A progressive optimization method for building contours based on multi-dimensional features and laser point cloud

Abstract

This invention discloses a progressive optimization method for building contours based on multi-dimensional features using laser point clouds. It employs a layer-by-layer optimization approach, refining the building contour extraction from coarse to fine. First, the building point cloud is segmented into facade point clouds (including trees and building walls) and top surface point clouds (including the building roof and trees). Then, based on the characteristics of each facade and top surface point cloud, different features are used for refined processing. This method enables rapid and accurate identification of building point clouds with varying heights and structural complexities. It can eliminate interference from non-building point clouds and the impact of missing facade point clouds while refining the contour point cloud extraction, thereby improving the accuracy of contour extraction.

Description

Technical Field

[0001] This invention belongs to the field of lidar data processing technology and laser scanning data processing technology, specifically involving a progressive optimization method for building outlines based on multi-dimensional features in laser point cloud. Background Technology

[0002] With the continuous development of new technologies such as "smart cities" and "digital twins," the rapid and accurate extraction of building outlines has become a research hotspot. Airborne LiDAR technology can quickly acquire high-precision 3D point cloud data of the Earth's surface and accurately obtain the geometric and topological relationships between ground features. It is an economical and reliable technology for acquiring 3D spatial data of the Earth's surface and has become an important means of obtaining 3D building models for cities, with broad application prospects. Therefore, researching how to quickly and accurately extract buildings from LiDAR point cloud data has significant practical implications.

[0003] Research on building contour extraction based on raw point clouds mainly focuses on how to quickly and accurately identify point clouds of buildings with different heights and structural complexities, while avoiding interference from tree point clouds, etc.

[0004] Currently, building contour extraction often relies on segmentation using one or more spatial geometric features of the point cloud, such as height difference, projection density, normal vector, and curvature, followed by clustering to obtain the point cloud of the building's roof surface. However, this process is prone to problems such as missing low-rise buildings, missing point cloud portions of the building's roof or facade, and misclassification of nearby buildings and trees. Alternatively, the point cloud information can be converted into an image, and then contour extraction can be performed using image processing algorithms. However, this approach faces challenges such as difficulty in determining the grid resolution scale and inaccurate boundary location extraction. In summary, current methods for rapidly extracting LOD1 (Level of Detail) 3D models of buildings from raw point clouds still have many problems that urgently need to be addressed.

[0005] Chinese patent application number 202210672480.0 discloses a method for extracting urban building attributes based on airborne laser point clouds. This method primarily involves calculating the geometric, height, and attribute features of each non-ground laser point and using a random forest classifier to label each point in the non-ground point cloud as a building or vegetation point. Then, given 3D distance thresholds, height distance thresholds, and intensity difference thresholds, region growing is performed on the laser points labeled as buildings to obtain multiple building point cloud clusters. Each building point cloud cluster is then fitted with a plane to obtain the building's structural height and outline information. This method has several drawbacks. First, the use of a random forest classifier for point-by-point labeling of the non-ground point cloud requires manual intervention, significantly reducing processing efficiency and increasing the risk of human error. Second, this method yields a complete point cloud of building categories, which inevitably leads to data redundancy when only LOD1 building outlines need to be extracted, reducing the efficiency of subsequent clustering and outline extraction algorithms. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of the prior art by providing a progressive optimization method for building outlines based on multi-dimensional features using laser point clouds, thereby solving the problems of low data processing efficiency and poor extraction accuracy in existing methods.

[0007] To solve the above-mentioned technical problems, one technical solution adopted by the present invention is to provide a progressive optimization method for building contours based on multi-dimensional features using laser point clouds, the method comprising the following steps:

[0008] (1) Separate non-facade point cloud: Filter the original airborne LiDAR point cloud data to separate ground points from non-ground points, thereby obtaining the point cloud of non-ground points;

[0009] (2) Preprocessing: Calculate the relative height of each point in the non-ground point cloud and remove points whose relative height is less than the set elevation threshold.

[0010] (3) Coarse segmentation: Calculate the normal vector nz of each point in the non-ground point cloud after the removal process, and use the normal vector threshold to perform preliminary segmentation of the building facade point cloud and the top point cloud, thereby achieving coarse segmentation of the building facade point cloud and the top point cloud.

[0011] (4) Refined extraction of facade point cloud: The coarsely segmented building facade point cloud is projected in two dimensions on the XOY plane. The projection range is determined with the corresponding projection point as the center and R2 as the radius. The projection density of each point is calculated. At the same time, the height difference of the point cloud is calculated within the projection range of each point, so as to obtain the density and height difference value of each point. Then, the density threshold and height difference threshold are used to refine the extraction of facade point cloud to remove the interference of tree point cloud.

[0012] (5) Top surface point cloud clustering: While finely extracting the building facade point cloud, first filter and denoise the top surface point cloud, and then cluster each point in the processed top surface point cloud to obtain multiple clustering objects;

[0013] (6) Fine extraction of top surface point cloud: Calculate the planarity of each point in the top surface point cloud and the projected area of the clustering object of the top surface point cloud on the XOY plane. Use the projected area threshold, planarity threshold, and the proportion threshold of the number of points in the planarity point cloud for the fine extraction of the top surface point cloud, and eliminate the clustering objects that do not meet the conditions, so as to obtain the top surface point cloud of the building without interference from other categories;

[0014] (7) Completion of top surface point cloud: Extract the boundary points of the top surface point cloud after fine processing, then search for building facade points near the boundary points of the top surface point cloud, and merge the searched facade points into the top surface point cloud, thus completing the extraction of the point cloud of the complete building top surface and part of the side surface;

[0015] (8) Building contour extraction: Perform a two-dimensional projection of the point cloud obtained in the previous step on the XOY plane, then use the alpha-shape method for boundary point extraction and merging, and then use the RANSAC method for line fitting, regularization, and merging processing to finally obtain the two-dimensional contour of the building.

[0016] In another embodiment of the present invention, the calculation radius selected when calculating the normal vector of each point in step (3) is R1, and the value of R1 is preferably 5.2 m. The normal vector threshold includes a high limit value and a low limit value, where the low limit value is preferably -0.3 and the high limit value is preferably 0.3. That is, the points with the normal vector meeting -0.3 < nz < 0.3 are facade points, otherwise they are top surface points.

[0017] In another embodiment of the present invention, when calculating the projection density of each point in step (4), the KD tree method is used, and the value of R2 is preferably 0.1 m; the projection density is the number of points within the projection range, the density threshold is preferably set to 3, and the height difference threshold is preferably 5 m. The points that simultaneously meet the conditions of projection density > 3 and height difference > 5 m are the qualified facade points and are retained; the points that do not meet the above conditions simultaneously are interference points and are eliminated.

[0018] In another embodiment of the present invention, the process of the KD tree method is as follows:

[0019] Ⅰ. Calculate the variance of each dimension in the k-dimensional data set, and select the dimension k with the largest variance from them:

[0020] <communicated="false">

[0021] Ⅱ. Arrange the data on dimension k in ascending order to obtain the data set in N k Given the number of data points in dimension k, calculate the median m in dimension k:

[0022]

[0023] III. Set the threshold to the median m obtained in the previous step to obtain two sets K. sub_low and K sub_high And create a tree node for storage, the set satisfies the requirements of the following equation (3):

[0024]

[0025] IV. Repeat the above steps until all subsets can no longer be divided; if a subset can no longer be divided, save the data in that subset to a leaf node with no child nodes.

[0026] In another embodiment of the present invention, the top surface point cloud clustering method in step (5) adopts Euclidean clustering or region growing clustering.

[0027] In another embodiment of the present invention, the Euclidean clustering method is as follows: For a point P in space, the k nearest neighbor search methods of KD tree are used to find the k points closest to point P. Among these points, those whose distance is less than a set threshold are clustered into set Q'. If the number of elements in Q' no longer increases, the entire clustering process ends; otherwise, points other than point P in set Q' must be selected and the above process is repeated until the number of elements in Q' no longer increases.

[0028] In another embodiment of the present invention, in step (6), the projection area threshold is preferably 30㎡, the flatness threshold is preferably 0.5, and the flatness point cloud number ratio threshold is 30%. That is, objects that simultaneously meet the conditions of projection area > 30㎡ and (the ratio of points with flatness planarity > 0.5) > 30% are cluster objects that meet the conditions and are retained; objects that do not simultaneously meet the above conditions are interfering cluster objects and are eliminated.

[0029] In another embodiment of the present invention, the alpha-shape method is used to extract the boundary points of the top surface point cloud in step (7); the building facade point search is achieved by building a KD tree in the refined facade point cloud and searching for building facade points near the boundary points of the top surface point cloud with a radius of 1m.

[0030] In another embodiment of the present invention, the method for extracting boundary points using alpha-shape is as follows:

[0031] ① For any point p(x, y), with a rolling circle radius α, search for the point set Q of all points within a distance of 2α from point p in the point cloud;

[0032] Select a point p1(x1, y1) in Q. Calculate the coordinates of the circle center using these two points and the radius α. The center points p2(x2, y2) and p3(x3, y3) are the center coordinates of two different cases passing through points p and p1 with a radius of α. The coordinate formulas are shown below:

[0033]

[0034]

[0035]

[0036]

[0037] in S 2 =(x-x1) 2 +(y-y1) 2

[0038] ② After removing point p1 from the point set Q, calculate the distances from the remaining points to points p2 and p3. If the distance from all points to these two points is greater than the radius α, then p is a boundary point. If the distance from the remaining points to p2 or p3 is not all greater than α, then traverse all points in the point set Q and treat them as points p in turn. If there is a point that satisfies the above two conditions, then the point is a boundary point, the judgment of the point is terminated, and the judgment of the next point is performed. If there are no points like p1 among all the neighboring points in Q, then p is a non-boundary point.

[0039] In another embodiment of the present invention, the method for linear fitting using the RANSAC method is as follows:

[0040] a. Randomly select two points, and use these two points to determine a straight line l;

[0041] b. Set a distance threshold t, determine the dataset S(l) whose geometric distance to line l is less than t, and call it the consistent set of line l;

[0042] c. Repeat the random selection several times to obtain lines l1, l2, ..., l n and the corresponding uniform sets S(l1), S(l2), ..., S(l n );

[0043] d. Use geometric distance to find the best-fit line of the largest consistent set, and use it as the best-matching line for the data points.

[0044] The beneficial effects of this invention are as follows: The progressive optimization method for building contours based on multi-dimensional features of laser point clouds adopts a layer-by-layer optimization and coarse-to-fine extraction of building contours. First, the building point cloud is segmented into building facade point clouds (including trees and building walls, etc.) and top surface point clouds (including building tops and trees, etc.). Then, according to the characteristics of the facade point clouds and top surface point clouds, different features are used to refine them. This allows for the rapid and accurate identification of building point clouds with different heights and structural complexities. While refining the extraction of contour point clouds, it can eliminate the interference of non-building point clouds and the influence of missing facade point clouds, thereby greatly improving the efficiency and accuracy of contour extraction.

[0045] In addition, this optimization method improves the recognition accuracy of building facade point clouds by combining the characteristics of high projection density and large height difference of building facade point clouds on the XOY plane.

[0046] Secondly, the flatness ratio is introduced into the clustering process of point clouds on building rooftops to determine the category of clustered objects, thereby improving the accuracy of point cloud recognition on building rooftops. Attached Figure Description

[0047] Figure 1 This is a flowchart of the progressive optimization method for building outlines based on multi-dimensional features using laser point clouds, as described in this invention.

[0048] Figure 2 This is a flowchart of the coarse segmentation of the point cloud of the building facade and the point cloud of the roof in this invention;

[0049] Figure 3 This is a flowchart of the refinement and contour extraction process for the point cloud of the building facade and the point cloud of the roof in this invention;

[0050] Figure 4 This is a flowchart of Euclidean clustering in an embodiment of the present invention;

[0051] Figure 5 This is a schematic diagram of boundary point extraction using the alpha-shape method in an embodiment of the present invention;

[0052] Figure 6 This is a schematic diagram of the linear fitting process of the RANSAC method in an embodiment of the present invention;

[0053] Figure 7 This is an intensity display diagram of the original point cloud in an embodiment of the present invention;

[0054] Figure 8 This is a height map showing the normal vector calculation results of the point cloud in an embodiment of the present invention;

[0055] Figure 9 This is a point cloud height map of the building facade after coarse segmentation according to an embodiment of the present invention;

[0056] Figure 10This is a point cloud height map of the building's top surface after coarse segmentation according to an embodiment of the present invention;

[0057] Figure 11 This is a refined point cloud extraction result of the building facade according to an embodiment of the present invention;

[0058] Figure 12 This is a refined point cloud extraction result image of the building roof surface according to an embodiment of the present invention;

[0059] Figure 13 This is a diagram showing the combined result of the facade and top surfaces of a building after the facade points have been completed, according to an embodiment of the present invention.

[0060] Figure 14 This is an image showing the result of extracting the outline of a complex rectangular building based on Alpha Shapes and RANSAC according to an embodiment of the present invention.

[0061] Figure 15 This is an image showing the result of extracting the outline of a circular building based on Alpha Shapes and RANSAC according to an embodiment of the present invention.

[0062] Figure 16 This is a diagram showing the result of extracting the outline of a complex building structure based on Alpha Shapes and RANSAC according to an embodiment of the present invention. Detailed Implementation

[0063] To facilitate understanding of the present invention, a more detailed description is provided below with reference to the accompanying drawings and specific embodiments. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0064] It should be noted that, unless otherwise defined, all technical and scientific terms used in this specification have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. The term "and / or" as used in this specification includes any and all combinations of one or more of the associated listed items.

[0065] like Figure 1 The flowchart shown is a progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to the present invention. As can be seen from the figure, the method includes the following steps:

[0066] (1) Separate non-facade point cloud: Filter the original airborne LiDAR point cloud data to separate ground points from non-ground points, thereby obtaining the point cloud of non-ground points;

[0067] (2) Preprocessing: Calculate the relative height of each point in the non-ground point cloud and remove points whose relative height is less than the set elevation threshold.

[0068] (3) Coarse segmentation: Calculate the normal vector nz of each point in the non-ground point cloud after the removal process, and use the normal vector threshold to perform preliminary segmentation of the building facade point cloud and the top point cloud, thereby achieving coarse segmentation of the building facade point cloud and the top point cloud.

[0069] (4) Refined extraction of facade point cloud: The coarsely segmented building facade point cloud is projected in two dimensions on the XOY plane. The projection range is determined with the corresponding projection point as the center and R2 as the radius. The projection density of each point is calculated. At the same time, the height difference of the point cloud is calculated within the projection range of each point, so as to obtain the density and height difference value of each point. Then, the density threshold and height difference threshold are used to refine the extraction of facade point cloud to remove the interference of tree point cloud.

[0070] (5) Top surface point cloud clustering: While extracting the point cloud of the building facade in detail, the top surface point cloud is first filtered and denoised, and then each point in the processed top surface point cloud is clustered to obtain multiple cluster objects.

[0071] (6) Refined extraction of top surface point cloud: Calculate the planarity of each point in the top surface point cloud and the projected area of ​​the clustered objects of the top surface point cloud on the XOY plane. Use the projection area threshold, planarity threshold and the planarity point cloud number ratio threshold to perform refined extraction of the top surface point cloud, so as to obtain the top surface point cloud of the building without interference from other categories.

[0072] (7) Top surface point cloud completion: Extract the boundary points of the refined top surface point cloud, then search for building facade points near the boundary points of the top surface point cloud, and merge the searched facade points into the top surface point cloud, thereby completing the extraction of the point cloud of the complete building top surface and part of the side surface.

[0073] (8) Building outline extraction: The point cloud obtained in the previous step is projected into the XOY plane in two dimensions. Then, the boundary points are extracted and merged using the alpha-shape method. Then, the random sampling consistency method RANSAC is used for line fitting, regularization and merging to finally obtain the two-dimensional outline of the building.

[0074] The optimization method of this invention mainly consists of two parts. First, coarse segmentation of the point cloud of the building facade and the point cloud of the roof is performed. The specific flowchart is as follows. Figure 2 As shown, the point clouds of the building's top surface and facade are then refined and their outlines extracted after coarse segmentation. The specific process is as follows: Figure 3 As shown.

[0075] In the above step (1), filtering the original airborne LiDAR point cloud data is to separate ground points and non-ground points from the point cloud, which is an important prerequisite for the building extraction step. There are various filtering methods here, such as mathematical morphology filtering, iterative densified triangular network filtering, multi-scale virtual grid filtering, etc. In this embodiment, the cloth filtering method is preferably used. As Figure 7 shown is the intensity display diagram of the original point cloud of the embodiment of the present invention. In addition to buildings, it also includes interference factors such as trees and dense vegetation.

[0076] In step (2), based on the separated non-ground point cloud, preprocessing is performed to remove some low-lying ground objects. Generally speaking, vegetation, low shrubs, and low protruding objects on the ground (such as cars, crowds, etc.) usually do not have a very high height, and most of their elevations are below 3m. Compared with buildings, the relative height is higher. Therefore, an elevation threshold of 3m is set here, which can effectively remove these low-lying ground object points.

[0077] After removing the low-lying ground object points, the ground object points mainly include buildings and some tall trees, etc. In order to be able to distinguish the wall points and top surface points of the building for refined processing, improve the accuracy of contour extraction, and avoid the missing of wall points, the present invention divides the building point cloud into building facade point cloud (including trees and building walls, etc.) and top surface point cloud (including building tops and trees, etc.), and then according to the respective characteristics of the facade point cloud and the top surface point cloud, different features are used to perform refined processing on them respectively, so as to improve the extraction accuracy.

[0078] In step (3), the present invention performs a rough segmentation on the preprocessed non-ground point cloud, mainly by comparing the normal vectors. First, the normal vector nz of each point in the non-ground point cloud is calculated. When calculating, a calculation radius R1 is selected. Here, the value of R1 is preferably 5.2m. The normal vector thresholds in this embodiment include a high limit value and a low limit value. Among them, the low limit value is preferably -0.3, and the high limit value is preferably 0.3. Compare the calculated normal vector value nz of each point with the set normal vector threshold. If -0.3 < nz < 0.3, then this point is a facade point, otherwise it is a top surface point. As Figure 8 shown is the height diagram of the normal vector calculation result of the building point cloud of the embodiment of the present invention, Figure 9 is the height diagram of the building facade point cloud result after rough segmentation, Figure 10 is the height diagram of the building top surface point cloud result after rough segmentation. At this time, the non-ground point cloud of the building is roughly segmented into two parts, one part is the facade point cloud (including trees and building walls, etc.), and the other part is the top surface point cloud (including building tops and trees, etc.).

[0079] After coarsely segmenting the top and facade point clouds, they need to be refined for extraction. In step (4), the facade point cloud is first refined for extraction, mainly by using projection density and elevation difference as dual features for judgment. First, the coarsely segmented facade point cloud is projected onto the XOY plane in two dimensions. The projection range is determined by taking the corresponding projection point as the center and R2 as the radius. The projection density of each point is calculated, that is, the number of points within the radius is counted. In this embodiment, the value of R2 is preferably 0.1m. At the same time, the elevation difference of the point cloud is calculated within the projection radius of each point, that is, the maximum and minimum elevation values ​​of all points within the projection radius are counted, and then the difference is calculated as the elevation difference of the point. After obtaining the density and elevation difference of each point, the density threshold and elevation difference threshold are compared. In this embodiment, the density threshold is preferably 3 and the elevation difference threshold is preferably 5m. Points that simultaneously satisfy the conditions of projection density > 3 and elevation difference > 5m are facade points that meet the conditions and are retained; points that do not simultaneously meet the above conditions are interference points and are discarded. Through the above operations, refined extraction of facade point clouds was achieved to remove interference from tree point clouds, such as... Figure 11 The image shown is a refined extraction result of the point cloud on the building facade according to an embodiment of the present invention.

[0080] Furthermore, to improve computational efficiency, this invention employs a KD-tree method when calculating projection density. A KD-tree is a binary tree that stores k-dimensional data. Constructing a KD-tree on a k-dimensional dataset represents a partition of the k-dimensional space formed by that dataset; each node in the tree corresponds to a k-dimensional hyperrectangular region. The KD-tree uses a divide-and-conquer approach, dividing the entire space into several smaller parts. This is a recursive process; by repeating the root node process on the data in the left and right subspaces, we obtain first-level child nodes, further subdividing the space and dataset. This process is repeated until the space contains only a single data point.

[0081] The process of the KD-tree method is as follows:

[0082] I. Calculate the variance of each dimension in the k-dimensional dataset, and select the dimension k with the largest variance:

[0083]

[0084] II. Arrange the data along dimension k in ascending order to obtain the dataset. in N k Given the number of data points in dimension k, calculate the median m in dimension k:

[0085]

[0086] III. Set the threshold to the median m obtained in the previous step to obtain two sets K. sub_low and K sub_high And create a tree node for storage, the set satisfies the requirements of the following equation (3):

[0087]

[0088] IV. Repeat the above steps until all subsets can no longer be divided; if a subset can no longer be divided, save the data in that subset to a leaf node with no child nodes.

[0089] After establishing the KD-tree data structure, it is generally only necessary to search for neighboring points in its child nodes and parent nodes, which can greatly reduce the extra computation caused by searching for neighboring points and improve search efficiency.

[0090] While refining the point cloud of the building facade, the point cloud of the top surface can be processed simultaneously. In step (5), the top surface point cloud is first filtered and denoised, and then each point in the processed top surface point cloud is clustered to obtain multiple clustered objects. The clustering method in this embodiment of the invention adopts Euclidean clustering or region growing clustering.

[0091] Euclidean clustering is a clustering method based on the Euclidean distance metric. In three-dimensional space, the Euclidean distance between points (x1, y1, z1) and (x2, y2, z2) is defined as d. E d E The calculation formula is: Nearest neighbor lookup based on KD-trees is a prerequisite for accelerating the Euclidean clustering process, such as... Figure 4 The diagram shows a flowchart of Euclidean clustering according to an embodiment of the present invention. As can be seen from the diagram, the Euclidean clustering method is as follows: For a point P in space, the k nearest neighbor search methods of KD tree are used to find the k points closest to point P. Among these points, those whose distance is less than a set threshold are clustered into set Q'. If the number of elements in Q' no longer increases, the entire clustering process ends; otherwise, points other than point P in set Q' must be selected and the above process is repeated until the number of elements in Q' no longer increases.

[0092] Furthermore, the region-growing clustering method is a conventional method in this field, and will not be described in detail here.

[0093] In step (6), after clustering the top surface point cloud, a refined extraction of the point cloud is performed. The planarity of each point in the top surface point cloud and the projected area of ​​the clustered objects on the XOY plane are calculated. In this embodiment, the preferred threshold for projected area is 30㎡, the preferred threshold for planarity is 0.5, and the threshold for the proportion of planarity points is 30%. The calculated projected area and planarity of each clustered object are compared with the corresponding thresholds. Objects that simultaneously meet the conditions of projected area > 30㎡ and (proportion of points with planarity > 0.5) > 30% are considered qualified clustered objects and are retained; points that do not simultaneously meet the above conditions are considered interfering clustered objects and are discarded. After the above refined extraction, a top surface point cloud of the building without interference from other categories is obtained. Figure 12 The image shown is a refined extraction result of point cloud data from the rooftop of some buildings according to an embodiment of the present invention.

[0094] The flatness calculation of point clouds in this invention uses covariance features, and the calculation method is a conventional technique in this field, which will not be described in detail here.

[0095] In addition, when refining the point cloud extraction of the top surface, besides using the dual features of projected area and flatness, the projected area can also be replaced by the point count feature, that is, the number of points in the clustered object. The extraction is judged by the dual features of point count and flatness. The point count threshold here is 200 to 300,000.

[0096] After identifying interference points in the top surface point cloud, boundary point extraction can be performed. In step (7), the boundary points of the refined top surface point cloud are first extracted, and then the building facade points are searched near the boundary points of the top surface point cloud. The searched facade points are then merged into the top surface point cloud, thus completing the extraction of the point cloud of the complete building top surface and part of the side surfaces. Figure 13 The image shown is a diagram of the combined facade and top surface points of a building according to an embodiment of the present invention.

[0097] Preferably, the alpha-shape method is used to extract boundary points from the top point cloud. The Alpha Shapes method is a simple, effective, and fast method for extracting boundary points, unaffected by the shape of the point cloud boundary points. It can quickly and accurately extract boundary points. Its principle is as follows: For a planar point cloud of arbitrary shape, if a circle of radius α is rolled around it, if the radius α is sufficiently small, then every point in the point cloud is a boundary point; if it is appropriately increased to a certain extent, it only rolls around the boundary points, and its rolling trajectory is the point cloud boundary. For example... Figure 5 The diagram shown illustrates the boundary point extraction method of the Alpha Shapes method in this invention. The method for boundary point extraction using alpha-shape is as follows:

[0098] ① For any point p(x, y), with a rolling circle radius α, search for the point set Q of all points within a distance of 2α from point p in the point cloud;

[0099] Select a point p1(x1, y1) in Q. Calculate the coordinates of the circle center based on the coordinates of these two points and the radius α. The center coordinates p2(x2, y2) and p3(x3, y3) are the center coordinates of two different cases passing through points p and p1 with a radius of α. The coordinate formulas are shown below:

[0100]

[0101]

[0102]

[0103]

[0104] in S 2 =(x-x1) 2 +(y-y1) 2

[0105] ② After removing point p1 from the point set Q, calculate the distances from the remaining points to points p2 and p3. If the distance from all points to these two points is greater than the radius α, then p is a boundary point. If the distance from the remaining points to p2 or p3 is not all greater than α, then traverse all points in the point set Q and treat them as points p in turn. If there is a point that satisfies the above two conditions, then the point is a boundary point, the judgment of the point is terminated, and the judgment of the next point is performed. If there are no points like p1 among all the neighboring points in Q, then p is a non-boundary point.

[0106] Furthermore, to improve search efficiency, the building facade point search is achieved by building a KD tree in the refined facade point cloud and searching for building facade points near the boundary points of the top surface point cloud with a radius of 1m.

[0107] In step (8), after completing the top surface point cloud, a two-dimensional projection of the completed point cloud onto the XOY plane is performed. Then, the alpha-shape method is used to extract and merge boundary points. Finally, the RANSAC method is used for line fitting, regularization, and merging to obtain the two-dimensional outline of the building. Figure 14-16 This is an image showing the contour extraction results of rectangular complex buildings, circular buildings, and complex structure buildings based on Alpha Shapes and RANSAC in an embodiment of the present invention.

[0108] The RANSAC method iteratively draws random samples from the model to find better model parameters that contain more supporting interior points. The model's remainder is then used to validate the drawn samples. After a certain number of iterations, the sample set with the highest probability of being close to a reasonable solution is considered the closest reasonable solution. Finally, the remainder of the samples is used to support the correctness of the obtained parameter solution. The method for line fitting using RANSAC is as follows:

[0109] a. Randomly select two points, and use these two points to determine a straight line l;

[0110] b. Set a distance threshold t, determine the dataset S(l) whose geometric distance to line l is less than t, and call it the consistent set of line l;

[0111] c. Repeat the random selection several times to obtain lines l1, l2, ..., l n and the corresponding uniform sets S(l1), S(l2), ..., S(l n );

[0112] d. Use geometric distance to find the best-fit line of the largest consistent set, and use it as the best-matching line for the data points.

[0113] like Figure 6 The diagram illustrates the line fitting process of the RANSAC method in this embodiment of the invention. First, two points are randomly selected from the point set, and the parameters of the line formed by them are solved. Then, the distances from the remaining points in the point set to the line are calculated. Points whose distances are less than a set distance threshold are considered interior points, and the number of interior points is counted. Next, two more points are randomly selected, and the number of interior points is counted again, and so on. The point set with the most interior points is the maximum consistent set. Finally, a line is fitted from the points in the maximum consistent set using least squares.

[0114] The progressive optimization method for building outlines of the present invention can quickly and accurately extract the complete building outline while effectively avoiding interference from tree point clouds and missing facade point clouds.

[0115] The above description is merely an embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are included within the patent protection scope of the present invention.

Claims

1. A progressive optimization method for building contours based on multi-dimensional features using laser point clouds, characterized in that, The method includes the following steps: (1) Separating non-vertical point clouds: Filter the original airborne LiDAR point cloud data to separate ground points from non-ground points, thereby obtaining the point cloud of non-ground points; (2) Preprocessing: Calculate the relative height of each point in the non-ground point cloud, and remove the points with a relative height less than the set elevation threshold; (3) Coarse segmentation: Calculate the normal vector nz of each point in the non-ground point cloud after the removal process, and use the normal vector threshold to perform a preliminary segmentation of the building facade point cloud and the top surface point cloud, thereby achieving the coarse segmentation of the building facade point cloud and the top surface point cloud; (4) Fine extraction of the facade point cloud: Perform a two-dimensional projection of the coarsely segmented building facade point cloud on the XOY plane, determine the projection range with the corresponding projection point as the center and R2 as the radius, calculate the projection density of each point, and at the same time calculate the height difference of the point cloud within the projection range of each point, thereby obtaining the density and height difference values of each point, and then use the density threshold and the height difference threshold to perform a fine extraction of the facade point cloud to remove the interference of the tree point cloud; (5) Clustering of the top surface point cloud: While performing a fine extraction of the building facade point cloud, first filter and denoise the top surface point cloud, and then cluster each point in the processed top surface point cloud to obtain multiple clustering objects; (6) Fine extraction of the top surface point cloud: Calculate the planarity of each point in the top surface point cloud and the projection area of the clustering object of the top surface point cloud on the XOY plane, and use the projection area threshold, the planarity threshold, and the proportion threshold of the number of flatness point clouds to perform a fine extraction of the top surface point cloud, and remove the clustering objects that do not meet the conditions, thereby obtaining the top surface point cloud of the building without interference from other categories; (7) Completing the top surface point cloud: Extract the boundary points of the finely processed top surface point cloud, then search for building facade points near the boundary points of the top surface point cloud, and merge the searched facade points into the top surface point cloud, thereby completing the extraction of the point cloud of the complete building top surface and part of the side surface; (8) Extracting the building contour: Perform a two-dimensional projection of the point cloud obtained in the previous step on the XOY plane, then use the alpha-shape method to extract and merge the boundary points, and then use the RANSAC method for line fitting, regularization, and merging processing, and finally obtain the two-dimensional contour of the building.

2. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 1, characterized in that, In step (3), the calculation radius selected when calculating the normal vector of each point is R1, and the value of R1 is 5.2 m. The normal vector threshold includes a high limit value and a low limit value, where the low limit value is -0.3 and the high limit value is 0.

3. That is, the points with a normal vector satisfying -0.3 < nz < 0.3 are facade points, otherwise they are top surface points.

3. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 2, characterized in that, In step (4), when calculating the projection density of each point, it is through the method of establishing a KD tree, and the value of R^2 is 0.1 m; the projection density is the number of points within the projection range, the density threshold is 3, and the height difference threshold is 5 m. The points that simultaneously satisfy the projection density > 3 and the height difference value > 5 m are the qualified facade points and are retained; the points that do not simultaneously meet the above conditions are interference points and are removed.

4. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 2 or 3, characterized in that, The process of the KD tree method is as follows: I. Calculate the variance of each dimension in the k-dimensional dataset, and select the dimension k with the largest variance: II. Arrange the data along dimension k in ascending order to obtain the dataset. in N k Given the number of data points in dimension k, calculate the median m in dimension k: III. Set the threshold to the median m obtained in the previous step to obtain two sets K. sub_low and K sub_high And create a tree node for storage, the set satisfies the requirements of the following equation (3): IV. Repeat the above steps until all subsets can no longer be partitioned; If a subset cannot be further divided, the data in that subset is saved to a leaf node that has no child nodes.

5. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 4, characterized in that, Step (5) The top surface point cloud clustering method adopts Euclidean clustering or region growing clustering.

6. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 5, characterized in that, The Euclidean clustering method is as follows: For a point P in space, the k nearest neighbor search methods of KD tree are used to find the k points closest to point P. Among these points, those whose distance is less than a set threshold are clustered into set Q'. If the number of elements in Q' no longer increases, the entire clustering process ends; otherwise, the above process must be repeated by selecting points other than point P in set Q' until the number of elements in Q' no longer increases.

7. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 6, characterized in that, In step (6), the projection area threshold is 30㎡, the flatness threshold is 0.5, and the flatness point cloud number ratio threshold is 30%. That is, objects that simultaneously meet the conditions of projection area > 30㎡ and (the ratio of points with flatness > 0.5) > 30% are cluster objects that meet the conditions and are retained; objects that do not simultaneously meet the above conditions are interfering cluster objects and are removed.

8. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 7, characterized in that, In step (7), the alpha-shape method is used to extract the boundary points of the top surface point cloud; the building facade point search is achieved by building a KD tree in the refined facade point cloud and searching for building facade points near the boundary points of the top surface point cloud with a radius of 1m.

9. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 8, characterized in that, The method for extracting boundary points using alpha-shape is as follows: ① For any point p(x,y), with a rolling circle radius α, search for the point set Q of all points within a distance of 2α from point p in the point cloud; Select a point p1(x1,y1) in Q. Calculate the coordinates of the circle center using these two points and the radius α. The center points p2(x2,y2) and p3(x3,y3) are the center coordinates of two different cases passing through points p and p1 with a radius of α. The coordinate formulas are shown below: in S 2 =(x-x1) 2 +(y-y1) ② After removing point p1 from the set of points Q, calculate the distances from the remaining points to points p2 and p3. If the distance from all points to these two points is greater than the radius α, then p is a boundary point. If the distance from the remaining points to points p2 or p3 is not all greater than α, then traverse all points in the set of points Q and take turns as point p. If there is a point that satisfies the above two conditions, then the point is a boundary point, the judgment of the point is terminated, and the judgment of the next point is performed. If there are no points like p1 among all the neighboring points in Q, then p is a non-boundary point.

10. The progressive optimization method for building contours based on multi-dimensional features using laser point clouds according to claim 9, characterized in that, The method for linear fitting using the RANSAC method is as follows: a. Randomly select two points, and use these two points to determine a straight line l; b. Set a distance threshold t, determine the dataset S(l) whose geometric distance to line l is less than t, and call it the consistent set of line l; c. Repeat the random selection several times to obtain lines l1, l2, ..., l n and the corresponding uniform sets S(l1), S(l2), ..., S(l n ); d. Use geometric distance to find the best-fit line of the largest consistent set, and use it as the best-matching line for the data points.

Images