is established. Then, the target scene point cloud is interpolated according to the plane interpolation model to obtain the grid plane fitting reference point cloud. The coordinates of each point in the plane interpolation model are shown in formula (13).

[0130] ax+by+cz+d=0(13)

[0131] Among them, a, b and c are all fitting plane parameters, and the fitting plane parameter constraints are a 2 +b2 +c 2 =1.

[0132] Step H: Use the slope angle grid to filter the grid plane fitting reference point cloud and obtain the filtered three-dimensional point cloud data. The specific steps are as follows:

[0133] like Figure 5 and Figure 6 As shown in step H1, the nine divided grids are taken as a whole to solve the slope angle of the adjacent grids of the target grid, as shown in formula (14),

[0134]

[0135] Among them, φ i is the slope angle of adjacent grids in the nine-square grid, (x i 、y i 、z i ) is the three-dimensional coordinate of the lowest point of the adjacent grid, (x ⑤ ,y ⑤ , z ⑤ ) is the three-dimensional coordinate of the lowest point to be determined in the target grid;

[0136] like Figure 7 As shown, in step H2, the surrounding grids of the target grid in the nine-square grid are integrated, and then the weighted slope angle corresponding to the target grid is obtained by distance weighting, as shown in formula (15),

[0137]

[0138] Among them, D i The Euclidean distance between the lowest point of the adjacent grid and the undetermined lowest point of the target grid, D is the weighted Euclidean distance, and γ is the slope angle of the target grid;

[0139] like Figure 8 As shown, in step H3, the slope angle threshold is set to calculate the first elevation threshold. The specific calculation of the first elevation threshold is shown in formula (16):

[0140]

[0141] Among them, h t is the first elevation threshold, α is the slope threshold corresponding to the slope angle, and L is the grid division length;

[0142] In step H4, the elevation plane equation is established using the elevation threshold and formula (13), and then the elevation plane equation is solved to obtain the second elevation threshold, as shown in formula (17).

[0143]

[0144] Among them, z t is the second elevation threshold;

[0145] Step H5: comparing the first elevation threshold and the second elevation threshold. If the second elevation threshold is lower than the first elevation threshold, the target grid's undetermined lowest point is a ground point. If the second elevation threshold is not lower than the first elevation threshold, the target grid's undetermined lowest point is a non-ground point.

[0146] like Figure 9 As shown, in step I, the edge contour points in the filtered three-dimensional point cloud data are checked and the three-dimensional point cloud data filtering operation is completed. Specifically, the violation points in the point cloud scene in step D and the cluster edge points in step C are re-taken as a whole point cloud set and the building point cloud in step F is removed therefrom, and the remaining point clouds are re-attributed to the corresponding clusters in step C. Then, the number of ground points obtained by step H within the set distance range of each point in the remaining point cloud is retrieved point by point. If the number of ground points within the set distance range meets the set threshold, it is further determined whether there is a point within the retrieval distance range whose height difference with the target grid's pending lowest point does not exceed the set threshold. If there is a point whose height difference with the target grid's pending lowest point does not exceed the set threshold, the target grid's pending lowest point is classified as a ground point.

[0147] A three-dimensional point cloud data filtering system based on adaptive clustering segmentation and gradient compensation includes a preprocessing module, a feature extraction module, an adaptive clustering segmentation module, a point cloud scene violation point retrieval module, a building violation point retrieval module, a building outline recognition module, a point cloud interpolation module, a point cloud filtering module and an edge contour point verification module. The preprocessing module is used to use a statistical filtering method to perform data preprocessing on the three-dimensional point cloud data and obtain preprocessed point cloud data; the feature extraction module is used to extract point cloud features in the preprocessed point cloud data and obtain point cloud feature data; the adaptive clustering segmentation module is used to perform adaptive clustering segmentation on the point cloud feature data and obtain clustering segmentation results; the point cloud scene violation point retrieval module is used to use a Delaunay triangulation to determine the spatial relationship between each point in the clustering segmentation result to obtain the point cloud scene violation points; The building violation point retrieval module is used to perform linear exponential solution to extract the violation points in the point cloud scene and obtain the building violation points; the building outline recognition module is used to identify the building outline based on the cluster segmentation results, the violation points in the point cloud scene and the building violation points and obtain the building point cloud; the point cloud interpolation module is used to remove the building point cloud from the three-dimensional point cloud data and obtain the remaining point cloud data, and then use the statistical filtering method to filter the remaining point cloud data and establish a plane interpolation model, and then perform point cloud interpolation on the target scene point cloud according to the plane interpolation model and obtain the grid plane fitting reference point cloud; the point cloud filtering module is used to use the slope angle grid to perform point cloud filtering on the grid plane fitting reference point cloud and obtain the filtered three-dimensional point cloud data; the edge contour point verification module is used to verify the edge contour points in the filtered three-dimensional point cloud data and complete the three-dimensional point cloud data filtering operation.

[0148] To sum up, the present invention provides a three-dimensional point cloud data filtering method and system based on adaptive clustering segmentation and gradient compensation, firstly, a statistical filtering method is used to preprocess the three-dimensional point cloud data and obtain preprocessed point cloud data, then the point cloud features in the preprocessed point cloud data are extracted and point cloud feature data are obtained, then the point cloud feature data are adaptively clustered and segmented to obtain clustering segmentation results, then the Delaunay triangulation is used to determine the spatial relationship of each point in the clustering segmentation results to obtain the point cloud scene violation points, then the linear exponential solution is performed on the point cloud scene to extract and obtain the building violation points, then based on the clustering segmentation results, the point cloud scene violation points and the building violation points, the building outline is identified and the building point cloud is obtained, then the building point cloud is removed from the three-dimensional point cloud data and the remaining point cloud data is obtained, then the statistical filtering method is used to filter the remaining point cloud data and a plane interpolation model is established, then the target scene point cloud is interpolated according to the plane interpolation model and a grid plane approximation is obtained. The 3D point cloud filtering method and system are effectively realized. The adaptive clustering segmentation algorithm based on density spatial distribution is used to establish multi-dimensional features including point cloud density, spatial autocorrelation and local curvature characteristics, and to dynamically cluster and segment the ground object-ground clusters. The linear feature parameters are introduced to separate the building cluster points, and the coarse-fine granularity grid hierarchical processing strategy is used to further separate the residual mixed point cloud, thereby separating the ground point cloud. It not only breaks through the technical limitations of traditional parameters with strong dependence on parameters, insufficient terrain adaptability and poor real-time performance, but also significantly improves the filtering accuracy and robustness in complex scenes through multi-dimensional collaborative optimization. This provides an efficient solution for the generation of high-precision digital elevation models with both theoretical innovation and engineering practical value.

[0149] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A 3D point cloud data filtering method based on adaptive clustering segmentation and gradient compensation, characterized by: The following steps are included: Step A, using a statistical filtering method to preprocess the three-dimensional point cloud data and obtain preprocessed point cloud data; Step B, extracting point cloud features from the pre-processed point cloud data and obtaining point cloud feature data; Step C, performing adaptive clustering segmentation on the point cloud feature data and obtaining the clustering segmentation result; Step D: Delaunay triangulation is used to determine the spatial relationship of each point in the cluster segmentation result, thereby obtaining the violation points in the point cloud scene; Step E, performing linear exponential solution to extract the violation points in the point cloud scene and obtaining the building violation points; Step F, identifying the building outline based on the cluster segmentation results, the violation points in the point cloud scene, and the violation points of the building to obtain the building point cloud; Step G: Remove the building point cloud from the 3D point cloud data to obtain the remaining point cloud data, then use the statistical filtering method to filter the remaining point cloud data and establish a plane interpolation model. Then, interpolate the target scene point cloud according to the plane interpolation model to obtain the grid plane fitting reference point cloud. Step H, using the slope angle grid to perform point cloud filtering on the grid plane fitting reference point cloud and obtain filtered three-dimensional point cloud data; Step I: Check the edge contour points in the filtered three-dimensional point cloud data and complete the three-dimensional point cloud data filtering operation.

2. The three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation according to claim 1, characterized in that: Step A: Preprocess the 3D point cloud data using a statistical filtering method to obtain preprocessed point cloud data. Specifically, the statistical filtering method calculates the average distance μ between each point in the 3D point cloud data and its neighboring points, and then obtains the distance threshold d from the average distance μ. threshold , as shown in formula (1), d threshold =μ+k·σ(1) Where k is a constant factor and σ is the standard deviation of the distance between each point and its neighboring points.

3. The three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation according to claim 2, characterized in that: Step B, extracting point cloud features from the pre-processed point cloud data and obtaining point cloud feature data, wherein the point cloud features include point cloud scene average density features, point cloud scene normal vector features, and point cloud scene neighboring point spatial distance features. The specific steps are as follows: Step B1: extract the average density feature of the point cloud scene. The average density feature of the point cloud scene is calculated as shown in formula (2): AD=SumD / NP(2) Among them, AD is the average density of the preprocessed point cloud data, SumD is the sum of the distances between each point and the nearest point in the preprocessed point cloud data, and NP is the number of point clouds in the preprocessed point cloud data; Step B2: extract the normal vector features of the point cloud scene, where the normal vector features include the first covariance matrix, eigenvalues, eigenvectors, and normal vector angles. The specific steps are as follows: Step B21, calculate the first covariance matrix, specifically, establish a KD-Tree and find the corresponding m adjacent points, then construct an m*3 matrix through the m point coordinates and calculate the first covariance matrix of the m*3 matrix, as shown in formula (3), Where Cov(X) is the first covariance matrix of matrix X, μ x is the average value of each dimension, and T is the matrix transpose; Step B22, calculate the eigenvalue and eigenvector, as shown in formula (4), Among them, λ is the eigenvalue and x is the eigenvector; Step B23, calculate the normal vector angle, as shown in formula (5), Among them, a1, b1, c1, a2, b2 and c2 are the components of the normal vector of each point, and θ is the angle between the normal vectors of two points; Step B3: extract the spatial distance features of neighboring points in the point cloud scene, as shown in formula (6): Among them, Dis is the distance between two points in space, (x1, y1, z1) and (x2, y2, z2) are the three-dimensional coordinates of the target point.

4. The method for filtering three-dimensional point cloud data based on adaptive clustering segmentation and gradient compensation according to claim 3, characterized in that: In step C, adaptive cluster segmentation is performed on the point cloud feature data to obtain the cluster segmentation result, wherein the clustering threshold of the adaptive cluster segmentation can be changed with the spatial characteristics of the point cloud, and the clustering threshold calculation process is shown in formula (7), Among them, YC is the clustering threshold of two points, and Δh is the height difference between two points.

5. The three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation according to claim 4, characterized in that: Step D: Use the Delaunay triangulation to determine the spatial relationship of each point in the cluster segmentation result, so as to obtain the violation points in the point cloud scene. The specific steps are as follows: Step D1, using the Delaunay triangulation to determine the spatial relationship of each point in the cluster segmentation result, specifically, calculating the normal vectors of adjacent triangles in the Delaunay triangulation, as shown in formula (8), in, and is the edge vector formed by two points in the triangle, (x a ,y a , z a )、(x b ,y b , z b ) and (x c ,y c , z c ) are the coordinates of the three vertices of the triangle, is the triangle normal vector; Step D2: Obtain the violation points in the point cloud scene, specifically using two triangular normal vectors Determine whether it is a violation triangle. If the two triangle normal vectors If the angle between the highest point and the lowest point of the violating triangle exceeds the set angle, it is a violating triangle. If the height difference between the highest point and the lowest point of the violating triangle meets the set threshold, the highest point of the violating triangle is taken as the violating point.

6. The three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation according to claim 5, characterized in that: Step E: extract the violation points in the point cloud scene by linear exponential solution and obtain the violation points of the building. Specifically, the violation points in the obtained point cloud scene are extracted and sorted into a point cloud set and each point in the point cloud set is judged point by point. The specific steps are as follows: Step E1: extract the K nearest neighboring points around the target point to construct a local neighborhood point set and translate the coordinate system so that the center of mass of the point cloud set is the origin, and then calculate the second covariance matrix as shown in formula (9): Among them, H is the second covariance matrix, n is the number of neighborhood points, x j is the three-dimensional coordinate of the j-th point, is the centroid coordinate of the neighborhood point; Step E2, decomposing the eigenvalues of the second covariance matrix H, as shown in formula (10), Where V is an orthogonal matrix, Λ is a diagonal matrix, λ1 and λ2 are the largest eigenvalue and the second largest eigenvalue of the covariance matrix H, respectively, λ3 is the eigenvalue of the second covariance matrix H, and diag is the square decomposition of the second covariance matrix H; In step E3, the obtained eigenvalues are subjected to linear index calculation to obtain the building violation points, wherein the linear index calculation is shown in formula (11): Among them, Linearity is the linear index of the violation point in the point cloud scene.

7. The three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation according to claim 6, characterized in that: Step F, based on the cluster segmentation results, the violation points in the point cloud scene and the building violation points, the building outline is identified and the building point cloud is obtained. Specifically, the cluster outer contour points in the cluster segmentation results are compared with the number of violation points in the point cloud scene belonging to the cluster. If the ratio of the number of violation points in the point cloud scene belonging to the cluster meets the set threshold, the number of building violation points is compared with the number of violation points in the point cloud scene. If the ratio of the number of building violation points meets the set threshold, the two-dimensional bounding box area of the cluster points in the cluster segmentation results is calculated to obtain the diagonal angle, and then the distance to the farthest point in the cluster is obtained. If the sine value of the diagonal angle is less than the set value, the cluster belongs to the building point cloud. The two-dimensional bounding box area is calculated using the built-in area calculation function polyarea in MATLAB as shown in formula (12). Where S is the area of the geometric figure, d is the side length, and β is the angle between the two sides.

8. The three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation according to claim 7, characterized in that: In step G, the building point cloud is removed from the three-dimensional point cloud data and the remaining point cloud data is obtained. The remaining point cloud data is filtered using the statistical filtering method and a plane interpolation model is established. Then, the target scene point cloud is interpolated according to the plane interpolation model to obtain the grid plane fitting reference point cloud. The coordinates of each point in the plane interpolation model are shown in formula (13). ax+by+cz+d=0 (13) Among them, a, b and c are all fitting plane parameters, and the fitting plane parameter constraints are a 2 +b 2 +c 2 =1.

9. The three-dimensional point cloud data filtering method based on adaptive clustering segmentation and gradient compensation according to claim 8, characterized in that: Step H: Use the slope angle grid to filter the grid plane fitting reference point cloud and obtain the filtered three-dimensional point cloud data. The specific steps are as follows: Step H1: Take the nine divided grids as a whole to solve the slope angle of the adjacent grids of the target grid, as shown in formula (14): Among them, φ i is the slope angle of adjacent grids in the nine-square grid, (x i 、y i 、z i ) is the three-dimensional coordinate of the lowest point of the adjacent grid, (x ⑤ ,y ⑤ , z ⑤ ) is the three-dimensional coordinate of the lowest point to be determined in the target grid; Step H2: Integrate the surrounding grids of the target grid in the nine-square grid, and then obtain the weighted slope angle corresponding to the target grid by distance weighting, as shown in formula (15): Among them, D i The Euclidean distance between the lowest point of the adjacent grid and the undetermined lowest point of the target grid, D is the weighted Euclidean distance, and γ is the slope angle of the target grid; Step H3: Set the slope angle threshold to calculate the first elevation threshold. The specific calculation of the first elevation threshold is shown in formula (16): Among them, h t is the first elevation threshold, α is the slope threshold corresponding to the slope angle, and L is the grid division length; In step H4, the elevation plane equation is established using the elevation threshold and formula (13), and then the elevation plane equation is solved to obtain the second elevation threshold, as shown in formula (17). Among them, z t is the second elevation threshold; Step H5: comparing the first elevation threshold and the second elevation threshold. If the second elevation threshold is lower than the first elevation threshold, the target grid's undetermined lowest point is a ground point. If the second elevation threshold is not lower than the first elevation threshold, the target grid's undetermined lowest point is a non-ground point. Step I, checks the edge contour points in the filtered three-dimensional point cloud data and completes the three-dimensional point cloud data filtering operation. Specifically, the violation points in the point cloud scene in step D and the cluster edge points in step C are re-taken as a whole point cloud set and the building point cloud in step F is removed therefrom. The remaining point cloud is then re-assigned to the corresponding cluster in step C. Then, the number of ground points obtained by step H within the set distance range of each point in the remaining point cloud is retrieved point by point. If the number of ground points within the set distance range meets the set threshold, it is further determined whether there is a point within the retrieval distance range whose height difference with the target grid's pending lowest point does not exceed the set threshold. If there is a point whose height difference with the target grid's pending lowest point does not exceed the set threshold, the target grid's pending lowest point is classified as a ground point.

10. A 3D point cloud data filtering system based on adaptive clustering segmentation and gradient compensation, wherein the specific filtering process of the 3D point cloud data filtering system is based on the 3D point cloud data filtering method according to any one of claims 1 to 9, characterized in that: It includes a preprocessing module, a feature extraction module, an adaptive clustering segmentation module, a point cloud scene violation point retrieval module, a building violation point retrieval module, a building outline recognition module, a point cloud interpolation module, a point cloud filtering module and an edge contour point verification module. The preprocessing module is used to use a statistical filtering method to preprocess the three-dimensional point cloud data and obtain preprocessed point cloud data; The feature extraction module is used to extract point cloud features from the pre-processed point cloud data and obtain point cloud feature data; The adaptive clustering segmentation module is used to perform adaptive clustering segmentation on point cloud feature data and obtain clustering segmentation results; The point cloud scene violation point retrieval module is used to determine the spatial relationship of each point in the cluster segmentation result using the Delaunay triangulation to obtain the violation points in the point cloud scene; The building violation point retrieval module is used to extract the violation points in the point cloud scene by linear exponential solution and obtain the building violation points; The building outline recognition module is used to identify the building outline and obtain the building point cloud based on the cluster segmentation results, the violation points in the point cloud scene and the violation points of the building; The point cloud interpolation module is used to remove the building point cloud from the three-dimensional point cloud data and obtain the remaining point cloud data, then use the statistical filtering method to filter the remaining point cloud data and establish a plane interpolation model, and then interpolate the target scene point cloud according to the plane interpolation model to obtain a grid plane fitting reference point cloud; The point cloud filtering module is used to perform point cloud filtering on the grid plane fitting reference point cloud using the slope angle grid and obtain filtered three-dimensional point cloud data; The edge contour point checking module is used to check the edge contour points in the filtered three-dimensional point cloud data and complete the three-dimensional point cloud data filtering operation.

Images