Non-uniform point cloud surface segmentation method based on improved region growing method

Abstract

The present application relates to a method for segmenting non-uniform point cloud domains based on an improved region growing method. The method includes: removing outlier data points of a three-dimensional laser point cloud to be segmented to obtain the main body of the point cloud; downsampling the main body of the point cloud, and performing rough segmentation to define the spatial coordinate range and area estimation of each domain; determining the grid type to which each grid of each domain belongs based on the eigenvalue of the covariance matrix of the k-neighborhood points; forming a mean point data set based on the points in all mean point grids of each domain, and determining the initial seed points corresponding to each domain; formulating growth rules based on curvature and normal vector information to determine the gross density of growing internal points of each domain; stopping the extraction of internal points of the domain for domains greater than or equal to the density threshold to obtain the internal points of the domain; performing consistency checks on domains less than the density threshold, extracting the non-growing internal points in the non-uniform point grid of the domain to form the internal points of the domain, and realizing the extraction of internal points of the domain with poor uniformity and continuity.

Description

Technical Field

[0001] The present application relates to the technical field of three-dimensional laser point cloud plane segmentation, and in particular to a non-uniform point cloud surface segmentation method based on an improved region growing method. Background Art

[0002] Laser 3D scanning technology is a non-contact intelligent sensing technology of machine vision. It can measure the outer contour of a 3D object with millimeter-level accuracy and generate high-precision spatial coordinates of a large number of surface points (also known as "point clouds") of the measured object to describe its refined geometric features. The use of this technology in the construction industry can quickly obtain data related to the overall and local geometric forms and spatial lines of components, assist workers in geometric shape verification and virtual construction of key components, and realize related digital modeling and archiving.

[0003] At present, the research on point cloud domain segmentation algorithms has made many progresses. Among them, the region growing method based on surface features has been widely used due to its high efficiency and low memory consumption. However, its disadvantages are: there is a lack of unified optimal criteria for the selection of initial seed points. Correct selection of seed points can ensure accurate and effective domain segmentation. At present, the point with the smallest curvature is generally selected as the seed point, but the segmentation effect is poor when there are many target planes. The algorithm itself has high requirements on the uniformity and continuity of point cloud density. In reality, due to the complex structure of the measured object, the large number of accessories that cause mutual occlusion, and the limited layout of the measuring station due to the on-site scanning environment conditions, the point cloud often appears uneven and discontinuous. The applicability of the traditional region growing method is weakened, and it is impossible to extract points in the domain with poor uniformity and continuity. Summary of the invention

[0004] Based on this, it is necessary to provide a non-uniform point cloud surface segmentation method based on an improved region growing method that can extract points within a surface with poor uniformity and continuity in order to address the above technical problems.

[0005] A non-uniform point cloud surface segmentation method based on an improved region growing method, the method comprising:

[0006] A filtering algorithm is used to remove outlier data points of the three-dimensional laser point cloud to be segmented to obtain the point cloud body;

[0007] Downsampling the point cloud body using a three-dimensional cube grid filter to obtain a downsampled point cloud body;

[0008] Roughly segmenting the face regions of the downsampled point cloud body, roughly defining the spatial coordinate range of each face region, estimating the area of ​​each face region, and numbering each face region;

[0009] Re-dividing each of the face areas into corresponding three-dimensional cubic grids, judging the uniformity of point cloud distribution in each grid of each of the face areas based on the eigenvalues ​​of the covariance matrix of the k-neighborhood points, and determining the grid type to which each grid of each of the face areas belongs, wherein the grid type includes uniform point grid and non-uniform point grid;

[0010] Constructing a mean point data set corresponding to each of the face regions according to the points in all mean point grids of each of the face regions, calculating the centroid point corresponding to each of the mean point data sets, and determining the point with the smallest Euclidean distance from the mean point data set of each of the face regions to the centroid point as the initial seed point corresponding to each of the face regions;

[0011] Formulate growth rules based on curvature and normal vector information, extract the growing inner points of each face domain according to the initial seed points corresponding to each face domain, calculate the ratio of the number of growing inner points of each face domain to the area of ​​the corresponding face domain, and determine the gross density of the growing inner points of each face domain;

[0012] Determine whether the gross density of the growing inner points of each of the face regions is greater than or equal to a density threshold, stop extracting face inner points for the face regions that are greater than or equal to the density threshold, and the growing inner points of the face regions are the face inner points obtained for the face regions;

[0013] The growing internal points corresponding to the face domain that is less than the density threshold are used as the target face domain internal points, and a consistency check is performed to extract the non-growing internal points in the non-uniform point grid of the face domain. The non-growing internal points and the growing internal points together constitute the face domain internal points of the face domain.

[0014] In one embodiment, the step of removing outlier data points of the three-dimensional laser point cloud to be segmented by using a filtering algorithm to obtain the point cloud body includes:

[0015] Treat each point p in the segmented 3D laser point cloud as i Find its k1 neighboring points, fit a local plane with the neighboring points, and calculate p i The distance to its local plane is calculated, and points whose distance exceeds a certain distance threshold are removed to obtain the point cloud body.

[0016] In one embodiment, the step of downsampling the point cloud body using a three-dimensional cube grid filter to obtain a downsampled point cloud body comprises:

[0017] A three-dimensional cube grid of the point cloud body is established, and points in the same cube are merged into a single point to obtain a downsampled point cloud body.

[0018] In one of the embodiments, the steps of re-dividing each of the face areas into corresponding three-dimensional cubic grids, judging the uniformity of point cloud distribution in each grid of each of the face areas based on the eigenvalues ​​of the k-neighborhood point covariance matrix, and determining the grid type to which each grid of each of the face areas belongs, wherein the grid type includes a uniform point grid and a non-uniform point grid, include:

[0019] Re-divide each of the face regions into corresponding three-dimensional cubic grids, and for each point of the downsampled point cloud body, take its k2 neighboring points to form a neighborhood point set, and determine the neighborhood point covariance matrix C of each point, which is expressed as:

[0020]

[0021] Where X, Y, and Z are vectors with k2 rows and 1 column, representing the three-dimensional rectangular coordinates of k2 neighborhood points, respectively. cov(X,Y) and cov(Y,X) represent the covariance of two vectors X and Y, respectively. cov(X,Z) and cov(Z,X) represent the covariance of two vectors X and Z, respectively. cov(Y,Z) and cov(Z,Y) represent the covariance of two vectors Z and Y, respectively.

[0022] Calculate the eigenvalue of the covariance matrix of the neighborhood points of each point, determine the maximum eigenvalue of the covariance matrix corresponding to each point, and determine the characteristic threshold value according to the calculation result of the maximum eigenvalue of each point;

[0023] The maximum eigenvalue corresponding to each point in each grid of the three-dimensional cube grid of each face area is counted, and the ratio of the number of points exceeding the characteristic threshold to the total number of points in the grid is calculated. The grid exceeding the ratio threshold is a non-uniform point grid, and the grid not exceeding the ratio threshold is an even point grid.

[0024] In one embodiment, the step of formulating growth rules based on curvature and normal vector information, extracting the growing inner points of each of the face regions according to the initial seed points corresponding to each of the face regions, calculating the ratio of the number of growing inner points of each of the face regions to the area of ​​the corresponding face region, and determining the gross density of the growing inner points of each of the face regions includes:

[0025] Calculating the normal vector and curvature of each point of each of the face regions using principal component analysis;

[0026] The initial seed points corresponding to the face regions are used as the initial seed points in the set of seed points to be searched for the face regions;

[0027] Sequentially use various sub-points in the seed point set to be searched as current search seed points, and search k3 neighboring points of the current search seed point;

[0028] Determine the neighboring points whose normal vector angles with the current search seed point are less than the angle threshold as points within the target surface area;

[0029] According to the curvature threshold, determine whether the curvature of each of the neighborhood points is less than the curvature threshold, add the neighborhood points less than the curvature threshold to the set of seed points to be searched, and delete the current search seed point from the set of seed points to be searched;

[0030] Return to the step of sequentially using various sub-points in the seed point set to be searched as the current search seed point and searching k3 neighborhood points of the current search seed point until the seed point set to be searched is an empty set.

[0031] In one embodiment, the density threshold is determined according to the edge length of the three-dimensional cube grid, and the density threshold calculation formula is:

[0032] Density_th=a*1 / (Edge1^2)

[0033] Among them, Density_th represents the density threshold, a represents the density reduction coefficient, and Edge1 is the edge length of the three-dimensional cube grid used for downsampling.

[0034] In one embodiment, the step of taking the growing interior points corresponding to the face area less than the density threshold as the target face area interior points, performing consistency check, extracting the non-growing interior points in the non-uniform point grid of the face area, and forming the face area interior points of the face area with the non-growing interior points and the growing interior points includes:

[0035] The growing inner points corresponding to the face domains less than the density threshold are taken as the target face domain inner points, and added to the inner point set, and the fitting plane equation is obtained by using the least square method for the points in the inner point set;

[0036] Extract all points in the non-uniform point grid in the surface space to form a candidate point data set;

[0037] Pop up the points in the candidate point data set one by one for consistency check;

[0038] When the distance between the currently popped-up point and the fitting plane is less than the distance threshold, and the angle between the normal vector of the currently popped-up point and the normal vector of the fitting plane is less than the angle threshold, the currently popped-up point is a non-growing inner point and is added to the inner point set;

[0039] Repeat the step of popping out points in the candidate point data set in sequence for consistency check until all points in the candidate point data set are popped out, and the points in the final internal point set are composed of the non-growing internal points and the growing internal points together forming the internal points of the face domain.

[0040] The above-mentioned non-uniform point cloud domain segmentation method based on the improved region growing method adopts a filtering algorithm to remove outlier data points of the three-dimensional laser point cloud to be segmented to obtain the point cloud body; adopts a three-dimensional cube grid filter to downsample the point cloud body to obtain the downsampled point cloud body; roughly segment the domain of the downsampled point cloud body, roughly define the spatial coordinate range of each domain, estimate the area of ​​each domain, and number each domain; re-divide each domain into a corresponding three-dimensional cube grid, judge the uniformity of the point cloud distribution in each grid of each domain based on the eigenvalue of the covariance matrix of the k-neighborhood points, and determine the grid type to which each grid of each domain belongs, and the grid type includes uniform point grid and non-uniform point grid; construct the uniform point data set corresponding to each domain according to the points in all the uniform point grids of each domain, and concentrate the uniform point data of each domain into a mass The point with the smallest Euclidean distance to the center point is determined as the initial seed point corresponding to each face domain; growth rules are formulated based on curvature and normal vector information, and the growing inner points of each face domain are extracted according to the initial seed points corresponding to each face domain, and the ratio of the number of growing inner points of each face domain to the corresponding face domain area is calculated to determine the gross density of the growing inner points of each face domain; it is judged whether the gross density of the growing inner points of each face domain is greater than or equal to the density threshold, and the face domain inner point extraction for the face domain greater than or equal to the density threshold is stopped, and the growing inner point of the face domain is the face domain inner point of the face domain; the growing inner point corresponding to the face domain less than the density threshold is used as the target face domain inner point, and a consistency check is performed to extract the non-growing inner points in the non-uniform point grid of the face domain, and the face domain inner points of the face domain are composed of the non-growing inner points and the growing inner points, so as to realize the extraction of face domain inner points with poor uniformity and continuity. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 A schematic diagram of a flow chart of a non-uniform point cloud surface segmentation method based on an improved region growing method in one embodiment;

[0042] Figure 2 It is a flowchart of a non-uniform point cloud surface segmentation method based on an improved region growing method in another embodiment;

[0043] Figure 3 A schematic diagram of a surface area of ​​a precast concrete box beam casting template in one embodiment;

[0044] Figure 4 The division result of the uniform point grid and the non-uniform point grid of the No. 1 surface area in an embodiment;

[0045] Figure 5 The division result of the uniform point grid and the non-uniform point grid of the surface area 2-3 in one embodiment;

[0046] Figure 6 The surface segmentation results of the left web and the right web in one embodiment;

[0047] Figure 7 The bottom plate surface segmentation result in one embodiment;

[0048] Figure 8 This is a diagram showing the overall effect of segmenting three face regions in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0050] In one embodiment, Figure 1 As shown, a non-uniform point cloud surface segmentation method based on an improved region growing method is provided, comprising the following steps:

[0051] Step S220: Using a filtering algorithm to remove outlier data points of the three-dimensional laser point cloud to be segmented, to obtain a point cloud body.

[0052] The three-dimensional laser point cloud to be segmented may be a three-dimensional laser point cloud obtained by manually removing block noise in the point cloud using professional point cloud processing software, such as removing noise such as debris on the scanned site.

[0053] In one embodiment, a step of using a filtering algorithm to remove outlier data points of the three-dimensional laser point cloud to be segmented and obtaining the point cloud body includes: for each point p in the three-dimensional laser point cloud to be segmented i Find its k1 neighboring points, fit a local plane with the neighboring points, and calculate p i The distance to its local plane is calculated, and the points whose distance exceeds a certain distance threshold are removed to obtain the point cloud body.

[0054] Wherein, k1 is a positive integer, and the value is selected according to the actual situation, such as k1 is 8, 15, 20, etc. The distance threshold is set according to the actual situation, such as 0.1m, 0.2m, 0.01m, etc.

[0055] Step S240: downsample the point cloud body using a three-dimensional cube grid filter to obtain a downsampled point cloud body.

[0056] In one embodiment, a three-dimensional cube grid filter is used to downsample a point cloud body to obtain a downsampled point cloud body, including: establishing a three-dimensional cube grid of the point cloud body, merging points in the same cube into a single point, and obtaining a downsampled point cloud body.

[0057] Among them, by establishing a three-dimensional cube grid of the point cloud body, the points in the same cube are merged into a single point to achieve the purpose of downsampling, better retain the original boundary of the point cloud, and remove the noise above and below the plane, which can be regarded as a further smoothing and filtering method.

[0058] Step S260 , roughly segment the face regions of the downsampled point cloud body, roughly define the spatial coordinate range of each face region, estimate the area of ​​each face region, and number each face region.

[0059] Step S280, re-divide each face area into a corresponding three-dimensional cube grid, judge the uniformity of the point cloud distribution in each grid of each face area based on the eigenvalue of the covariance matrix of the k-neighborhood points, and determine the grid type to which each grid of each face area belongs, and the grid types include uniform point grid and non-uniform point grid.

[0060] In one embodiment, each face region is re-divided into a corresponding three-dimensional cubic grid, and the uniformity of the point cloud distribution in each grid of each face region is judged based on the eigenvalue of the covariance matrix of the k-neighborhood points, and the grid type to which each grid of each face region belongs is determined, and the grid type includes a uniform point grid and a non-uniform point grid, including:

[0061] Re-divide each face area into the corresponding three-dimensional cube grid, and for each point of the downsampled point cloud body, take its k2 neighboring points to form a neighborhood point set, and determine the neighborhood point covariance matrix C of each point, which is expressed as:

[0062]

[0063] Where X, Y, and Z are vectors with k2 rows and 1 column, representing the three-dimensional rectangular coordinates of k2 neighborhood points, respectively. cov(X,Y) and cov(Y,X) represent the covariance of two vectors X and Y, respectively. cov(X,Z) and cov(Z,X) represent the covariance of two vectors X and Z, respectively. cov(Y,Z) and cov(Z,Y) represent the covariance of two vectors Z and Y, respectively.

[0064] The eigenvalue of the covariance matrix of the neighborhood points of each point is calculated, the maximum eigenvalue of the covariance matrix corresponding to each point is determined, and the characteristic threshold is determined according to the calculation result of the maximum eigenvalue of each point; the maximum eigenvalue corresponding to each point in each grid of the three-dimensional cube grid of each face domain is counted, and the proportion of the number of points exceeding the characteristic threshold to the total number of points in the grid. The grid exceeding the proportion threshold is a non-uniform point grid, and the grid not exceeding the proportion threshold is an even point grid.

[0065] Among them, a point exceeding the characteristic threshold indicates that the neighborhood points of the point have obvious dispersion in the direction of the eigenvector corresponding to the maximum eigenvalue, and it is considered that the distribution of the point in each direction within its neighborhood is relatively uneven.

[0066] Step S300, construct a mean point data set corresponding to each face domain based on the points in all the mean point grids of each face domain, calculate the centroid point of each mean point data set, and determine the point with the smallest Euclidean distance from the mean point data set of each face domain to the centroid point as the initial seed point corresponding to each face domain.

[0067] Step S320, formulate growth rules based on curvature and normal vector information, extract the growing internal points of each face domain according to the initial seed points corresponding to each face domain, calculate the ratio of the number of growing internal points of each face domain to the corresponding face domain area, and determine the gross density of the growing internal points of each face domain.

[0068] In one embodiment, the steps of formulating growth rules based on curvature and normal vector information, extracting the growing inner points of each face region according to the initial seed points corresponding to each face region, calculating the ratio of the number of growing inner points of each face region to the area of ​​the corresponding face region, and determining the gross density of the growing inner points of each face region include:

[0069] The principal component analysis method is used to calculate the normal vector and curvature of each point in each face domain; the initial seed point corresponding to each face domain is used as the initial seed point in the seed point set to be searched for each face domain; various sub-points in the seed point set to be searched are used as the current search seed point in turn, and the k3 neighboring points of the current search seed point are searched; the neighboring points whose angles with the normal vectors of the current search seed point are less than the angle threshold are determined as points in the target face domain; according to the curvature threshold, it is determined whether the curvature of each neighboring point is less than the curvature threshold, and the neighboring points less than the curvature threshold are added to the seed point set to be searched, and the current search seed point is deleted from the seed point set to be searched; return to execute the steps of using various sub-points in the seed point set to be searched as the current search seed point in turn, and searching the k3 neighboring points of the current search seed point, until the seed point set to be searched is an empty set.

[0070] Step S340, determine whether the gross density of the growing internal points of each face region is greater than or equal to the density threshold, stop extracting the face region internal points for the face region that is greater than or equal to the density threshold, and the growing internal points of the face region are the face region internal points of the face region.

[0071] In one embodiment, the density threshold is determined according to the edge length of the three-dimensional cube grid, and the density threshold calculation formula is:

[0072] Density_th=a*1 / (Edge1^2)

[0073] Among them, Density_th represents the density threshold, a represents the density reduction coefficient, and Edge1 is the edge length of the three-dimensional cube grid used for downsampling.

[0074] Step S360, taking the growing internal points corresponding to the face domain less than the density threshold as the target face domain internal points, performing consistency check, extracting the non-growing internal points in the non-uniform point grid of the face domain, and the non-growing internal points and the growing internal points together constitute the face domain internal points of the face domain.

[0075] In one embodiment, the step of taking the growing interior points corresponding to the face domains less than the density threshold as the target face domain interior points, performing consistency check, extracting the non-growing interior points in the non-uniform point grid of the face domain, and forming the face domain interior points of the face domain by the non-growing interior points and the growing interior points includes:

[0076] The growing inner points corresponding to the face domain less than the density threshold are taken as the target face domain inner points and added to the inner point set corresponding to the face domain. The least square method is used to obtain the fitting plane equation for the points in the inner point set; all points in the non-uniform point grid in the face domain space are extracted to form a candidate point data set; the points in the candidate point data set are popped out in sequence for consistency check; when the distance between the current popped out point and the fitting plane is less than the distance threshold, and the angle between the normal vector of the current popped out point and the normal vector of the fitting plane is less than the angle threshold, the current popped out point is a non-growing inner point and is added to the inner point set; the steps of popping out the points in the candidate point data set in sequence for consistency check are repeated until all the points in the candidate point data set are popped out, and the points in the final inner point set are composed of non-growing inner points and growing inner points together to form the face domain inner points of the face domain.

[0077] In one embodiment, Figure 2 As shown, a non-uniform point cloud surface segmentation method based on the improved region growing method is specifically applied to, Figure 3 The 3D laser point cloud (i.e. the 3D laser point cloud to be segmented) of the webs and bottom plates on both sides of the precast concrete box beam casting formwork (hereinafter referred to as the precast formwork) shown in the figure is segmented for the surface area, as shown in the figure:

[0078] Step 1: for each point p in the prefabricated template point cloud P i Find its k1 neighboring points, fit a local plane with the neighboring points, and calculate p i If the distance to its local plane exceeds a certain distance threshold Dist_th_denoise, the point is defined as an outlier and removed to obtain a point cloud Q. The remaining steps are performed on the point cloud Q. In this embodiment, k1 is 8 and Dist_th_denoise is 0.1 m.

[0079] The precast formwork point cloud P is the three-dimensional laser point cloud of the webs and bottom plates on both sides of the precast concrete box girder casting formwork (hereinafter referred to as the precast formwork). The point cloud is pre-processed with professional point cloud processing software to remove block noise in the point cloud. The block noise in the point cloud includes the pedestal under the precast formwork and debris on the scanning site.

[0080] Step 2: Use a 3D cube grid filter to downsample the point cloud body to obtain the downsampled point cloud body.

[0081] Among them, a three-dimensional uniform cube grid is established for the point cloud Q, and the edge length Edge1 of the cube grid is set to 0.02m. Points in the same cube are merged into a single point to achieve the purpose of downsampling. This method better retains the original boundary of the point cloud and removes noise points above and below the plane, which can be regarded as a further smoothing and filtering method.

[0082] Step 3: Roughly define the spatial coordinate range of each face area of ​​the downsampled point cloud body and estimate its area A i , number each face area (i=1,2,3…).

[0083] In this embodiment, the center of the bottom surface of the prefabricated template is used as the coordinate origin, and the number, name, spatial range and area of ​​each area are listed in Table 1.

[0084] Table 1 Area delineation results

[0085]

[0086]

[0087] Step 4: Re-divide each roughly delineated area into several three-dimensional cubic grids, and judge the uniformity of the point cloud distribution in each grid based on the eigenvalue of the covariance matrix of the k-neighborhood points. Based on this, divide each grid into two categories: uniform point grid and non-uniform point grid;

[0088] Among them, for each point of the point cloud body Q, take its k2 neighboring points to form a neighborhood point set, and make the neighborhood point covariance matrix C:

[0089]

[0090] Where X, Y, and Z are all vectors with k2 rows and 1 column, representing the three-dimensional rectangular coordinates of k2 neighborhood points respectively. cov(X,Y) and cov(Y,X) represent the covariance of two vectors X and Y. cov(X,Z) and cov(Z,X) represent the covariance of two vectors X and Z. cov(Y,Z) and cov(Z,Y) represent the covariance of two vectors Z and Y.

[0091] Furthermore, the eigenvalue of the covariance matrix of the neighborhood points of each point is calculated, the maximum eigenvalue of each point is determined, and the characteristic threshold is determined according to the calculation result of the maximum eigenvalue. If the maximum eigenvalue of a point exceeds the characteristic threshold, it means that the neighborhood points of the point have obvious dispersion in the direction of the eigenvector corresponding to the maximum eigenvalue, and it is considered that the distribution of the point in each direction within its neighborhood is relatively uneven.

[0092] Furthermore, the three face domains are re-divided into several cubic grids, with the edge length designated as Edge2. The maximum eigenvalue corresponding to each point in each cubic grid is counted. If the ratio of the number of points exceeding the characteristic threshold to the total number of points in the grid exceeds a certain ratio threshold, the grid is defined as a non-uniform point grid, otherwise it is an even point grid. In this embodiment, Edge2 is 0.1m, k2 is 8, the characteristic threshold is 0.0004, and the ratio threshold is 80%. In this embodiment, only one measuring station is set for on-site scanning. The farther away from the laser scanner, the lower the point cloud density, and the worse the uniformity and continuity of the point cloud. Figure 4 and Figure 5 As shown, the division results of the uniform point grid and the non-uniform point grid of the surface areas 1-3 of this embodiment are shown.

[0093] Step 5: All points in the mean point grid of each area constitute the mean point data set D i , calculate the centroid point M of the mean point data set i , D i The point S with the smallest Euclidean distance to the centroid i Defined as the initial seed point of face region i.

[0094] Among them, the points in the mean point grids in the three surface domain spaces are extracted respectively to form three mean point data sets D i (i=1,2,3), calculate D i The center of mass M i (i=1,2,3) coordinates, find the mean point data set D i From the center to the center of mass M i The point S with the smallest Euclidean distance i (i=1, 2, 3) is defined as the initial seed point of the i-th face region, so as to ensure that the point cloud distribution near the initial seed point is uniform and continuous.

[0095] Step 6: Formulate growth rules based on curvature and normal vector information, extract points inside each face domain, called growth interior points, and calculate the number of growth interior points N i And the corresponding area A i The ratio of is called the growing inner point gross density Density_gro.

[0096] Among them, the principal component analysis method is used to calculate the normal vector and curvature of each point; the initial seed point corresponding to each face area is used as the initial seed point in the seed point set to be searched for each face area; various sub-points in the seed point set to be searched are used as the current search seed point in turn, and the k3 neighboring points of the current search seed point are searched. If the angle between the normal vector of the neighboring point and the seed point is less than the angle threshold, the neighboring point is considered to be a point in the target face area. If the curvature of the neighboring point is less than the curvature threshold, the neighboring point is added to the seed point set to be searched. After the k3 neighboring points are judged, the current seed point is deleted, and various sub-points in the seed point set to be searched are used as the current search seed point in turn, and the k3 neighboring points of the current search seed point are searched until the seed point set to be searched is an empty set. In this embodiment, k3 is 8, the curvature threshold is 0.05, and the angle threshold is 5°.

[0097] Furthermore, the number of growing internal points N of the three face regions is counted i (i=1,2,3), calculate the number of growing internal points N i and the area A i The ratio of is the growing inner point gross density Density_gro. The growing inner point gross density of the three face regions in this embodiment is calculated and listed in Table 2.

[0098] Table 2 Growth point hair density

[0099]

[0100] Step 7: If the gross density of the growing inliers is greater than the density threshold, it is considered that the extraction of the inliers in the target face domain is completed, and the growing inliers are the inliers in the target face domain.

[0101] Furthermore, in this embodiment, the density threshold is determined by the edge length Edge1 of the cube grid in step 2.

[0102] Density_th=a*1 / (Edge1^2)

[0103] In this embodiment, Edge1 = 0.02m. The distance between the scanner and the object to be measured will affect the point cloud density, and density reduction needs to be considered. According to the scanner accuracy and the on-site measurement distance, the density reduction coefficient a in this embodiment is taken as 0.85. Density_th = 2125.

[0104] The gross density of the growing inner points of the left and right ventral plate domains is greater than 2125, and the extraction of the inner points of the domains has been completed. The extraction of the inner points of the bottom plate domain, that is, domain No. 1, needs to be completed through further consistency testing. Figure 6 As shown, the left and right webs, i.e., the segmentation results of face regions 2 and 3, are shown in this embodiment.

[0105] Step 8: If the gross density of the growing inliers is less than the density threshold, further consistency check is required, and the obtained growing inliers are set as the target face domain inliers. Verify whether the points in the candidate data set have similar characteristics to the set target face domain inliers, i.e. consistency check. If the test conditions are met, they are considered to be the set target face domain inliers. Thus, the non-growing inliers in the face domain are extracted, and together with the growing inliers, they constitute the target face domain inliers.

[0106] The obtained growing internal points are set as the internal points of the target surface domain and added to the internal point set I1 corresponding to the surface domain No. 1. The least square method is used to obtain the fitting plane equation for the points in I1. All points in the non-uniform point grid in the bottom plate surface domain space are extracted to form the candidate point data set E1, and the points in E1 are popped out one by one for consistency check.

[0107] Furthermore, the consistency check conditions include: the distance between the current pop-up point and the fitting plane is less than a certain distance threshold, and the angle between the normal vector of the current pop-up point and the normal vector of the fitting plane is less than a certain angle threshold. If the current pop-up point meets the consistency check conditions, it will pass the consistency check and be considered as an inlier point in the target surface domain, that is, a non-growing inlier point, and will be added to the inlier point set I1. In this embodiment, the distance threshold is 0.01m and the angle threshold is 10°. Repeat the above steps until all points in the candidate point data set E1 are popped out, and the final inlier point set I1 is the inlier point in the target surface domain. Figure 7 As shown in FIG. 1 , the segmentation result of the bottom plate region, namely, region No. 1, is shown in FIG. Figure 8 As shown, it is a three-dimensional laser point cloud surface segmentation effect diagram of the web and bottom plates on both sides of the precast concrete box girder casting formwork (hereinafter referred to as precast formwork) of this embodiment.

[0108] It should be understood that although Figure 1 The steps in the flowchart are shown in sequence as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 At least part of the steps may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily executed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least part of the sub-steps or stages of other steps.

[0109] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0110] The above embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A non-uniform point cloud surface segmentation method based on an improved region growing method, characterized in that: The method comprises: A filtering algorithm is used to remove outlier data points of the three-dimensional laser point cloud to be segmented to obtain the point cloud body; Downsampling the point cloud body using a three-dimensional cube grid filter to obtain a downsampled point cloud body; Roughly segmenting the face regions of the downsampled point cloud body, roughly defining the spatial coordinate range of each face region, estimating the area of ​​each face region, and numbering each face region; Re-dividing each of the face areas into corresponding three-dimensional cubic grids, judging the uniformity of point cloud distribution in each grid of each of the face areas based on the eigenvalues ​​of the covariance matrix of the k-neighborhood points, and determining the grid type to which each grid of each of the face areas belongs, wherein the grid type includes uniform point grid and non-uniform point grid; According to the points in all the mean point grids of each face area, a mean point data set corresponding to each face area is formed, the centroid point of each mean point data set is calculated, and the point with the smallest Euclidean distance from the mean point data set of each face area to the centroid point is determined as the initial seed point corresponding to each face area; Formulate growth rules based on curvature and normal vector information, extract the growing inner points of each face domain according to the initial seed points corresponding to each face domain, calculate the ratio of the number of growing inner points of each face domain to the area of ​​the corresponding face domain, and determine the gross density of the growing inner points of each face domain; Determine whether the gross density of the growing inner points of each of the face regions is greater than or equal to a density threshold, stop extracting face region inner points for the face regions that are greater than or equal to the density threshold, and the growing inner points of the face regions are the face region inner points of the face regions; The growing internal points corresponding to the face domain that is less than the density threshold are used as the target face domain internal points, and a consistency check is performed to extract the non-growing internal points in the non-uniform point grid of the face domain. The non-growing internal points and the growing internal points together constitute the face domain internal points of the face domain.

2. The method according to claim 1, characterized in that The step of using a filtering algorithm to remove outlier data points of the three-dimensional laser point cloud to be segmented to obtain the point cloud body includes: Treat each point p in the segmented 3D laser point cloud as i Find its k1 neighboring points, fit a local plane with the neighboring points, and calculate p i The distance to its local plane is calculated, and points whose distance exceeds a certain distance threshold are removed to obtain the point cloud body.

3. The method according to claim 1, characterized in that The step of downsampling the point cloud body by using a three-dimensional cube grid filter to obtain a downsampled point cloud body comprises: A three-dimensional cube grid of the point cloud body is established, and points in the same cube are merged into a single point to obtain a downsampled point cloud body.

4. The method according to claim 1, characterized in that: The steps of re-dividing each of the face areas into corresponding three-dimensional cubic grids, judging the uniformity of point cloud distribution in each grid of each of the face areas based on the eigenvalues ​​of the covariance matrix of the k-neighborhood points, and determining the grid type to which each grid of each of the face areas belongs, wherein the grid type includes an even point grid and an uneven point grid, include: Re-divide each of the face regions into corresponding three-dimensional cubic grids, and for each point of the downsampled point cloud body, take its k2 neighboring points to form a neighborhood point set, and determine the neighborhood point covariance matrix C of each point, which is expressed as: Where X, Y, and Z are vectors with k2 rows and 1 column, representing the three-dimensional rectangular coordinates of k2 neighborhood points, respectively. cov(X,Y) and cov(Y,X) represent the covariance of two vectors X and Y, respectively. cov(X,Z) and cov(Z,X) represent the covariance of two vectors X and Z, respectively. cov(Y,Z) and cov(Z,Y) represent the covariance of two vectors Z and Y, respectively. Calculate the eigenvalue of the covariance matrix of the neighborhood points of each point, determine the maximum eigenvalue of the covariance matrix corresponding to each point, and determine the characteristic threshold value according to the calculation result of the maximum eigenvalue of each point; The maximum eigenvalue corresponding to each point in each grid of the three-dimensional cube grid of each face area is counted, and the ratio of the number of points exceeding the characteristic threshold to the total number of points in the grid is calculated. The grid exceeding the ratio threshold is a non-uniform point grid, and the grid not exceeding the ratio threshold is an even point grid.

5. The method according to claim 1, characterized in that The step of formulating growth rules based on curvature and normal vector information, extracting the growing inner points of each face region according to the initial seed points corresponding to each face region, calculating the ratio of the number of growing inner points of each face region to the area of ​​the corresponding face region, and determining the gross density of the growing inner points of each face region includes: Calculating the normal vector and curvature of each point of each of the face regions using principal component analysis; The initial seed points corresponding to the face regions are used as the initial seed points in the set of seed points to be searched for the face regions; Sequentially use various sub-points in the seed point set to be searched as current search seed points, and search k3 neighboring points of the current search seed point; Determine the neighboring points whose normal vector angles with the current search seed point are less than the angle threshold as points within the target surface area; According to the curvature threshold, determine whether the curvature of each of the neighborhood points is less than the curvature threshold, add the neighborhood points less than the curvature threshold to the set of seed points to be searched, and delete the current search seed point from the set of seed points to be searched; Return to the step of sequentially using various sub-points in the seed point set to be searched as the current search seed point and searching k3 neighborhood points of the current search seed point until the seed point set to be searched is an empty set.

6. The method according to claim 1, characterized in that The density threshold is determined according to the edge length of the three-dimensional cube grid, and the density threshold calculation formula is: Density_th=a*1 / (Edge1^2) Among them, Density_th represents the density threshold, a represents the density reduction coefficient, and Edge1 is the edge length of the three-dimensional cube grid used for downsampling.

7. The method according to claim 1, characterized in that The step of taking the growing interior points corresponding to the face domains less than the density threshold as the target face domain interior points, performing consistency check, extracting the non-growing interior points in the non-uniform point grid of the face domain, and forming the face domain interior points of the face domain with the non-growing interior points and the growing interior points comprises: The growing interior points corresponding to the face domain less than the density threshold are taken as the target face domain interior points, and added to the interior point set corresponding to the face domain, and the least square method is used to obtain the fitting plane equation for the points in the interior point set; Extract all points in the non-uniform point grid in the surface space to form a candidate point data set; Pop up the points in the candidate point data set one by one for consistency check; When the distance between the currently popped-up point and the fitting plane is less than the distance threshold, and the angle between the normal vector of the currently popped-up point and the normal vector of the fitting plane is less than the angle threshold, the currently popped-up point is a non-growing inner point and is added to the inner point set; Repeat the step of popping out points in the candidate point data set in sequence for consistency check until all points in the candidate point data set are popped out, and the points in the final internal point set are composed of the non-growing internal points and the growing internal points, which together constitute the internal points of the face domain.

Images