Outdoor scene building extraction method based on multi-feature constraints

Abstract

The present invention discloses a method for extracting buildings from outdoor scenes based on multi-feature constraints. First, a voxel grid filtering method is used to downsample a large-scale point cloud scene, and a cloth filtering algorithm is used to remove ground points from the downsampled scene. Then, a density-based clustering algorithm and a KD tree are combined to cluster non-ground points into different point clusters. Finally, the building point cloud is separated from the non-building point cloud by combining the point cloud height, volume, direction, dimension and color features, and finally the building point cloud is extracted from the outdoor scene. The present invention effectively improves the problem of low extraction accuracy using traditional methods, and can successfully extract most buildings in outdoor scenes, providing a corresponding processing idea for the extraction of buildings in large-scale scenes.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud scene segmentation and relates to a method for extracting buildings from outdoor scenes based on multi-feature constraints. Background Art

[0002] The ultimate goal of digital cities is to transform and describe extracted building information into digital models that computers can understand. While significant progress has been made in the reconstruction of small-scale indoor scenes and small objects, the reconstruction of objects in large-scale outdoor scenes still presents many challenges that need to be addressed.

[0003] Building extraction, as the premise and foundation of building reconstruction, refers to the task of dividing a given dataset into non-overlapping homogeneous regions and identifying buildings from these regions, so that the buildings will not be affected by other factors during the reconstruction process. Building extraction needs to fully consider the characteristics of building objects that distinguish them from other categories of objects, so as to separate buildings from non-buildings. For a long time, large-scale and complex outdoor scenes have contained a large number of different types of land objects, such as vegetation, vehicles, trees of similar height to buildings, trees that are very close to buildings in spatial distance or even connected to buildings, etc., making it difficult to automatically extract buildings. Traditional extraction methods have problems with poor extraction accuracy and low efficiency, which will further affect the quality of building reconstruction models. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for extracting buildings from outdoor scenes based on multiple feature constraints, which effectively improves the problem of low extraction accuracy existing in traditional methods.

[0005] The technical solution adopted by the present invention is a method for extracting buildings from outdoor scenes based on multiple feature constraints, which is specifically implemented in the following steps:

[0006] Step 1, data preprocessing: input the original point cloud data, use the voxel grid filtering method to downsample the original scene to obtain the preprocessed point cloud data;

[0007] Step 2, ground point filtering: Use the cloth filtering algorithm to filter out the ground points in the scene;

[0008] Step 3, clustering of non-ground points: construct a KD (K-Dimensional) tree, use the index method to find neighboring points, and implement clustering of non-ground points through a density-based clustering algorithm;

[0009] Step 4, building point extraction: Consider the five aspects of point cloud height features, point cloud volume features, direction features, dimension features and color features to distinguish between buildings and non-buildings and complete building point extraction.

[0010] The present invention is also characterized in that:

[0011] Step 1 is as follows:

[0012] Step 1.1, calculate the center point of each voxel grid

[0013] Draw a voxel grid on the original point cloud data, assuming the current input point is , the voxel grid side length , then the center point coordinates of the voxel grid for,

[0014] (1)

[0015] The position of each voxel grid in space is obtained according to the side length of the voxel grid and the coordinates of the center point of the voxel grid;

[0016] Step 1.2, calculate the centroid of each voxel grid

[0017] Assume that the center of gravity of the grid is , then for The centroid of each voxel grid is represented as:

[0018] (2)

[0019] In step 1.3, the centroid of each voxel grid replaces all points in the corresponding voxel grid, and the remaining points other than the centroid are filtered out to obtain the preprocessed point cloud data.

[0020] Step 2 is as follows:

[0021] Step 2.1, invert the pre-processed point cloud data, set the initial cloth mesh side length, and construct the initial cloth mesh, assuming that particles only move in the vertical direction;

[0022] Step 2.2, project the grid particles and the point cloud to the same horizontal plane, find the neighboring points of each particle in the grid and record the elevation value HIV of the corresponding point;

[0023] Step 2.3: For each movable particle, calculate the distance moved by the particle after being affected by gravity. The calculation formula is as shown in formula (3);

[0024] (3)

[0025] In formula (3), Indicates that the particles The location at the moment, represents the time step, represents the gravitational constant, represents the mass of the particle;

[0026] Step 2.4: Compare the position of the particle after movement with the elevation value HIV of its corresponding neighboring point: If the particle's height is less than or equal to the elevation value HIV of the neighboring point, move the particle to the HIV height and set it as an immovable point.

[0027] Step 2.5, for each particle, calculate the movement distance caused by the internal driving factor; if the two connected particles are both movable particles, move the two particles to the average elevation value of the two particles; if one is an immovable point and the other is a movable particle, move the movable particle; if both particles are immovable points, do not move;

[0028] The distance generated by particle movement is calculated by formula (4);

[0029] (4)

[0030] In formula (4), represents the displacement vector of the particle, If it is a movable particle, the value is 1, otherwise it is 0. represents the position of the movable particle, Indicates the position of the current particle adjacent to the particle. represents the unit vector in the vertical direction;

[0031] Step 2.6, iterate steps 2.3 to 2.5 until the maximum elevation change of all particles is less than the threshold Or when the number of iterations reaches the preset value, the simulation process is stopped;

[0032] Step 2.7: Calculate the height difference between the point cloud and the particle. If the distance between the point in the point cloud and the particle is less than the height threshold, the point in the point cloud is considered to be a ground point, otherwise it is considered to be a non-ground point.

[0033] Step 3 is as follows:

[0034] Step 3.1, construct the KD tree;

[0035] Step 3.2, select an unprocessed point from the point cloud , unprocessed points refer to points that have not been marked as processed, and are searched according to the index Points in the neighborhood, if the neighborhood contains at least Point, proof As a core point, the point is marked as processed and a new cluster is created ,Will of Neighborhood points added to Otherwise, the point is marked as processed and a new point is selected until the core point is selected;

[0036] Step 3.3, traverse the cluster For the unprocessed points in the traversal process, if there is a point that is also a core point under a given threshold, then the point All points in the neighborhood are added to the cluster and mark the point as processed;

[0037] Step 3.4, repeat step 3.3 until no new points are added to the cluster ;

[0038] Step 3.5: Repeat steps 3.2 to 3.4 until all points in the point cloud are processed, and clustering is completed.

[0039] Step 4 is as follows:

[0040] Step 4.1: Use the elevation value to remove low vegetation in the scene: point clusters with elevation values less than 2m are low objects and need to be removed, while point clusters above 2m are tall objects such as buildings and trees. These point clusters are retained for the next calculation. The elevation value calculation is as shown in formula (5);

[0041] (5)

[0042] In formula (5), Indicates the elevation value, Indicates the maximum value of the point cloud in the direction perpendicular to the ground. Represents point cloud ground points Coordinates, using ground points The mean value of the coordinates is used instead;

[0043] Step 4.2: Use the minimum convex hull volume to eliminate objects with a small area and a height higher than low vegetation but not higher than buildings in the scene: solve the minimum convex hull of the clustered point cluster, and further calculate the volume of the minimum convex hull to distinguish buildings from other objects. The volume threshold According to the volume of buildings in different scenes;

[0044] Step 4.3, use directional features to remove tall trees in the scene: construct and analyze the point cloud normal vector and the Z-axis direction vector The cosine value of the angle is used to determine the category to which the current point cluster belongs;

[0045] Step 4.4: Use dimensional features to remove linear and spherical objects from the scene, and analyze the point cloud using principal component analysis.

[0046] In step 4.5, use color features to remove sparsely populated trees and trees connected to buildings in the scene.

[0047] Step 4.3 is as follows:

[0048] Step 4.3.1: First, divide the cosine value of the angle between the point cloud normal vector and the Z-axis direction vector into multiple intervals. The interval size is in the range of [0,1], and set the group distance. , then construct the distribution histogram of the angle cosine value;

[0049] Step 4.3.2, calculate the probability of each interval in the cosine value distribution of the angle between the point cloud normal vector and the Z-axis direction vector and the standard deviation of all probabilities, and count the number of intervals with a probability greater than the standard deviation The number of intervals whose sum probability is less than the standard deviation ,if , proving that the cosine value distribution of this point cluster is relatively discrete and it is not a building, so it is eliminated.

[0050] Step 4.4 is as follows:

[0051] Step 4.4.1, calculate the eigenvalues of the covariance matrix of each cluster of points ;

[0052] Step 4.4.2, let , use formula (6) to calculate the probability that a point in the space belongs to the three dimensions respectively;

[0053] (6)

[0054] In formula (6), represents linearity, that is, the probability that the point cloud belongs to a one-dimensional feature, represents the flatness, that is, the probability that the point cloud belongs to a two-dimensional feature, Represents the degree of scatter, that is, the probability that the point cloud belongs to a three-dimensional feature;

[0055] In step 4.4.3, the dimensional features of the current point are obtained according to the size of the probability values, and the category to which the point belongs is further obtained. If the proportion of surface points in a point cluster is the highest, it is determined to belong to a building. On the contrary, if the proportion of scattered points or line points is higher than that of surface points, it is determined to be a non-building, thereby extracting all building point clouds in the scene.

[0056] Step 4.5 is as follows:

[0057] In step 4.5.1, the excess green index EXG (Excess Green) is introduced into the 3D point cloud scene as the basis for removing sparse trees. Its calculation formula is as shown in formula (7);

[0058] (7)

[0059] In formula (7), , Respectively represent the color values of the point cloud;

[0060] Step 4.5.2, if Index greater than threshold , then judge the point All points in the neighborhood Index and count its number ,if Greater than the point threshold , then the point is a tree point, otherwise it is a building point, and the threshold Set to 0.08, threshold The adaptive one is set to half of the number of point clouds in the neighborhood, and the building points are retained to complete the building point extraction.

[0061] The beneficial effects of the present invention are:

[0062] The present invention is based on a multi-feature constrained outdoor scene building extraction method. The ground points of the downsampled scene are filtered out through a cloth filtering algorithm. The method has stronger applicability and can be applied to a variety of terrains with complex features. After clustering non-ground points, features with strong distinguishing ability are selected for building extraction. It can effectively eliminate low and small-area objects, and also has a good elimination effect on sparsely dense trees and trees connected to buildings. It effectively improves the problem of low extraction accuracy using traditional methods, and can successfully extract most buildings in outdoor scenes, providing a corresponding processing idea for building extraction in large-scale scenes. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a block diagram of an outdoor scene building extraction method based on multi-feature constraints;

[0064] Figure 2(a)-Figure 2(b) It is the result of filtering out ground points using a cloth filtering algorithm using a multi-feature-constrained outdoor scene building extraction method.

[0065] Figure 3 It is the result of clustering non-ground points using a multi-feature-constrained outdoor scene building extraction method combined with a density clustering algorithm and a KD tree;

[0066] Figure 4(a)-Figure 4(b) It is the result of constructing the minimum convex hull of the outdoor scene building extraction method based on multiple feature constraints;

[0067] Figure 5(a)-Figure 5(d) The outdoor scene building extraction method based on multi-feature constraints is the distribution histogram of the cosine value of the angle between different point clusters;

[0068] Figure 6 It is a histogram of the point category distribution in different point clusters of the outdoor scene building extraction method based on multi-feature constraints;

[0069] Figure 7(a)-Figure 7(b) It is the result of building point extraction based on the outdoor scene building extraction method with multi-feature constraints. DETAILED DESCRIPTION

[0070] The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] Example

[0072] This embodiment provides a method for extracting buildings from outdoor scenes based on multiple feature constraints, such as Figure 1 As shown, the specific steps are as follows:

[0073] Step 1, data preprocessing: input the original point cloud data, use the voxel grid filtering method to downsample the original scene to obtain the preprocessed point cloud data;

[0074] Step 1.1, calculate the center point of each voxel grid

[0075] Draw a voxel grid on the original point cloud data, assuming the current input point is , the voxel grid side length , then the center point coordinates of the voxel grid for,

[0076] (1)

[0077] The position of each voxel grid in space is obtained according to the side length of the voxel grid and the coordinates of the center point of the voxel grid;

[0078] Step 1.2, calculate the centroid of each voxel grid

[0079] Assume that the center of gravity of the grid is , then for The centroid of each voxel grid is represented as:

[0080] (2)

[0081] In step 1.3, the centroid of each voxel grid replaces all points in the corresponding voxel grid, and the remaining points other than the centroid are filtered out to obtain the preprocessed point cloud data.

[0082] Step 2, ground point filtering: Use the cloth filtering algorithm to filter out the ground points in the scene;

[0083] Step 2.1, invert the pre-processed point cloud data, set the initial cloth mesh side length, and construct the initial cloth mesh, assuming that particles only move in the vertical direction;

[0084] Step 2.2, project the grid particles and the point cloud to the same horizontal plane, find the neighboring points of each particle in the grid and record the elevation value HIV of the corresponding point;

[0085] Step 2.3: For each movable particle, calculate the distance moved by the particle after being affected by gravity. The calculation formula is as shown in formula (3);

[0086] (3)

[0087] In formula (3), Indicates that the particles The location at the moment, represents the time step, represents the gravitational constant, represents the mass of the particle;

[0088] Step 2.4: Compare the position of the particle after movement with the elevation value HIV of its corresponding neighboring point: If the particle's height is less than or equal to the elevation value HIV of the neighboring point, move the particle to the HIV height and set it as an immovable point.

[0089] Step 2.5, for each particle, calculate the movement distance caused by the internal driving factor; if the two connected particles are both movable particles, move the two particles to the average elevation value of the two particles; if one is an immovable point and the other is a movable particle, move the movable particle; if both particles are immovable points, do not move;

[0090] The distance generated by particle movement is calculated by formula (4);

[0091] (4)

[0092] In formula (4), represents the displacement vector of the particle, If it is a movable particle, the value is 1, otherwise it is 0. represents the position of the movable particle, Indicates the position of the current particle adjacent to the particle. represents the unit vector in the vertical direction;

[0093] Step 2.6, iterate steps 2.3 to 2.5 until the maximum elevation change of all particles is less than the threshold Or when the number of iterations reaches the preset value, the simulation process is stopped;

[0094] Step 2.7: Calculate the height difference between the point cloud and the particle. If the distance between the point in the point cloud and the particle is less than the height threshold, the point in the point cloud is considered to be a ground point. Otherwise, it is considered to be a non-ground point. The results of cloth filtering are shown in Figure 2 (a)-Figure 2 (b).

[0095] Step 3, clustering of non-ground points: construct a KD (K-Dimensional) tree, use the index method to find neighboring points, and implement clustering of non-ground points through a density-based clustering algorithm;

[0096] Step 3.1, construct the KD tree;

[0097] Step 3.2, select an unprocessed point from the point cloud , unprocessed points refer to points that have not been marked as processed, and are searched according to the index Points in the neighborhood, if the neighborhood contains at least Point, proof As a core point, the point is marked as processed and a new cluster is created ,Will of Neighborhood points added to Otherwise, the point is marked as processed and a new point is selected until the core point is selected;

[0098] Step 3.3, traverse the cluster For the unprocessed points in the traversal process, if there is a point that is also a core point under a given threshold, then the point All points in the neighborhood are added to the cluster and mark the point as processed;

[0099] Step 3.4, repeat step 3.3 until no new points are added to the cluster ;

[0100] Step 3.5, repeat steps 3.2 to 3.4 until all points in the point cloud are processed, and clustering is completed. The clustering result is as follows: Figure 3 shown.

[0101] Step 4: Building point extraction: Consider the five aspects of point cloud height, point cloud volume, direction, dimension, and color to distinguish between buildings and non-buildings and complete building point extraction;

[0102] Step 4.1: Use the elevation value to remove low vegetation in the scene: point clusters with elevation values less than 2m are low objects and need to be removed, while point clusters above 2m are tall objects such as buildings and trees. These point clusters are retained for the next calculation. The elevation value calculation is as shown in formula (5);

[0103] (5)

[0104] In formula (5), Indicates the elevation value, Indicates the maximum value of the point cloud in the direction perpendicular to the ground. Represents point cloud ground points Coordinates, using ground points The mean value of the coordinates is used instead;

[0105] Step 4.2, use the minimum convex hull volume to eliminate the objects in the scene that are small in area and higher than low vegetation but not higher than buildings: solve the minimum convex hull of the clustered point cluster. The results of the minimum convex hull construction are shown in Figure 4 (a)-Figure 4 (b). Further calculate the volume of the minimum convex hull to distinguish buildings from other objects. The volume threshold According to the volume of buildings in different scenes;

[0106] Step 4.3, use directional features to remove tall trees in the scene: construct and analyze the point cloud normal vector and the Z-axis direction vector The cosine value of the angle is used to determine the category to which the current point cluster belongs;

[0107] Step 4.3 is as follows:

[0108] Step 4.3.1: First, divide the cosine value of the angle between the point cloud normal vector and the Z-axis direction vector into multiple intervals. The interval size is in the range of [0,1], and set the group distance. , and then construct the distribution histogram of the cosine value of the angle as shown in Figure 5 (a) to Figure 5 (d);

[0109] Step 4.3.2, calculate the probability of each interval in the cosine value distribution of the angle between the point cloud normal vector and the Z-axis direction vector and the standard deviation of all probabilities, and count the number of intervals with a probability greater than the standard deviation The number of intervals whose sum probability is less than the standard deviation ,if , proving that the cosine value distribution of this point cluster is relatively discrete and it is not a building, so it is eliminated.

[0110] Step 4.4: Use dimensional features to remove linear and spherical objects from the scene, and analyze the point cloud using principal component analysis.

[0111] Step 4.4 is as follows:

[0112] Step 4.4.1, calculate the eigenvalues of the covariance matrix of each cluster of points ;

[0113] Step 4.4.2, let , use formula (6) to calculate the probability that a point in the space belongs to the three dimensions respectively;

[0114] (6)

[0115] In formula (6), represents linearity, that is, the probability that the point cloud belongs to a one-dimensional feature, represents the flatness, that is, the probability that the point cloud belongs to a two-dimensional feature, Represents the degree of scatter, that is, the probability that the point cloud belongs to a three-dimensional feature;

[0116] Step 4.4.3, according to the different probability values, the dimensional features of the current point are obtained, and the category to which the point belongs is further obtained. The distribution of point categories in different point clusters is as follows: Figure 6 As shown in the figure, the proportion of surface points in clusters 1 and 2 is the highest, while the proportion of scattered points in the other three clusters is high. Therefore, clusters 1 and 2 are considered to belong to buildings, while clusters 3, 4, and 5 are non-buildings. Thus, all building point clouds in the scene are extracted.

[0117] Step 4.5: Use color features to remove sparsely populated trees and trees connected to buildings in the scene.

[0118] Step 4.5 is as follows:

[0119] In step 4.5.1, the excess green index EXG (Excess Green) is introduced into the 3D point cloud scene as the basis for removing sparse trees. Its calculation formula is as shown in formula (7);

[0120] (7)

[0121] In formula (7), , Respectively represent the color values of the point cloud;

[0122] Step 4.5.2, if Index greater than threshold , then judge the point All points in the neighborhood Index and count its number ,if Greater than the point threshold , then the point is a tree point, otherwise it is a building point, and the threshold Set to 0.08, threshold The adaptive value is set to half of the number of point clouds in the neighborhood, and the building points are retained to complete the building point extraction. The extraction results are shown in Figure 7 (a)-Figure 7 (b).

[0123] Through the above method, it can be seen that the present invention is based on a method for extracting buildings from outdoor scenes with multi-feature constraints. First, a voxel grid filtering method is used to downsample the large-scale point cloud scene, and a cloth filtering algorithm is used to remove ground points from the downsampled scene. Then, the non-ground points are clustered based on a density-based clustering algorithm and a KD tree, and the non-ground points are clustered into different point clusters. Finally, the building point cloud is separated from the non-building point cloud by combining the point cloud height, volume, direction, dimension and color features, and finally the building point cloud is extracted from the outdoor scene. The present invention effectively improves the problem of low extraction accuracy using traditional methods, has a complete technical route, and can successfully extract most buildings in outdoor scenes, providing a corresponding processing idea for the extraction of buildings in large-scale scenes.

Claims

1. A method for extracting buildings from outdoor scenes based on multi-feature constraints, characterized by: The specific implementation steps are as follows: Step 1, data preprocessing: input the original point cloud data, use the voxel grid filtering method to downsample the original scene to obtain the preprocessed point cloud data; Step 2, ground point filtering: Use the cloth filtering algorithm to filter out the ground points in the scene; Step 3, clustering of non-ground points: construct a KD tree, use the index method to find the neighborhood points, and implement clustering of non-ground points through a density-based clustering algorithm; Step 4: Building point extraction: Consider the five aspects of point cloud height, point cloud volume, direction, dimension, and color to distinguish between buildings and non-buildings and complete building point extraction; The step 4 is specifically as follows: Step 4.1: Use the elevation value to remove low vegetation in the scene: point clusters with elevation values less than 2m are low objects and need to be removed, while point clusters with elevation values higher than 2m are high objects, including buildings and trees. These high objects are retained for the next calculation. The elevation value calculation is as shown in formula (5); (5) In formula (5), Indicates the elevation value, Indicates the maximum value of the point cloud in the direction perpendicular to the ground. Represents point cloud ground points Coordinates, using ground points The mean value of the coordinates is used instead; Step 4.2: Use the minimum convex hull volume to eliminate objects with a small area and a height higher than low vegetation but not higher than buildings in the scene: solve the minimum convex hull of the clustered point cluster, and further calculate the volume of the minimum convex hull to distinguish buildings from other objects. The volume threshold According to the volume of buildings in different scenes; Step 4.3, use directional features to remove tall trees in the scene: construct and analyze the point cloud normal vector and the Z-axis direction vector The cosine value of the angle is used to determine the category to which the current point cluster belongs; Step 4.4: Use dimensional features to remove linear and spherical objects from the scene, and analyze the point cloud using principal component analysis. In step 4.5, use color features to remove sparsely populated trees and trees connected to buildings in the scene.

2. The outdoor scene building extraction method based on multi-feature constraints according to claim 1 is characterized in that: The step 1 is specifically as follows: Step 1.1, calculate the center point of each voxel grid Draw a voxel grid on the original point cloud data, assuming the current input point is , the voxel grid side length , then the center point coordinates of the voxel grid for, (1) The position of each voxel grid in space is obtained according to the side length of the voxel grid and the coordinates of the center point of the voxel grid; Step 1.2, calculate the centroid of each voxel grid Assume that the center of gravity of the grid is , then for The centroid of each voxel grid is represented as: (2) In step 1.3, the centroid of each voxel grid replaces all points in the corresponding voxel grid, and the remaining points other than the centroid are filtered out to obtain the preprocessed point cloud data.

3. The outdoor scene building extraction method based on multi-feature constraints according to claim 2 is characterized in that: The step 2 is specifically as follows: Step 2.1, invert the pre-processed point cloud data, set the initial cloth mesh side length, and construct the initial cloth mesh, assuming that particles only move in the vertical direction; Step 2.2, project the grid particles and the point cloud to the same horizontal plane, find the neighboring points of each particle in the grid and record the elevation value HIV of the corresponding point; Step 2.3: For each movable particle, calculate the distance moved by the particle after being affected by gravity. The calculation formula is as shown in formula (3); (3) In formula (3), Indicates that the particles The location at the moment, represents the time step, represents the gravitational constant, represents the mass of the particle; Step 2.4: Compare the position of the particle after movement with the elevation value HIV of its corresponding neighboring point: If the particle's height is less than or equal to the elevation value HIV of the neighboring point, move the particle to the HIV height and set it as an immovable point. Step 2.5, for each particle, calculate the movement distance caused by the internal driving factor; if the two connected particles are both movable particles, move the two particles to the average elevation value of the two particles; if one is an immovable point and the other is a movable particle, move the movable particle; if both particles are immovable points, do not move; The distance generated by particle movement is calculated by formula (4); (4) In formula (4), represents the displacement vector of the particle, If it is a movable particle, the value is 1, otherwise it is 0. represents the position of the movable particle, Indicates the position of the current particle adjacent to the particle. represents the unit vector in the vertical direction; Step 2.6, iterate steps 2.3 to 2.5 until the maximum elevation change of all particles is less than the threshold Or when the number of iterations reaches the preset value, the simulation process is stopped; Step 2.7: Calculate the height difference between the point cloud and the particle. If the distance between the point in the point cloud and the particle is less than the height threshold, the point in the point cloud is considered to be a ground point, otherwise it is considered to be a non-ground point.

4. The outdoor scene building extraction method based on multi-feature constraints according to claim 3 is characterized in that: The step 3 is specifically as follows: Step 3.1, construct the KD tree; Step 3.2, select an unprocessed point from the point cloud , unprocessed points refer to points that have not been marked as processed, and are searched according to the index Points in the neighborhood, if the neighborhood contains at least Point, proof As a core point, the point is marked as processed and a new cluster is created ,Will of Neighborhood points added to Otherwise, the point is marked as processed and a new point is selected until the core point is selected; Step 3.3, traverse the cluster For the unprocessed points in the traversal process, if there is a point that is also a core point under a given threshold, then the point All points in the neighborhood are added to the cluster and mark the point as processed; Step 3.4, repeat step 3.3 until no new points are added to the cluster ; Step 3.5: Repeat steps 3.2 to 3.4 until all points in the point cloud are processed, and clustering is completed.

5. The outdoor scene building extraction method based on multi-feature constraints according to claim 1 is characterized in that: The step 4.3 is specifically as follows: Step 4.3.1: First, divide the cosine value of the angle between the point cloud normal vector and the Z-axis direction vector into multiple intervals. The interval size is in the range of [0,1], and set the group distance. , then construct the distribution histogram of the angle cosine value; Step 4.3.2, calculate the probability of each interval in the cosine value distribution of the angle between the point cloud normal vector and the Z-axis direction vector and the standard deviation of all probabilities, and count the number of intervals with a probability greater than the standard deviation The number of intervals whose sum probability is less than the standard deviation ,if , proving that the cosine value distribution of this point cluster is relatively discrete and it is not a building, so it is eliminated.

6. The method for extracting buildings from outdoor scenes based on multi-feature constraints according to claim 5, characterized in that: The step 4.4 is specifically as follows: Step 4.4.1, calculate the eigenvalues of the covariance matrix of each cluster of points ; Step 4.4.2, let , use formula (6) to calculate the probability that a point in the space belongs to the three dimensions respectively; (6) In formula (6), represents linearity, that is, the probability that the point cloud belongs to a one-dimensional feature, represents the flatness, that is, the probability that the point cloud belongs to a two-dimensional feature, Represents the degree of scatter, that is, the probability that the point cloud belongs to a three-dimensional feature; In step 4.4.3, the dimensional features of the current point are obtained according to the size of the probability values, and the category to which the point belongs is further obtained. If the proportion of surface points in a point cluster is the highest, it is determined to belong to a building. On the contrary, if the proportion of scattered points or line points is higher than that of surface points, it is determined to be a non-building, thereby extracting all building point clouds in the scene.

7. The method for extracting buildings from outdoor scenes based on multi-feature constraints according to claim 6, characterized in that: The step 4.5 is specifically as follows: Step 4.5.1: Introduce the overgreen index EXG into the 3D point cloud scene as the basis for removing sparse trees. Its calculation formula is as shown in formula (7); (7) In formula (7), , Respectively represent the color values of the point cloud; Step 4.5.2, if Index greater than threshold , then judge the point All points in the neighborhood Index and count its number ,if Greater than the point threshold , then the point is a tree point, otherwise it is a building point, and the threshold Set to 0.08, threshold The adaptive one is set to half of the number of point clouds in the neighborhood, and the building points are retained to complete the building point extraction.

Images