A building facade point cloud feature line extraction method

Abstract

This invention belongs to the field of point cloud feature extraction technology, specifically relating to a method for extracting feature lines from point cloud data of building facades. The method first acquires point cloud data of the building, extracts feature points including boundary points and fold points, then uses the extracted feature points to perform straight line fitting to obtain feature lines, and finally determines the building feature lines based on the obtained feature lines. When extracting fold points, the relative angle defined in this invention is used. When calculating the relative angle, it is not necessary to simultaneously calculate the normal vectors of the target point and its nearest neighbors; only the normal vector of the target point needs to be calculated. This reduces the impact of actual normal vector calculation deviations on the extraction results, enabling more effective extraction of fold points. The precision, reliability, and accuracy of feature point extraction are higher, thus improving the accuracy of building feature line extraction.

Description

Technical Field

[0001] This invention belongs to the field of point cloud feature extraction technology, specifically relating to a method for extracting feature lines from point clouds on building facades. Background Technology

[0002] Automatic acquisition of feature lines from building facade point clouds can not only intuitively express the shape, structure, and complex topological relationships of building facades, but also effectively simplify the description of building facade point clouds, thus facilitating the reconstruction of richer building geometric models. Therefore, effectively mining and extracting feature lines from building facade point clouds is of great significance. Currently, feature line extraction from two-dimensional images has been extensively and deeply studied, but related algorithms for point cloud feature line extraction still have some problems, such as poor adaptability and low automation capabilities.

[0003] In the field of laser point cloud analysis, the purpose of feature line extraction is to find feature points belonging to the point cloud object and to regularize these feature points. Currently, there is no unified standard for the definition of feature lines. For example, some scholars define the boundaries of an object's surface as feature lines; others divide feature lines in the field of architecture into boundary lines and fold lines: boundary lines are the edges of a building's surface (mainly roof outlines and facade outlines), while fold lines are the lines where building planes intersect (mainly sharp feature lines and broken lines), such as... Figure 1 As shown; in recent years, some scholars have referred to all boundary lines and contour lines in point clouds as three-dimensional edge lines.

[0004] Currently, commonly used methods for extracting feature lines from point clouds of building facades can be categorized into four types: image-based, point-based, plane-based, and deep learning-based feature line extraction methods. 1) Image-based feature line extraction methods generally utilize the corresponding 2D image of the point cloud or convert the point cloud projection into a 2D image. Then, image processing methods are used to extract 2D line segments, which are then mapped to 3D space using reprojection to obtain the corresponding 3D line segments. While image-based feature line extraction methods can incorporate mature line extraction algorithms from the image domain, the optimal projection plane or projection viewpoint is difficult to determine when converting to a 2D image. Furthermore, some information loss occurs during the 2D-3D conversion process, making it unsuitable for complex scenarios. 2) Point-based feature line extraction methods directly extract 3D feature line information from the original point cloud data. This often involves directly obtaining feature points through local geometric features, and then fitting these feature points to obtain 3D feature lines. While point-based feature line extraction methods can effectively extract the geometric features of feature points, they suffer from poor noise resistance, are prone to missed or incorrect extractions, and their extraction results depend heavily on the selection of local features, resulting in low feature line extraction accuracy. 3) Since most buildings are segmented planes, many researchers have proposed plane-based feature line extraction methods based on point-based methods. The main idea is to first segment the point cloud into planes, then extract the intersection points of adjacent planes or the boundary points of each plane to extract building feature points, and finally fit the feature lines of the building point cloud. For feature line extraction methods that directly segment the point cloud into planes, the quality of the extraction result directly depends on the plane segmentation effect, and this method tends to extract larger intersecting planes, thus potentially losing some detailed features of the building. Conversely, feature line extraction methods that oversegment into small planes are prone to more incorrect extractions, resulting in low feature line extraction accuracy. 4) Deep learning methods can extract feature lines that are more suitable for human cognition, but this method requires pre-labeling a large number of feature points and spending a lot of training time, and it is difficult to apply to complex and ever-changing real-world scenarios. Compared with other methods, its advantages are still not significant. In summary, current feature line extraction methods suffer from difficulties in applying to complex and ever-changing real-world scenarios and have relatively low feature line extraction accuracy. Summary of the Invention

[0005] The purpose of this invention is to provide a method for extracting feature lines from point clouds of building facades, in order to solve the problem of low extraction accuracy in existing feature line extraction methods.

[0006] To solve the above technical problems, this invention provides a method for extracting feature lines from point cloud data of building facades, comprising the following steps: 1) acquiring point cloud data of the building and extracting feature points from it; the feature points include boundary points and fold points, wherein the boundary points are measurement points reflecting the boundary features of the building, and the fold points are intersection points between adjacent planes or curved surfaces, and the fold points are determined by the following method: solving for target point O i nearest neighbor P i j The normal vector n of the target point is solved by using the fitted plane obtained from all nearest neighbor points. i Let i ∈ {1,2,…,n}, j ∈ {1,2,…,k}; find an optimal projection plane M, which utilizes the normal vector n that forms the maximum included angle. i sum vector O i P1 is obtained, where P1 represents the nearest neighbor point corresponding to the vector that forms the largest angle with the normal vector; projecting all nearest neighbor points onto the optimal projection plane M yields each vector O. i P j Projection vector on the optimal projection plane M Calculate the normal vector n i With each projection vector 1) Find the angle between the points and calculate the average value, and use the result as the relative angle; 2) Determine the target point whose relative angle is greater than the set relative angle threshold as the folding point; 3) Use the extracted feature points to perform straight line fitting to obtain the feature line; 4) Determine the building feature line based on the obtained feature line.

[0007] Its beneficial effects are as follows: This invention fully explores the topological relationships of the neighborhood space between feature points and non-feature points, proposes a more robust feature point extraction method, and classifies feature points into two categories: boundary points and fold points, to better distinguish different feature information of buildings. Furthermore, by using the relative angle defined in this invention to extract fold points, it is not necessary to calculate the normal vectors of the target point and its nearest neighbors simultaneously when calculating the relative angle; only the normal vector of the target point needs to be calculated. This reduces the impact of actual normal vector calculation deviations on the extraction results, enabling more effective extraction of fold points. The precision, reliability, and accuracy of feature point extraction are higher, improving the accuracy of building feature line extraction. Moreover, experimental results show that this method can still better preserve the feature information at the fold points in the intersection of different planes / surfaces, resulting in better extraction performance.

[0008] Furthermore, the boundary points mentioned in step 1) are determined using the following method: ① Solve for the target point O i nearest neighbor P i j The normal vector n of the target point is solved by using the fitted plane obtained from all nearest neighbor points. iProjecting all nearest neighbor points onto the fitting plane yields the corresponding nearest neighbor projection points. ② Project all nearest neighbor points Perform coordinate transformation to a plane parallel to the two-dimensional plane; ③ Use the convex hull algorithm to solve for the nearest neighbor convex hull point set of the target point, and make the following judgment on the target point: If the target point belongs to the convex hull point set, then the target point is a boundary point; if the target point does not belong to the convex hull point set, then use one of the following two methods to continue the judgment process to determine whether the target point is a boundary point: Method 1: Calculate the distance from the target point to each side of the convex hull, and define the minimum distance as the convex hull distance value of the target point; determine the target point whose convex hull distance value is less than the set distance threshold as a boundary point, otherwise it is not a boundary point; Method 2: Use the target point as a triangle A triangle is formed by taking one vertex of a shape and using two adjacent convex hull points as the other two vertices of the triangle, thus obtaining multiple triangles with the target point as one vertex. The nearest neighbor ratio and vertex angle value of each triangle are determined. The nearest neighbor ratio is the ratio of the number of nearest neighbors in the triangle to the total number of nearest neighbors of the target point. The vertex angle value is the vertex angle value of the triangle with the target point as its vertex. If there is a target point with a nearest neighbor ratio less than a set ratio threshold and the largest vertex angle value among the vertex angle values ​​of triangles with a nearest neighbor ratio less than the set ratio threshold is greater than the set vertex angle threshold, then the target point is determined to be a boundary point; otherwise, it is not a boundary point.

[0009] The beneficial effects are as follows: Boundary points can be extracted using either of two methods. The first method utilizes convex hull distance values, which can effectively extract boundary points inside and outside buildings. Experimental results show that this method can extract more significant boundary feature points, better preserve boundary feature information, and achieve better extraction results, solving the problem that existing technologies cannot extract boundary points of internal voids. The second method utilizes the proportion of nearest neighbors and vertex angle values. This method can extract potential boundary feature points describing boundary characteristics, better preserve feature information, and is less affected by different facade types and point cloud density variations, exhibiting strong adaptability. Both methods can utilize the convex hull method to find any missed boundary points.

[0010] Further, step 3) includes merging the feature lines of the fitted lines using the following method to obtain the building feature lines based on the merged feature lines: 3-1) Determine the endpoints of the feature lines obtained in step 2), and perform a neighborhood search on the endpoints of the feature lines to find the endpoints that cover other feature lines within the neighborhood. The endpoints of other feature lines covered within the neighborhood are called the nearest neighbor endpoints, and the endpoints for neighborhood search are called the original endpoints; 3-2) Determine the included angle and tilt distance between the two feature lines corresponding to the original endpoints and the nearest neighbor endpoints using the following method: If the two feature lines are coplanar straight lines, then directly calculate the included angle between the two feature lines; [The text abruptly ends here, likely due to an incomplete translation or a missing section.] The line is translated to one end of another feature line. After translation, the distance from the other end of one feature line to the other feature line is recorded as the inclination distance between the two feature lines. If the two feature lines are skew lines, one feature line is projected onto the plane formed by one end of one feature line and the other feature line to obtain a projected line. The projected line and the other feature line are processed according to the coplanar line processing method to determine the included angle and inclination distance between the two feature lines. 3-3) If the included angle between the two feature lines is less than the set included angle threshold and the inclination distance between the two feature lines is less than the set inclination threshold, the two feature lines are merged to obtain the merged feature line.

[0011] Furthermore, step 3) also includes optimizing the merged feature lines using the following method to obtain the building feature lines: 3-4) Performing a neighborhood search on the endpoints of the merged feature lines to determine the number of nearest neighbor endpoints covered within the neighborhood; 3-5) Determining the common endpoint of all merged feature lines within the neighborhood based on the number of nearest neighbor endpoints and performing corresponding processing, and optimizing the merged feature lines based on the common endpoints to obtain the building feature lines; the common endpoint is determined using the following method: if there is only one nearest neighbor endpoint, then there are two merged feature lines within the neighborhood; if the two merged feature lines are coplanar lines, then the intersection of the two merged feature lines is directly taken as the two merged feature lines. Common endpoint; if there is one nearest neighbor endpoint, then there are two merged feature lines in the neighborhood. If the two merged feature lines are skew lines, then one of the feature lines is projected onto the plane formed by one endpoint of the merged feature line and the other merged feature line to obtain the projected line. The projected line and the other merged feature line are then processed according to the coplanar line processing method to obtain the common endpoint of the two skew lines. If there are two nearest neighbor endpoints, then there are three merged feature lines in the neighborhood. The intersection point of one merged feature line to the plane formed by the other two merged feature lines is calculated, and the average of the coordinates of the three intersection points is taken as the common endpoint of the three merged feature lines.

[0012] Its beneficial effects are as follows: In the face of problems such as discontinuous feature lines and non-closed endpoints of nearby feature lines, the above merging and optimization operations can make the feature line extraction results at corners such as windows more complete, making the extracted feature lines more practical.

[0013] Further, in step 2), the least squares method with robust estimation is used to fit the line to obtain the characteristic line. The least squares method with robust estimation includes the following steps: establishing a parametric adjustment function model of the spatial line equation; determining the error equation based on several feature points on a line segment; obtaining the parametric approximate solution of the spatial line equation using the least squares principle; and optimizing the obtained parametric approximate solution using the IGGⅢ weight function as an equivalent weight function to obtain the final parametric solution. The spatial line equation corresponding to the final parametric solution corresponds to the expression of the characteristic line.

[0014] Furthermore, before performing line fitting in step 2), the feature points need to be clustered.

[0015] Its beneficial effects are as follows: In the feature point fitting stage, in the face of the problem that the fitted feature lines have deviations, which affect the results of subsequent feature line merging and optimization, the weighted iterative least squares method based on the IGGⅢ weight function as the equivalent weight function can reduce the impact of coarse errors contained in the feature point clustering region on the accuracy of line fitting, and obtain a more accurate line fitting result, thus laying a corresponding foundation for subsequent feature line merging and optimization.

[0016] Furthermore, clustering is performed using the following method: the RANSAC algorithm is used to calculate the line direction vector of the nearest neighbor set of each feature point, and this vector is used as the direction vector of the corresponding feature point; the RG clustering algorithm is then used to cluster all feature points by combining the direction vectors of each feature point.

[0017] Its beneficial effects are: the RANSAC algorithm can estimate the direction vector of each point and has good robustness to noise points.

[0018] Furthermore, after clustering all feature points using the RG clustering algorithm to obtain the initial clustering result, the method also includes a step of further clustering outliers and small regions in the initial clustering result using a re-clustering method, wherein the small regions are clustering regions where the total number of feature points is less than a set threshold.

[0019] Its beneficial effect is that re-clustering is performed after the initial clustering, which improves the clustering accuracy.

[0020] Furthermore, the KD-tree method is used to solve for the target point O. i nearest neighbor P i j .

[0021] Furthermore, the Rodriguez rotation formula is used to project all nearest neighbor points. Perform coordinate transformation. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the building's facade feature lines;

[0023] Figure 2 This is an overall flowchart of the method for extracting feature lines from point clouds of building facades according to the present invention;

[0024] Figure 3(a) and Figure 3(b) are schematic diagrams of convex hull solving in the ideal case and the general case, respectively;

[0025] Figure 4 This is a schematic diagram of nearest neighbor point projection and coordinate transformation;

[0026] Figures 5(a) and 5(b) are schematic diagrams of the convex hull triangulation of the boundary points and interior points, respectively.

[0027] Figure 6(a) and Figure 6(b) are respectively the neighborhood distribution map of a planar point and the projection distribution map of a planar point and its nearest neighbor points;

[0028] Figures 6(c) and 6(d) are the neighborhood distribution map of the fold point and the projection distribution map of the fold point and its nearest neighbor points, respectively.

[0029] Figures 7(a) and 7(b) are schematic diagrams of the merging of two feature lines located on coplanar lines and skew lines, respectively.

[0030] Figure 8(a) and Figure 8(b) are schematic diagrams of the optimization of characteristic lines for finding the common endpoint of two coplanar lines and two skew lines, respectively.

[0031] Figure 8(c) is a schematic diagram of the optimization of the characteristic line for finding the common endpoint of three straight lines;

[0032] Figures 9(a), 9(b), 9(c), and 9(d) are the original point cloud data of Data 1, Data 2, Data 3, and Data 4 in Experiment 1, respectively.

[0033] Figure 10(a) and Figure 10(b) are the point cloud data map and boundary point truth map of Data 1 in Experiment 1, respectively;

[0034] Figures 10(c), 10(d), and 10(e) are the boundary point extraction results of Data 1 using comparison method 1, comparison method 2, and the method of the present invention, respectively.

[0035] Figure 11(a) and Figure 11(b) are the point cloud data map and boundary point truth map of Data 2 in Experiment 1, respectively;

[0036] Figures 11(c), 11(d), and 11(e) are the boundary point extraction results of Data 2 using comparison method 1, comparison method 2, and the method of the present invention, respectively.

[0037] Figure 12 This is a graph showing the statistical results of the post-test unit weight mean square error for the two methods in Experiment 1;

[0038] Figure 13(a) shows the feature point extraction results in Experiment 1;

[0039] Figures 13(b) and 13(c) are the fitting results of the least squares method and the method of the present invention in Experiment 1, respectively.

[0040] Figures 14(a) and 14(b) are the feature line fitting results and feature line merging and optimization results in Experiment 1, respectively.

[0041] Figures 15(a) and 15(b) are the matching results of the feature lines extracted by comparison method 5 and the method of the present invention in Experiment 1, respectively, with the model;

[0042] Figure 16 This is the point cloud data map of Data 4 from Experiment 1;

[0043] Figures 17(a) and 17(b) are the results of extracting building facade feature points from Data 4 point cloud data using the comparative methods (including comparative method 2 and comparative method 4) and the method of the present invention, respectively, in Experiment 1.

[0044] Figures 18(a) and 18(b) are the results of extracting building facade feature lines from Data 4 point cloud data using comparative method 5 and the method of the present invention, respectively, in Experiment 1.

[0045] Figures 19(a) and 19(b) are the original point cloud data map and the true value map of boundary feature points for Data 1 in Experiment 2, respectively.

[0046] Figures 19(c), 19(d), and 19(e) are the boundary feature point extraction results of Data 1 extracted using comparison method 6, comparison method 7, and the method of the present invention, respectively, in Experiment 2.

[0047] Figures 20(a) and 20(b) are the original point cloud data map and the true value map of boundary feature points for Data 2 in Experiment 2, respectively.

[0048] Figures 20(c), 20(d), and 20(e) are the boundary feature point extraction results of Data 2 extracted using comparison method 6, comparison method 7, and the method of the present invention, respectively, in Experiment 2.

[0049] Figure 21(a) is the original point cloud data of Data 3 in Experiment 2;

[0050] Figures 21(b), 21(c), and 21(d) are the boundary feature point extraction results of Data 3 extracted using comparison method 6, comparison method 7, and the method of the present invention, respectively, in Experiment 2.

[0051] Figure 22(a) is the original point cloud data of Data 3 in Experiment 2;

[0052] Figures 22(b), 22(c), and 22(d) are the boundary feature point extraction results of Data 4 using comparison method 6, comparison method 7, and the method of the present invention, respectively, in Experiment 2.

[0053] Figure 23(a) is the original point cloud data of Data 3 in Experiment 2;

[0054] Figures 23(b), 23(c), and 23(d) are the results of extracting boundary feature points of Data 5 using comparison method 6, comparison method 7, and the method of the present invention, respectively, in Experiment 2. Detailed Implementation

[0055] This invention delves into the different spatial topological relationships between feature points and non-feature points, proposing a method for extracting feature lines on building facades. The main contributions of this method include: 1) Fully exploring the neighborhood spatial topological relationships between feature points and non-feature points, resulting in a more robust feature point extraction method. For boundary point extraction, a method based on convex hull distance and a method based on convex hull triangulation are proposed; for folded points, a method based on relative angles is proposed. 2) Introducing a feature line fitting method and proving the effectiveness of the weighted iterative least squares fitting method based on the IGGⅢ weight function. 3) Proposing a more comprehensive feature line endpoint search method to merge and optimize feature lines, improving the problems of discontinuity in the original feature lines and the lack of common endpoints between neighboring feature lines.

[0056] Method Implementation Examples:

[0057] An embodiment of the method for extracting feature lines from point clouds of building facades according to the present invention, the overall process of which is as follows: Figure 2 As shown.

[0058] Step 1: Obtain building point cloud data and extract feature points, including boundary points and fold points.

[0059] 1. Boundary Point Extraction. This invention defines boundary points as measurement points that reflect the boundary characteristics of a building, such as facade outline points, internal hole points, and other measurement points with boundary properties. The extraction effect often depends on the quality of the building point cloud and the performance of the extraction method. As can be seen from this definition, the expected effect of boundary point extraction is to extract measurement points that can accurately reconstruct the building boundary as much as possible.

[0060] A two-dimensional convex hull is the smallest convex polygon on a plane that contains a given set of points. In an ideal plane, the boundary points are distributed in a continuous straight line. Taking a nearest neighbor count of k=8 as an example, the schematic diagram of the convex hull of the boundary points and interior points (non-boundary points) is shown in Figure 3(a). From this figure, it can be seen that the convex hull point set obtained by using the nearest neighbors of the boundary points (A, B, C) contains the boundary points, while the convex hull point set of the interior points (D, E) does not contain them. In actual data acquisition, the point cloud data obtained is non-uniform, and its boundary points are often disordered. As shown in Figure 3(b), observing the convex hulls of the boundary points (A′, C′), potential boundary points (B′), and interior points (D′, E′), it can be found that the boundary points and interior points still retain their original characteristics, while the convex hull point set obtained by using the nearest neighbors of the potential boundary points does not contain the boundary points. However, the distribution of the nearest neighbors of the potential boundary points is biased to one side. When the number of neighborhoods is appropriate, there will be a significant difference between the minimum distances from potential boundary points and interior points to the edges of the convex hull set. Therefore, it can be assumed that if the convex hull set of the target point contains that point, then that point must be a boundary point; if the convex hull set of the target point does not contain that point, further filtering is possible. There are two filtering methods: one is to use the minimum distances from the target point to the edges of the convex hull (called the convex hull distance value) for filtering, and the other is to use the constructed convex hull triangulation for filtering. The specific steps for obtaining boundary points are as follows:

[0061] 1) Project the nearest neighbors of the target point to obtain the target point O. i In the fitting plane L i The nearest neighbor projection point.

[0062] For any target point O in the point cloud data i (i∈{1,2,…,n}) Construct a KD-tree and perform a k-nearest neighbor search to obtain its k-nearest neighbor set P. i j (j∈{1,2,…,k}), and then obtain the least squares fitting plane L through the nearest neighbor points. i To fit the plane L i The normal vector is used as the target point O. i normal vector n i Principal component analysis (PCA) is used to calculate the fitted plane L. iThe problem of normal vectors is transformed into the problem of finding the eigenvector corresponding to the minimum eigenvalue of the covariance matrix Q obtained from the nearest neighbor points, as shown in Equation (1).

[0063]

[0064] In the formula, p j (j=1,2,…,k) are the nearest neighbors of point p. Let e1 be the geometric centroid of the nearest neighbor point, and e2, e3 be the eigenvectors corresponding to the eigenvalues.

[0065] Assume target point O i =(x i ,y i ,z i ), whose nearest neighbor is P. i j =(x j ,y j ,z j Nearest neighbor points are projected onto the fitting plane L. i The nearest neighbor projection points are The normal vector is n i = (a, b, c). Based on geometric relationships, the fitting plane L passing through this point can be derived. i Specifically, it is expressed as:

[0066]

[0067] Further solving yields the nearest neighbor projection points. The coordinates are:

[0068]

[0069] 2) Perform coordinate transformation on the projected nearest neighbor points. To use the convex hull algorithm in the 2D plane, rotate the 3D plane points to be parallel to the xoy plane. Rotation based on Rodrigues's rotation formula is chosen, and the main steps are as follows:

[0070] Step 1: Based on the target point normal vector n i With the normal vector n of the xoy plane z = (0,0,1) Calculate the rotation angle and the rotation axis l.

[0071]

[0072] Step 2: Based on rotation angle Given the rotation axis l = (p, q, t), solve for the rotation matrix R using Rodriguez's formula, and then multiply the rotation matrix R by left to obtain the nearest neighbor projection points parallel to the xoy plane. The nearest neighbor projection and coordinate transformation are shown below. Figure 4 As shown.

[0073]

[0074]

[0075] 3) Assume the target point O i The set of convex hull points is Ω iλ (λ∈{1,2,…,r}, where r is the number of points in the convex hull set of the i-th target point. If a target point belongs to this convex hull set, it is considered a boundary point, and the convex hull distance value Γ is set. iλ It is 0, that is, O i ∈Ω iλ Otherwise, you can use any of the following methods to determine whether the target point is a boundary point:

[0076] Method 1: Further processing based on the convex hull distance value: Calculate the distance from the target point to each edge of the convex hull, and define the minimum distance as the convex hull distance value Γ. iλ The distance threshold η is set (the specific value is introduced in the subsequent experiment) to determine whether it is a boundary point. The judgment condition is shown in equation (7).

[0077]

[0078] Method 2 involves further processing based on the proportion of nearest neighbors and the vertex value:

[0079] Step 1: Construct convex hull triangles with the target point as the center, as shown in Figure 5(a) and Figure 5(b). That is, a triangle is formed with the target point as one vertex of the triangle and two adjacent convex hull points as the other two vertices of the triangle, thus obtaining multiple triangles with the target point as one vertex of the triangle.

[0080] Step 2: Calculate the percentage of nearest neighbors δ for each triangle. m and vertex value m represents the triangle number. The nearest neighbor percentage δ is also represented. m t represents the number of nearest neighbors contained in the triangle numbered m. m The ratio of the number of all nearest neighbors k of the target point, i.e. vertex value Let m be the vertex angle value of the triangle with the target point as its vertex.

[0081] Step 3: Set a percentage threshold δ0 and a vertices threshold. If the proportion of nearest neighbors of each triangle is δ mIf all the nearest neighbors of a target point are greater than or equal to the set proportion threshold δ0, then the target point is considered to be evenly distributed in the vicinity of the target point, and the target point is directly determined to be an interior point, as shown in Figure 5(b); for triangles with a proportion less than the set proportion threshold δ0, the maximum vertex angle value is obtained statistically. (As shown in the convex hull triangulation of the potential boundary feature points in Figure 5(a), when the proportion of multiple triangles such as ① and ② is less than or equal to the set proportion threshold δ0, the maximum vertex value is used for the next step of screening.) If the vertex value Greater than the set vertex threshold If the target point is not a boundary point, it is considered a boundary point; otherwise, it is considered a non-boundary point.

[0082] 2. Fold Point Extraction. Fold points are the intersections of adjacent planes or surfaces, and each point has an intersecting structure of adjacent surfaces in its neighborhood. Based on this, such as... Figures 6(a) to 6(d) As shown, by analyzing and mining the topological differences in the neighborhood space between folded and non-folded points, it can be found that: after finding a projection surface that reflects the distribution of the nearest points on both sides of the target point (ideally, it can be considered as a plane passing through the target point and perpendicular to the folding line), the target point and its nearest points are connected and projected. For planar points, after projection, the nearest points are roughly distributed near a straight line, and the angle between the line connecting each nearest point to the target point and the normal vector is approximately 90°, as shown in Figures 6(a) and 6(b); for folded points, after projection, the nearest points are roughly distributed near two straight lines, and the angle between the line connecting each nearest point to the target point and the bidirectional normal vector is significantly greater than / less than 90° (this angle depends on the intersection between adjacent planes or surfaces), as shown in Figures 6(c) and 6(d). Therefore, this invention uses this angle as a significant indicator (called the relative angle) to describe folded points, thereby intuitively extracting folded points. The specific steps for obtaining the folding angle are as follows:

[0083] 1) Find the nearest neighbors of the target point and their normal vectors. Refer to step 1) of "Boundary Point Finding" for the specific method to find the target point O. i P, the nearest neighbor of (i∈{1,2,…,n}) i j (j∈{1,2,…,k}) and its normal vector n i .

[0084] 2) Find an optimal projection surface. Since the target point O... i normal vector n i With vector O i P i The angle between The larger the value, the closer the plane formed by the two is to the target point O. i And a plane perpendicular to the fold line. Therefore, first calculate the included angle. And select the maximum included angle θ i Then, using the normal vector n that forms the maximum included angle... i sum vector O i P1 obtains the optimal projection plane M (assuming n). i With O i The angle between P1 is the largest, and the solution process is shown in Figure 6.

[0085]

[0086] 3) Calculate the relative angles and extract the folding points. First, project the nearest neighbor points onto the optimal projection plane M to obtain the vector O. i P j Projection vector on the optimal projection plane M The calculation process is shown in equation (9). Then calculate the normal vector n. i with vector The angle between the two points is taken as the average to obtain the angle ω. i Define ω i Let ω be the relative angle, and the calculation process is shown in equation (10). Finally, by combining the set relative angle threshold ω0 (the specific value will be introduced in the subsequent experiments), we can determine whether the point is a folding point based on the size of the relative angle, and determine the points with relative angles greater than the set relative threshold as folding points. At the same time, considering the bidirectionality of the normal vector, in order to facilitate calculation and extraction, when ω i When the angle is ≥90°, subtract the angle from 180° (the result of the subtraction is still recorded as ω). i ); when ω i When the angle is less than 90°, it remains unchanged.

[0087]

[0088]

[0089] Step 2: Based on the feature points extracted in Step 1, further fit the feature points of the building to obtain feature lines.

[0090] 1) Feature Point Clustering. The feature points extracted in Step 1 are still unordered and may contain noise points that meet the extraction criteria. Therefore, clustering is required to facilitate subsequent feature point fitting. The RANSAC algorithm can estimate the direction vector of each point and has good robustness to noise points. In this step, the RANSAC algorithm is first used to estimate the direction vector of each point, and then the RG algorithm is used to cluster the feature points based on the known direction vectors of each point.

[0091] Step 1: Under the constraint of the line model threshold, use the RANSAC algorithm to estimate the line direction vector calculated by the nearest neighbor set of each target point as the direction vector of the target point.

[0092] Step 2: Cluster the feature points using the RG method by combining the direction vectors of each target point. Let V i and V j Let be the direction vector of the i-th feature point and the direction vector of the j-th nearest neighbor feature point, respectively, and let δ be the angle between the two direction vectors. ij The calculation method is shown in equation (10). If the included angle δ ij <15° or δ ij If the angle is greater than 165° (the threshold set in this embodiment can be adjusted according to specific circumstances), then the two points are considered to belong to the same feature line clustering region.

[0093]

[0094] Step 3: After the initial clustering, there are still some outliers and small regions (clustering regions where the total number of points is less than a set threshold T, which is generally 5 to 10 in this embodiment). To eliminate the influence of such point sets on the subsequent feature line extraction results, the outliers and small regions are further clustered using the re-clustering approach.

[0095] 2) Feature Point Fitting. Among common line fitting methods, the RANSAC algorithm and the least squares method are widely used. The RANSAC algorithm determines the optimal parameters of the linear model from data that may contain noise points through iterative solutions. Although it has strong robustness, it is difficult to reasonably set the model error threshold. The least squares method has good fitting accuracy, but it lacks robustness and is sensitive to outliers. Therefore, to reduce the impact of outliers contained in cluster regions on the accuracy of line fitting, this invention introduces robust estimation, constructs a weight factor function based on standardized residuals, and uses the IGGⅢ weight function as an equivalent weight function to construct a weighted iterative least squares line fitting method. The specific process is as follows:

[0096] Step 1: Based on the clustering of feature points, the feature points are divided into points on several line segments. According to the relevant knowledge of spatial line equations, a parametric adjustment function model is established, as shown in equation (12).

[0097]

[0098] In the formula, a1, b1, a2, b2 are the parameters of the equation of the spatial straight line.

[0099] Therefore, there exist m characteristic points {x} on a certain line segment. i ,y i ,z i}(i=1,2,…,m), the error equation is:

[0100]

[0101] Equation (13) can be further written in matrix form as follows:

[0102]

[0103] In the formula,

[0104] Using the least squares principle and equation (14), the parametric approximation of the equation of a spatial line can be obtained:

[0105] X0=(B T B) -1 B T L (15)

[0106] In the formula, X0 = [ 0 a1 0 b1 0 a2 0 b2] T .

[0107] Step 2: After obtaining approximate parameter values, an iterative solution method is constructed to obtain a more accurate equation of the line. The error equation and its corresponding matrix form can be expressed as follows:

[0108]

[0109]

[0110] In the formula,

[0111] Based on the least squares criterion function:

[0112] V k T P k V k =min (18)

[0113] Furthermore, we can obtain:

[0114]

[0115] In the formula, k is the number of iterations.

[0116] When k=1, without introducing robust estimation, and using unit weight P1 for initial weighting, the following can be obtained:

[0117]

[0118] In the formula,

[0119] Step 3: Calculate the equivalent weights using the IGGⅢ equivalent weight function, and reweight each point to obtain the weight function matrix P2. The weight factor expression for the IGGⅢ equivalent weight function is as follows:

[0120]

[0121] In the formula, It is the standardized residual. k0 and k1 are harmonic coefficients. k0 is usually selected from 1.0 to 1.5, and k1 is usually selected from 2.5 to 8.0 (in this embodiment, k0 is 1.0 and k1 is 4.0).

[0122] For the k-th iteration, we have:

[0123] [k] p i =γ i [k-1] p i (twenty two)

[0124] Adjustment is performed again according to formula (19), if Then stop the iteration and output the final parameter solution. Here, ε0 is a very small positive number; otherwise, repeat the previous steps to calculate the corresponding equivalent weights until the condition is met. (In this embodiment, ε0 is taken as 10) -6 If the iteration stops and the final parameter solution is obtained, then the iteration stops.

[0125] After obtaining the linear parameter solutions corresponding to each feature line region through the above steps, the corresponding spatial linear equation can be obtained by combining it with equation (12).

[0126] Step 3: Based on the feature line extraction results obtained in Step 2, further merge and optimize the feature lines using the feature line endpoint neighborhood search method.

[0127] 1) Feature line merging. Due to the uneven distribution of point clouds, feature lines are prone to discontinuity, meaning multiple feature lines are extracted from the same feature line, as shown in Figures 7(a) and 7(b). To overcome this problem, feature lines need to be merged to obtain clearer feature lines. The main steps are as follows:

[0128] Step 1: After obtaining the endpoints of the fitted feature lines, perform a spherical neighborhood search on the endpoints of the fitted lines to locate feature line regions that may have discontinuities. The endpoints of other feature lines covered within the neighborhood are called nearest neighbor endpoints, and the endpoints for which the neighborhood search is performed are called original endpoints.

[0129] Step 2: Determine whether the two feature lines corresponding to the original endpoint and the nearest endpoint are coplanar lines: If so, directly calculate the angle ξ between the two feature lines according to equation (23). 1,2 And the tilt distance ρ (the distance from the other end of feature line L2 to feature line L1 after translating feature line L2 to a point where one end of feature line L1 is the starting point), as shown in Figure 7(a); if not, it is treated as a skew line, and feature line L2 needs to be projected onto the plane formed by one end of feature line L2 and feature line L1 to obtain the projected line, and then the ξ corresponding to feature line L1 is determined according to the treatment method of coplanar lines. 1,2 And the tilt distance ρ, as shown in Figure 7(b).

[0130]

[0131] Step 3: By setting relevant thresholds (ξ in this embodiment) 1,2 The corresponding set angle threshold is selected between 5° and 20°, and the corresponding set slope threshold is selected between 0.050m and 0.200m. The feature lines with small angles and close tilt distances between the fitted feature lines are merged.

[0132] Step 4: After processing, perform the next nearest neighbor endpoint search, and repeat Step 2 to Step 3 until all nearest neighbor endpoints have been searched and processed accordingly.

[0133] 2) Feature line optimization. Because feature points at internal holes and corners are easily lost during feature point extraction, it's difficult to form continuous feature lines during subsequent feature point fitting, resulting in some missing or misaligned lines. Therefore, it's necessary to properly repair these inflection points. The main steps are as follows:

[0134] Step 1: Use the spherical neighborhood search method to find the corresponding missing inflection point and determine the number of nearest neighbor endpoints.

[0135] Step 2: For two feature lines (with only one nearest neighbor endpoint), first determine if they are coplanar lines. If so, directly find the intersection point of the two feature lines as their common endpoint, as shown in Figure 8(a). If they are not coplanar lines, randomly select one endpoint of one feature line to form a plane with the other feature line, then project the feature line and process it as two coplanar lines, as shown in Figure 8(b). If there are three feature lines (with two nearest neighbor endpoints), calculate the intersection point of each feature line with the plane formed by the other two feature lines, and take the average coordinate of the three intersection points as the common endpoint of these three feature lines, as shown in Figure 8(c). Furthermore, building feature lines rarely involve four feature lines.

[0136] The simulation experiments are conducted below. Experiment 1 is used to verify the effectiveness of the entire method. In Experiment 1, the method based on the convex hull distance value is used to extract boundary points. Experiment 2 is used to verify the effectiveness of the feature point extraction method based on the convex hull triangulation algorithm.

[0137] Experiment 1:

[0138] 1. Experimental Platform and Data. All experiments were conducted on a computer with an Intel(R) Core(TM) i7-8700 CPU@3.20GHz and 8GB RAM. Three sets of simulated / model point cloud data and one set of measured point cloud data were selected to verify the effectiveness of the method of this invention. Data 1 is simulated point cloud data of a building facade with uneven density generated using Matlab software, used to test the extraction effect of the method of this invention on boundary points, as shown in Figure 9(a); Data 2 is widely used Block point cloud data, used to test the extraction effect of the method of this invention on fold points, as shown in Figure 9(b); Data 3 is point cloud data generated by constructing a BIM model of the building facade, used to analyze the extraction accuracy and precision of the method of this invention on feature lines, as shown in Figure 9(c); Data 4 is point cloud data of a building facade collected on a campus using a Riegl VZ-400 terrestrial laser scanner, used to verify the feasibility of the method of this invention in extracting feature lines of the building facade, as shown in Figure 9(d). Other relevant information for the four sets of data is shown in Table 1.

[0139] Table 1. Relevant information on experimental data of the method of the present invention.

[0140]

[0141] 2. Parameter Settings. The parameters involved in the method of this invention are mainly distributed in three parts: feature point extraction, feature line generation, and feature line merging and optimization. The parameter settings for feature line generation and feature line merging and optimization are relatively simple, and some parameters have been introduced above. The feature point extraction part mainly includes three parameters: the number of nearest neighbors k, the set distance threshold η, and the set relative angle threshold ω0. Among them, η is an important indicator used to determine whether it is a boundary point. Its main purpose is to describe the distribution of the nearest neighbor set of the target point. That is, the closer the potential boundary point is to the nearby boundary point, the smaller this distance indicator is. After multiple experiments, it can be selected with reference to the average spacing ρ of the point cloud. In this experiment, it is mainly selected between 0.5ρ and 3ρ. ω0 refers to the largest relative angle in the point cloud scene. Theoretically, the relative angle between planes is (0°, 90°). Because relatively gentle or steep fold lines are difficult to extract accurately due to various factors, this experiment selected parameters according to the experimental scenario. When the scenario was more complex, i.e., the relative angle range was larger, the parameter range was mainly selected between 40° and 60°. If the scenario was simpler, i.e., the relative angle range was smaller, a value could be estimated based on the angle of the folding of the two planes. For example, the folding angle of a cube is 90°, so the relative angle index could be selected at around 45° during the experiment. Table 2 shows the parameter settings for the corresponding parts of feature point extraction in the four sets of experimental data in this experiment.

[0142] Table 2. Test parameter settings for the method of the present invention.

[0143]

[0144] 3. Evaluation Indicators. To quantitatively evaluate the performance of the method of this invention, the accuracy of feature point extraction is evaluated by the degree of overlap between the feature points and the corresponding manually extracted standard feature points; the accuracy of feature line extraction is evaluated by the matching rate between the feature lines and the manually labeled feature lines. Three indicators—Correctness, Completeness, and F1 score—are used to evaluate the extraction results. The specific calculation method is shown in Equation (24).

[0145]

[0146] In the formula, TP represents the number of building facade feature points / lines correctly extracted by the method, FP represents the number of building facade feature points / lines incorrectly extracted by the method, and FN represents the number of building facade feature points / lines not extracted by the method.

[0147] 4. During the experiment, several key steps in the feature line extraction process were quantitatively evaluated and analyzed using simulated point cloud data.

[0148] 1) Quantitative Analysis of Feature Point Extraction. This experiment quantitatively analyzed two types of feature points sequentially, using Data1 and Data2 respectively. The two sets of point cloud data were relatively regularly distributed. Data1 used the manually extracted outermost boundary points as the ground truth, while Data2 used the manually extracted more significant folding points as the ground truth.

[0149] ① Quantitative Analysis of Boundary Point Extraction. Two typical boundary point extraction methods were compared: the first method (referred to as Comparison Method 1) uses an improved convex hull algorithm to find boundary points by determining the angle between the line connecting the current point and its nearest neighbors and a fixed direction; the second method (referred to as Comparison Method 2) determines whether a point is a boundary point by calculating the distance from the centroid of the nearest neighbor set to the current point. The boundary point extraction results of the present invention method and the two comparison methods are shown in Table 3, and the extraction effect is as follows: Figure 10(a)~10(e) As shown.

[0150] Table 3 Comparison of boundary point extraction results using different methods

[0151]

[0152] From Table 3 and Figure 10(a)~10(e) It can be observed that Comparison Method 1 can accurately extract the convex hull points containing the entire point set, but it can only extract the outer boundary points of the building and cannot extract the inner boundary points of the facade (boundary points at windows). Comparison Method 2 identifies boundary points by the distance between the centroid and feature points. Although it can extract the inner and outer boundary points of the building, it cannot effectively identify and detect boundary points at significant density change boundaries (this invention also classifies these feature points as boundary points). The method of this invention can effectively extract both inner and outer boundary points of the building, overcoming the limitation of Comparison Method 1's convex hull approach, which can only extract outer boundary points, and can more intuitively reflect the shape and structure of the building. In summary, the method of this invention is superior in terms of correctness, completeness, and F1 score. The extraction results achieve a high degree of consistency with manually extracted boundary points, can extract more significant boundary feature points, and can better preserve boundary feature information.

[0153] ② Quantitative Analysis of Fold Point Extraction. Two fold point extraction methods were compared: Method 3 calculates the normal vectors of the target point and its nearest neighbor set, clusters the normal vectors, and extracts fold points based on the number of clusters; Method 4 determines whether a point is a fold point by calculating the maximum angle difference between the normal vector of the target point and the normal vectors of its nearest neighbor set. The fold point extraction results of the present invention method and the two comparative methods are shown in Table 4, and the extraction effect is as follows: Figure 11(a)~11(e) As shown.

[0154] Table 4 Comparison of Fold Point Extraction Results Using Different Methods

[0155]

[0156] From Table 4 and Figure 11(a)~11(e) It can be observed that while the idea of ​​clustering based on normal vectors in Comparison Method 3 is easy to understand and implement, it is susceptible to the influence of point cloud distribution, resulting in numerous instances of incorrect extraction. Comparison Method 4 identifies folding points by using the maximum angle difference between normal vectors, which can extract folding points relatively well, but it is also prone to incorrectly extracting points surrounding the folding points. Unlike the above two methods, the method of this invention does not require simultaneous calculation of the normal vectors of the target point and its nearest neighbors; it only needs to calculate the normal vector of the target point. This reduces the impact of actual normal vector calculation deviations on the extraction results, enabling more effective extraction of folding points. In summary, the method of this invention outperforms Comparison Methods 3 and 4 in terms of correctness, completeness, and F1 score, exhibiting higher extraction accuracy and better preservation of folding feature information even in the case of intersections of different planes / surfaces, resulting in better extraction performance.

[0157] 2) Quantitative analysis of feature line extraction.

[0158] ① Feature point fitting result analysis. To evaluate the advantages of the weighted iterative least squares method used in the feature point fitting stage, after clustering the feature points of Data 3 data and fitting spatial lines using the least squares method and the method of this invention, the posterior unit weight mean square error is used to evaluate the fitting results (the smaller the value, the higher the fitting accuracy). Figure 12 Table 5 shows the statistical results of the maximum, minimum, and average values ​​of the posterior unit weighted mean errors of the corresponding cluster regions obtained after fitting the cluster regions of each feature point using the two methods. Figures 13(a) to 13(c) The image shows the fitting results of the two methods on the clustered feature points.

[0159] Table 5 Evaluation of feature point fitting results for the two methods

[0160]

[0161] From Table 5, Figure 12 and Figures 13(a) to 13(c)As can be seen, both methods can fit the clustered feature points to obtain the final feature point fitting result. However, due to deviations in the calculation of direction vectors at corners during clustering, some feature points at corners are clustered incorrectly. Consequently, in the feature point fitting stage, the least squares method cannot resist such gross errors, resulting in deviations in the fitted feature lines and affecting the subsequent feature line merging and optimization results. In contrast, the weighted iterative least squares method based on the IGGⅢ weight function as the equivalent weight function can reduce the impact of gross errors in the feature point clustering region on the accuracy of line fitting, obtaining more accurate line fitting results and laying the foundation for subsequent feature line merging and optimization.

[0162] ② Feature line extraction result analysis. The fitted feature lines are then merged and optimized to obtain the final feature lines, as shown in the following figure. Figures 14(a) to 14(b) As shown, merging and optimization operations can make the feature line extraction results at corners such as windows more complete, thus making the extracted feature lines more practical. To further evaluate and analyze the feature line extraction accuracy of the method of this invention, feature lines in Data 3 were extracted using both comparison method 5 and the method of this invention, and the lengths of different categories of feature lines were calculated, including boundary feature lines of different lengths (containing a large number of window feature lines) and folded feature lines. Among them, comparison method 5 is an image-based feature line extraction method. First, the point cloud is segmented into planes through region growing and region merging. Then, the points on each plane are projected into two-dimensional images. Then, the contours of the images are extracted and fitted using the least squares method to obtain two-dimensional line segments. Finally, the two-dimensional line segments are reprojected onto a three-dimensional plane to obtain the corresponding feature lines. Using the BIM model design value corresponding to Data 3 as the true value, the number and length of the extracted feature lines are compared with the true value, and the maximum error, minimum error, and average error of the feature line length are calculated. The statistical results are shown in Table 6. In addition, the feature line extraction results of comparison method 5 and the method of the present invention are matched with the BIM model, wherein the model color is assigned gray and the feature lines are assigned black. The matching results are as follows: Figures 15(a) to 15(b) As shown.

[0163] Table 6 Comparison of feature line length results using different methods

[0164]

[0165] Analysis Table 6 and Figures 15(a) to 15(b)It can be observed that the feature lines extracted by the comparative method 5 exhibit significant missing features and incomplete endpoint closures, and two feature lines are mistakenly extracted. In contrast, the method of this invention extracts more complete feature lines effectively, achieving a more ideal matching result with the corresponding model, particularly at corner points, without any missed or incorrect feature line extraction. Regarding extraction accuracy, both methods achieve centimeter-level accuracy for both boundary and folded feature lines. Using Data 3 data, ignoring the accuracy loss caused by randomness in the simulation point cloud data generation process, the average error of the comparative method 5 (0.0960m) is approximately 7.62 times the average point cloud spacing (0.0126m), while the average error of the method of this invention (0.0223m) is approximately 1.76 times the average point cloud spacing. Therefore, the method of this invention achieves higher feature line extraction accuracy, demonstrating a clear advantage.

[0166] 5. Comparative Experiment. To better analyze the effectiveness and feasibility of the method of the present invention, the feature point and feature line extraction results of the method of the present invention on the measured point cloud data Data 4 were compared with existing methods. Among them, comparison method 2 and comparison method 4 are feature point extraction methods with the same approach as the method of the present invention, and their specific principles have been used and introduced in the simulation experiment in Section 4; comparison method 5 is a feature line extraction method based on the image, and its specific principles have been used and introduced in the simulation experiment in Section 4. Since there are many cluttered point clouds inside the window in Data 4 (as shown in Figure 8(d)), which easily interfere with the extraction of feature lines, the data was preprocessed according to the point cloud reflection intensity before the experiment. The preprocessed Data 4 is shown in Figure 8(d). Figure 16 As shown in Figures 17(a) and 17(b), feature points were extracted using comparative methods 2 and 4, as well as the method of this invention. The extraction results and magnified local details are shown in Figures 18(a) and 18(b). Feature lines were extracted using comparative method 5 and the method of this invention. The extraction results and magnified local details are shown in Figures 18(a) and 18(b). By using manually extracted feature points and manually marked feature lines from Data 4 as ground truth, the feature point / line extraction results of different methods were statistically analyzed, and the results are shown in Tables 7 and 8.

[0167] Table 7 Comparison of Feature Point Extraction Results by Different Methods

[0168]

[0169] Table 8 Comparison of feature line extraction results using different methods

[0170]

[0171] Combining Tables 7 and 8, Figures 16–18(b)The results of extracting feature points / lines of buildings in Data 4 using different methods show that: ① In terms of feature point extraction, the method of this invention is significantly better than comparative methods 2 and 4 in terms of accuracy. This is mainly because comparative methods 2 and 4 have difficulty setting the feature point extraction threshold well for buildings with uneven density, which easily leads to mis-extraction or omission of feature points in the magnified local details shown in Figure 17(a), making it difficult to extract relatively accurate feature points. Comparative methods 2 and 4 can only obtain feature points with a certain width (otherwise, there will be serious missing features). In terms of F1 score, the method of this invention is still better than comparative methods 2 and 4, which further illustrates the reliability and accuracy of the method of this invention in feature point extraction. ② In terms of feature line extraction, comparative method 5 is cumbersome. As can be seen from Figure 18(a), accuracy loss is inevitable in the process of converting from 2D to 3D space. Feature lines in multiple windows and other areas in Data 4 are missed, and relatively closed feature lines cannot be obtained. The method of this invention, on the other hand, performs fitting and optimization based on feature point extraction. The extraction effect mainly depends on the feature point extraction effect and the feature point clustering situation. The method of this invention outperforms the comparative method 5 in overall extraction performance of Data 4. The proposed point-based method is simpler in principle, and its feature line extraction capability can reach the same level (or even higher) as the image-based method. Furthermore, it exhibits fewer omissions / mis-extractions, and the extracted feature lines match the original point cloud more closely and with higher accuracy, thus better restoring the detailed feature information of the building. ③ Overall, the method of this invention achieves good feature line extraction performance from building point clouds, demonstrating advantages in both feature point and feature line extraction accuracy. It can effectively represent the shape features and geometric structure of buildings, thereby achieving a simplified representation of building objects.

[0172] Experiment 2:

[0173] 1. Experimental Platform and Data. All experiments were conducted on a computer with an Intel(R) Core(TM) i7-8700 CPU@3.20GHz and 8GB RAM. One set of model sampling point cloud data and four sets of measured point cloud data were selected to verify the effectiveness of the method. Data 1, Data 2, and Data 4 were point cloud data of partial building facades collected on a school campus using a Riegl VZ-400 terrestrial laser scanner; Data 3 was point cloud data generated by constructing a BIM model of the building facade; Data 5 was a building facade data selected from the open-source dataset Robotic 3D Scan Repsitory-12, which exhibited numerous irregular local protrusions and a complex data distribution. Relevant information for the five sets of experimental data is shown in Table 9.

[0174] Table 9. Related Information on Experimental Data

[0175]

[0176] 2. Evaluation Indicators. To quantitatively evaluate the performance of the method of this invention, four indicators, namely simplification rate (SR), precision (P), recall (R), and F1 score, are used to evaluate the results of extracting feature points of building facade boundaries. The specific calculation method is shown in Equation (25).

[0177]

[0178] In the formula, N B The method represents the number of building facade boundary feature points extracted; N0 represents the total number of building facade point clouds; T P The number of building facade boundary feature points correctly extracted by the method; F P The number of building facade boundary feature points incorrectly extracted by the method; F N The number of building facade boundary feature points not extracted by the method.

[0179] 3. Parameter Settings. The method of this invention extracts feature points of building facade boundaries from experimental data. The main parameters involved include the number of nearest neighbors k, the threshold percentage of nearest neighbors in the triangulation network δ0, and the threshold for setting vertex angles. Three parameters. The proportion threshold δ0 is an important indicator used to determine whether a point is a boundary feature point. Its main purpose is to describe the distribution of the nearest neighbor set of the target point. That is, the closer a potential boundary point is to a nearby boundary feature point, the fewer the nearest neighbors in the triangular network, and the lower the corresponding proportion. Therefore, the threshold is generally chosen to be a small value; set the vertex threshold. These parameters are further filtered under the premise of meeting the proportion threshold δ0. The larger the angle, the closer it is to the nearby boundary feature point, that is, the more likely it is to be a potential boundary feature point. Table 10 shows the parameter settings corresponding to the final extraction effect diagrams of the five sets of experimental data in this experiment.

[0180] Table 10 Test parameter settings for the method of the present invention

[0181]

[0182] 4. Quantitative Experiment. To verify and evaluate the performance of the method of this invention, experiments were conducted using measured point cloud data of partial building facades from Data 1 and Data 2. The two sets of point cloud data have relatively regular distributions, and the outermost boundary feature points extracted manually were used as the ground truth. The improved convex hull algorithm of Comparison Method 6 and the point-based boundary feature point extraction method of Comparison Method 7 were introduced for comparative analysis with the method of this invention. Comparison Method 6 uses an improved convex hull algorithm to find boundary feature points by judging the angle between the line connecting the current point and its nearest neighbor and a fixed direction. Comparison Method 7 proposes a simple and effective detection method for boundary feature points, namely, determining whether a point is a boundary feature point by calculating the distance from the centroid of the nearest neighbor set to the current point. Table 3 summarizes the point cloud quantity, simplification rate (SR), precision (P), recall rate (R), and F1 score of Comparison Method 6, Comparison Method 7, and the method of this invention before and after boundary feature point extraction for the two sets of point cloud data. The extraction results are as follows: Figure 19(a)~19(e) and Figure 20(a)~20(e) As shown.

[0183] Table 11 Quantitative Evaluation Results of Two Types of Point Cloud Data

[0184]

[0185] Comprehensive analysis of Table 11 and Figure 19(a)~19(e) and Figure 20(a)~20(e) It can be observed that: ① Data 1 and Data 2 are point cloud data of curved and planar types, respectively. All three methods can extract the boundary feature points of the two facade shapes, but comparative method 6 cannot extract the boundary feature points inside the facade (as shown in the ellipse in Figure 20(c)). In addition, since the distance between the centroid and the feature point of the boundary feature point at the corner (such as the corner in the circle in 20(d)) is smaller than that of other feature points, comparative method 7 has difficulty extracting the boundary feature points at the corner well. ② All three methods have a high simplification rate, which can simplify the original point cloud data well. Combined with the F1 score, the method of this invention can retain the boundary feature information better while simplifying the data, achieving a better extraction effect. ③ The method of this invention outperforms comparative methods 6 and 7 in terms of the comprehensive performance of precision, recall and F1 score. The extraction results have a high degree of consistency with the manually extracted boundary feature points, and can extract more significant boundary feature points.

[0186] 5. Comparative Experiment. To further demonstrate the effectiveness of building facade boundary feature point extraction, boundary feature points were extracted from Data 3, Data 4, and Data 5 using comparative methods 6 and 7, as well as the method of this invention. The extraction results are as follows: Figures 21(a) to 21(d) , Figures 22(a) to 22(d) , Figures 23(a) to 23(d) As shown. By Figures 21(a) to 21(d) , Figures 22(a) to 22(d), Figures 23(a) to 23(d) The extraction results show that comparative method 6 can only extract the outer boundary feature points of the building facade, while the method of the present invention and comparative method 7 can effectively extract the inner and outer boundary feature points of the building facade, which can more intuitively reflect the shape and structure of the building facade. Among them, the boundary feature point extraction effects of Data 3 and Data 4 are comparable, but for Data 5, comparative method 7 has more serious omissions or mis-extractions, as shown in Figure 23(c). The extracted boundary feature points obviously have boundary missing or local mis-extraction phenomena, while the extraction results of the method of the present invention are more complete and the extracted boundary feature points are more accurate. The main reason for this is that the judgment method of comparative method 7 based on distance index is more suitable for the detection of planar boundary feature points with regular point cloud distribution. For building facades with obvious differences in point cloud density, it is difficult to set a suitable distance threshold, so boundary missing or local mis-extraction is easy to occur. The method of the present invention performs coordinate projection and transformation operations on the nearest neighbor points and uses proportion and angle indexes, which is less affected by local protrusions and differences in point cloud density.

[0187] In summary, this invention addresses the need for reconstructing building facade geometric models. Considering the lack of spatial topological relationships between points in point cloud data, it proposes a method for extracting building facade feature lines that takes into account neighborhood spatial topological relationships. For feature point extraction, boundary points and fold points are extracted using the convex hull distance and relative angle defined in this invention. For feature line generation, a weighted iterative least squares method is used to fit feature points, reducing gross errors introduced during feature point clustering and resulting in more accurate feature lines. Regarding feature line merging and optimization, to address potential issues such as discontinuous feature lines and missing common endpoints, a feature line endpoint sphere neighborhood search method is used, employing relevant geometric constraints to obtain more complete feature lines. Simulation results show that the feature point extraction method of this invention achieves higher accuracy, reliability, and precision compared to similar methods, demonstrating significant advantages and better preservation of building facade feature information. When matching the feature line extraction results with the actual model, the accuracy of this invention's method is significantly higher than the comparative methods, exhibiting a higher degree of fit. Comparative experimental results show that, compared with image-based feature line extraction methods, the method of this invention can obtain more complete building facade feature line extraction results and has strong practicality. Furthermore, this invention proposes a convex hull triangulation algorithm for extracting building facade boundary feature points based on the traditional convex hull algorithm. Experimental results show that the method of this invention can effectively extract building facade boundary feature points, better preserve feature information, and is less affected by different facade types and point cloud density variations, exhibiting strong adaptability.

Claims

1. A method for extracting feature lines from point clouds of building facades, characterized in that, Includes the following steps: 1) Obtain point cloud data of the building and extract feature points including boundary points and fold points; boundary points are measurement points reflecting the boundary features of the building, and fold points are intersection points between adjacent planes or curved surfaces, and are determined by the following method: Solve for the target point nearest neighbor The normal vector of the target point is solved by using the fitted plane obtained from all nearest neighbor points. , , ; Find the optimal projection plane M, which uses the normal vector that forms the maximum included angle. sum vector get, This represents the nearest neighbor point corresponding to the vector that forms the largest angle with the normal vector; Projecting all nearest neighbor points onto M yields the vectors. Projection vector on M ,calculate With each The angle between the two points is calculated, and the average value is taken as the relative angle; the target points with a relative angle greater than the set relative angle threshold are determined as folding points; 2) Use the extracted feature points to fit a straight line to obtain the feature line; 3) Merge the feature lines and use the merged feature lines to obtain the building feature lines; The merging method is as follows: 3-1) Determine the endpoints of the feature lines and perform a neighborhood search to find the endpoints within the neighborhood that cover other feature lines. These endpoints are called nearest neighbor endpoints, and the endpoints for the nearest neighbor search are the original endpoints. 3-2) Determine the angle between the two feature lines containing the original endpoint and the nearest neighbor endpoint. and tilt distance : If the two feature lines are coplanar, then calculate directly. Translate one endpoint of one feature line to the endpoint of another feature line, and denot the distance from the other endpoint of one feature line to the other feature line after the translation as . ; If two feature lines are skew, one feature line is projected onto the plane formed by one endpoint of the first feature line and the other feature line to obtain a projected line; the projected line and the other feature line are then determined using the method for treating coplanar lines. and ; 3-3) If Less than the set included angle threshold and If the tilt is less than the set tilt threshold, the two feature lines will be merged.

2. The method for extracting feature lines from point clouds of building facades according to claim 1, characterized in that, The boundary points mentioned in step 1) are determined using the following method: ① Solve for the target point nearest neighbor The normal vector of the target point is solved by using the fitted plane obtained from all nearest neighbor points. Projecting all nearest neighbor points onto the fitting plane yields the corresponding nearest neighbor projection points. ; ② Project all nearest neighbor points Perform a coordinate transformation to transform to a plane parallel to the two-dimensional plane; ③ The convex hull algorithm is used to solve for the nearest neighbor convex hull point set of the target point, and the target point is judged as follows: if the target point belongs to the convex hull point set, then the target point is a boundary point; if the target point does not belong to the convex hull point set, then one of the following two methods is used to continue the judgment process to determine whether the target point is a boundary point: Method 1: Calculate the distance from the target point to each edge of the convex hull, and define the minimum distance as the convex hull distance value of the target point; determine the target point whose convex hull distance value is less than the set distance threshold as the boundary point, otherwise it is not a boundary point; Method 2: Construct a triangle with the target point as one vertex and two adjacent convex hull points as the other two vertices, thus obtaining multiple triangles with the target point as one vertex; determine the nearest neighbor ratio and vertex angle value of each triangle. The nearest neighbor ratio is the ratio of the number of nearest neighbors in the triangle to the total number of nearest neighbors of the target point, and the vertex angle value is the vertex angle value of the triangle with the target point as its vertex angle; if there is a target point with a nearest neighbor ratio less than a set ratio threshold and the largest vertex angle value among the triangles with a nearest neighbor ratio less than the set ratio threshold is greater than the set vertex angle threshold, then the target point is determined to be a boundary point; otherwise, it is not a boundary point.

3. The method for extracting feature lines from point clouds of building facades according to claim 1, characterized in that, Step 3) also includes optimizing the feature lines after merging using the following method to obtain the building feature lines: 3-4) Perform a neighborhood search on the endpoints of the merged feature lines to determine the number of nearest neighbor endpoints covered within the neighborhood; 3-5) Based on the number of nearest neighbor endpoints and after corresponding processing, determine the common endpoints of all merged feature lines within the neighborhood, and optimize the merged feature lines based on the common endpoints to obtain the building feature lines; the common endpoints are determined using the following method: If there is only one nearest neighbor endpoint, then there are two merged feature lines in the neighborhood. If the two merged feature lines are coplanar lines, then the intersection of the two merged feature lines is directly taken as the common endpoint of the two merged feature lines. If there is only one nearest neighbor endpoint, then there are two merged feature lines in the neighborhood. If the two merged feature lines are skew lines, then one of the feature lines is projected onto the plane formed by one endpoint of the merged feature line and the other merged feature line to obtain the projected line. The projected line and the other merged feature line are then processed according to the coplanar line processing method to obtain the common endpoint of the two skew lines. If there are two nearest neighbor endpoints, then there are three merged feature lines in the neighborhood. Calculate the intersection point of one of the merged feature lines with the plane formed by the other two merged feature lines, and take the average of the coordinates of the three intersection points as the common endpoint of the three merged feature lines.

4. The method for extracting feature lines from point clouds of building facades according to claim 3, characterized in that, In 3-4), a spherical neighborhood search method is used for neighborhood search.

5. The method for extracting feature lines from point clouds of building facades according to claim 1, characterized in that, Step 2) employs a robust least squares method to fit the line and obtain the characteristic line. The robust least squares method includes the following steps: Establish a parametric adjustment function model for the equation of a spatial straight line; determine the error equation based on several characteristic points on a straight line segment; and obtain the parametric approximate solution of the equation of the spatial straight line using the least squares principle. Using the IGGⅢ weight function as an equivalent weight function, the obtained approximate parametric solution is optimized to obtain the final parametric solution. The spatial line equation corresponding to the final parametric solution is the expression of the characteristic line.

6. The method for extracting feature lines from point clouds of building facades according to claim 1 or 5, characterized in that, Before performing line fitting in step 2), the feature points need to be clustered.

7. The method for extracting feature lines from point clouds of building facades according to claim 6, characterized in that, Clustering is performed as follows: the RANSAC algorithm is used to calculate the line direction vector of the nearest neighbor set of each feature point, and this vector is used as the direction vector of the corresponding feature point; the RG clustering algorithm is then used to cluster all feature points by combining the direction vectors of each feature point.

8. The method for extracting feature lines from point clouds of building facades according to claim 7, characterized in that, After clustering all feature points using the RG clustering algorithm, an initial clustering result is obtained. The method also includes a step of further clustering outliers and small regions in the initial clustering result using a re-clustering method, wherein the small regions are clustering regions where the total number of feature points is less than a set threshold.

9. The method for extracting feature lines from point clouds of building facades according to claim 1, characterized in that, Using the KD-tree method to solve for the target point nearest neighbor .

10. The method for extracting feature lines from point clouds of building facades according to claim 2, characterized in that, Apply Rodriguez's rotation formula to all nearest neighbor projection points Perform coordinate transformation.

Images