CN115099304A

CN115099304A - Building facade point cloud extraction method

Info

Publication number: CN115099304A
Application number: CN202210564805.3A
Authority: CN
Inventors: 熊峰; 李宗春; 付永健; 汪文琪; 冉佳欢; 何华; 黄哲琨
Original assignee: Information Engineering University of PLA Strategic Support Force
Current assignee: Information Engineering University of PLA Strategic Support Force
Priority date: 2022-05-23
Filing date: 2022-05-23
Publication date: 2022-09-23

Abstract

The invention belongs to the technical field of point cloud data processing, and particularly relates to a building facade point cloud extraction method. Firstly, ground points are separated in a self-adaptive mode by utilizing the elevation distribution characteristic based on a virtual grid, and low ground object point clouds are filtered according to the grid height difference. And then, on the basis of solving the optimal neighborhood, carrying out DBSCAN clustering according to the local point cloud density of each point. And finally, performing fine extraction on the point cloud of the facade of the building by combining semantic rules such as the central line of the clustering block, the proportion and the size of the planar points and the like. The results of tests with two different scene data sets show that the method can obtain better building facade extraction results, the extraction precision is higher than that of the MRG, RealWorks and EPSB comparison methods, F1 scores are better than 97.69%, and reliable building facade information can be provided for building reconstruction, city fine management and other applications.

Description

Building facade point cloud extraction method

Technical Field

The invention belongs to the technical field of point cloud data processing, and particularly relates to a building facade point cloud extraction method.

Background

With the introduction of concepts such as smart cities, Building Information Models (BIMs), City Information Models (CIMs), and the like and the deep development of related technologies, the demand for accurately and effectively reconstructing three-dimensional models of cities is increasing. LiDAR point cloud data is used as a mainstream modeling data source, and plays an increasingly important role in the fields of city fine management, city safety analysis, three-dimensional change detection, navigation and the like.

Buildings are main components in urban scenes, and how to extract building point clouds from the scenes is always a research direction which is concerned by many scholars. Compared with an Airborne Laser Scanning (ALS) system, a ground Laser Scanning (TLS) system and a Mobile Laser Scanning (MLS) system can acquire point clouds with high density of building facades, thereby intuitively presenting rich detailed features and being used for three-dimensional reconstruction of buildings. At present, scholars at home and abroad develop a lot of research on how to accurately and efficiently extract the building facade from the TLS/MLS point cloud, and the research can be summarized into the following three categories:

1) two-dimensional based methods. The point cloud is projected to a two-dimensional plane, and the point cloud is extracted by utilizing the difference of projection density or converted into a two-dimensional image by utilizing an image processing technology. For the projection density method, the projection density method is often influenced by other ground object targets, the threshold value is difficult to set, and the roof information is easily lost; for the image method, a certain amount of information loss is caused in the conversion process, and the method is difficult to be applied to complex scenes.

2) A segmentation-based approach. In the method, a building is mostly assumed to be composed of a plurality of planes, the scene point cloud is firstly segmented to obtain plane segmentation blocks, and then the semantic rules of the building are utilized to identify plane areas, so that the building facade point cloud is extracted. However, the extraction effect of the method depends on a segmentation method and semantic rules.

3) A classification based approach. Generally, a scene point cloud is classified, and then a point cloud marked as a building is extracted from a classification result. Although the method can classify the scene point cloud into different ground objects and is convenient for extracting the facade of the building, various different classification characteristics need to be selected and extracted, a large number of thresholds need to be set manually, and different scenes are difficult to meet.

In conclusion, most of the existing building facade extraction algorithms have the problems that parameters are difficult to reasonably set, extraction accuracy is low and the like.

Disclosure of Invention

The invention aims to provide a building facade point cloud extraction method, which is used for solving the problem of low building facade point cloud extraction precision in the prior art.

In order to solve the technical problem, the invention provides a building facade point cloud extraction method, which comprises the following steps:

1) acquiring point cloud data, and determining the optimal number of neighborhood points of each point in the point cloud data and the farthest distance of each point, wherein the farthest distance of each point refers to the maximum value of the distance between the point and each neighboring point under the optimal number of the neighborhood points;

2) determining the local point cloud density of each point according to the optimal neighborhood point number and the farthest distance of each point:

in the formula, LPD _i Local point cloud density, k, representing point i _i Number of optimal neighborhood points, d, representing point i _kimax Represents the farthest distance of point i;

3) dividing all the points into N types according to the local point cloud density of each point, wherein N is more than or equal to 2; for a kind of point cloud, determining the median of local point cloud densities of all points in the kind of point cloud, selecting the farthest distance of the points corresponding to the median of the local point cloud densities as the neighborhood search radius for carrying out DBSCAN clustering on each point in the kind of point cloud, and further determining the neighborhood search radius of each point in each kind of point cloud;

4) clustering the point cloud data by adopting a DBSCAN clustering method according to the neighborhood search radius of each point;

5) and extracting the point cloud of the facade of the building according to the clustering result.

The beneficial effects are as follows: the method calculates the local point cloud density of each point by combining the optimal neighborhood of each point, applies the local point cloud density to the DBSCAN clustering algorithm, and takes the local point cloud density as the basis for the subsequent calculation of the neighborhood search radius of each point so as to prevent the phenomenon that a plurality of clustering segmentation blocks appear in a region with small density when the clustering is directly carried out by using the DBSCAN clustering algorithm, thereby improving the precision of the subsequent building facade point cloud extraction. Moreover, considering the problem of calculation efficiency, the invention divides all point clouds into a plurality of categories according to the local point cloud density, the neighborhood search radius of the point clouds in one category is the same, and compared with the method that each point adopts different neighborhood search radii, the invention greatly reduces the calculation cost and simplifies the calculation process; moreover, the neighborhood search radius of the point clouds in a class is the median of the local point cloud density of all the points in the class, namely a compromise processing method is selected to ensure the precision of the point clouds. Experiments prove that the method has higher building facade extraction precision, and can better extract the building facades with uneven point cloud density distribution and different shape complexity.

Further, the following method is adopted in step 1) to determine the optimal number of neighborhood points of a certain point:

setting the number k of initial neighborhood points of a point p ₀ (ii) a Taking the number of the initial neighborhood points as the number of the current neighborhood points;

solving three characteristic values lambda of point p under current neighborhood point number ₁ 、λ ₂ 、λ ₃ And λ ₁ ＞λ ₂ ＞λ ₃ ；

Thirdly, calculating the entropy value E of the point p under the current neighborhood point number according to the following formula _λ ：

E _λ ＝-L ₁ ln(L ₁ )-L ₂ ln(L ₂ )-L ₃ ln(L ₃ )

Fourthly, repeating the second step and the third step according to the traversal range and the traversal step length of the number of the neighborhood points, and thus calculating to obtain the entropy value of each point in the traversal range; and selecting the number of the neighborhood points corresponding to the minimum entropy value as the optimal number of the neighborhood points of the point p.

Further, the traversal range is [ k ] _min ,k _max ]The traversal step length is as follows: at [ k ] _min ,k ₁ ]∪[k ₂ ,k _max ]The traversal step length in the range is k _△1 In (k) ₁ ,k ₂ ) The step size of traversal in the range is k _△2 ，k _△1 ＜k _△2 。

The beneficial effects are as follows: the traversal step lengths in different ranges are different, so that the total traversal range can be designed to be larger to set a larger traversal range to find the most suitable number of the optimal neighborhood points, moreover, the step length set in the range where the optimal neighborhood points are most likely to appear is larger, and the step length set in the range where the optimal neighborhood points are less likely to appear is smaller, so that the calculation cost of the search range is reduced.

Further, k is _△1 ＝1，k _△2 ＝2。

Further, before determining the optimal number of neighborhood points of each point in the point cloud data in step 1), the obtained point cloud data needs to be screened to remove short feature points.

The beneficial effects are as follows: before clustering, the depopulated object points are removed to improve the efficiency of subsequent clustering calculation.

Further, the screening method comprises the following steps: projecting the point cloud data to a corresponding two-dimensional grid plane along the z-axis direction, dividing the grid according to a set size, and determining the grid where each point is located; determining the elevation values of all points, finding the elevation value corresponding to the lowest point of all the points, calculating the elevation difference value from the lowest point to the lowest point of all the points in each grid, and subtracting the elevation difference value corresponding to the grid from the elevation value of each point in the grid to obtain the result which is the elevation distribution statistical value of each point; dividing an interval between the maximum value and the minimum value of the elevation distribution statistical value into M equal-interval intervals, wherein M is larger than 1, counting the number of points in each equal-interval, drawing a frequency histogram by taking the equal-interval intervals as horizontal coordinates and the number of points as vertical coordinates, connecting a starting point and an end point of a curve in the frequency histogram to obtain a straight line, calculating the distance between each point on the curve and the straight line, and counting the elevation value of a point corresponding to the maximum distance value to be used as an elevation threshold value for screening ground points and non-ground points; and screening out points with the elevation value smaller than the elevation threshold value.

The beneficial effects are as follows: the elevation distribution characteristic is used for adaptively separating the ground points based on the grid method, so that the influence of terrain fluctuation on elevation distribution can be overcome.

Further, the extraction of the point cloud of the facade of the building in the step 5) needs to meet the following three conditions:

h _hight ≥h _Δ

l _width ≥l _Δ

in the formula, n _linearity 、n _planarity 、n _scattering Respectively the number of linear, planar and spherical points in the cluster block, delta _n Is a linear-planar point proportion threshold value h in a clustering block _hight 、l _width Height difference and maximum width of clustering block, h _Δ 、l _Δ Respectively a height difference threshold and a maximum width threshold.

The beneficial effects are as follows: the method is used for carrying out fine extraction on the building facade based on the characteristics of high proportion of building linear-surface-shaped points and large size of clustering blocks so as to improve the accuracy of point cloud extraction of the building facade.

Further, N is 2, and is divided as follows: and determining the median of the local point cloud densities of all the points, and dividing the points with the local point cloud densities larger than the median of the local point cloud densities of all the points into high-density point clouds, or else, dividing the points into low-density point clouds.

The beneficial effects are as follows: and all point clouds are classified into two types, so that the calculation efficiency is improved.

Further, when screening, points in the grid where the elevation difference value is smaller than the set elevation difference threshold and the lowest elevation in the grid is larger than the average elevation of all the points need to be reserved.

The beneficial effects are as follows: the mode can keep the roof surface point cloud to avoid the condition that the roof surface point cloud is rejected as a low ground point.

Drawings

FIG. 1 is a flow chart of a building facade point cloud extraction method of the present invention;

FIG. 2(a) is a frequency histogram of the present invention;

FIG. 2(b) is a schematic diagram of adaptively selecting an elevation threshold value according to the present invention;

FIG. 3 is a schematic diagram of DBSCAN algorithm;

FIG. 4(a) is a diagram illustrating the selection of the neighborhood search radius Eps when the LPD is low;

FIG. 4(b) is a schematic diagram of the neighborhood search radius Eps selection when the LPD is higher;

FIG. 5(a) is a Data 1 Data map;

FIG. 5(b) is a Data 2 Data map;

fig. 6(a) is a graph of clustering results using the DBSCAN algorithm;

FIG. 6(b) is a graph of the clustering results using the LPD-DBSCAN algorithm;

FIG. 7(a) is a Data 1 point cloud true value map of the building facade;

FIG. 7(b) is a Data 2 point cloud true value map of the building facade;

FIG. 8(a) is a diagram showing the result of Data 1 extraction by the MRG method;

FIG. 8(b) is a diagram showing the result of Data 2 extraction by the MRG method;

FIG. 9(a) is a diagram showing the result of extracting Data 1 by the Realworks method;

FIG. 9(b) is a diagram showing the result of extracting Data 2 by the Realworks method;

FIG. 10(a) is a diagram showing the result of Data 1 extraction by the EPSB method;

FIG. 10(b) is a diagram showing the result of Data 2 extraction by the EPSB method;

FIG. 11(a) is a diagram showing the result of Data 1 extraction using the method of the present invention;

FIG. 11(b) is a diagram showing the result of Data 2 extraction by the method of the present invention.

Detailed Description

Aiming at the problems that parameters are difficult to reasonably set, the extraction precision is low and the like in a building facade extraction algorithm in the point cloud data processing of a laser radar (Light Detection and Ranging), the building facade point cloud extraction method combining a virtual grid and a DBSCAN (sensitivity-Based Spatial Clustering of Applications with Noise) Clustering algorithm considering local point cloud Density is provided. The main contributions of the method are: 1) an elevation threshold value determining method for adaptively separating ground points according to elevation distribution characteristics is provided, and low short feature point cloud data are further filtered by adopting a grid method so as to reduce the time complexity of subsequent data processing; 2) the DBSCAN algorithm is improved, an LPD-DBSCAN clustering algorithm is provided, the point cloud density of different areas can be considered well while parameter setting is reduced, and a better clustering effect is obtained.

The following describes a method for extracting a point cloud of a facade of a building in detail with reference to the accompanying drawings and embodiments.

The embodiment of the method for extracting the point cloud of the facade of the building comprises the following steps:

the whole flow of the building facade point cloud extraction method is shown in figure 1. It should be noted that specific values of the partial parameter threshold used in this embodiment will be described in the final simulation test.

The method comprises the steps of firstly, acquiring original point cloud data, and utilizing elevation distribution characteristics to adaptively separate ground points based on a grid method so as to remove low and short feature point clouds.

1. And constructing a virtual grid. The core idea of virtual grid construction is to project point cloud data to a corresponding two-dimensional grid plane along the z-axis direction, and to divide the grid according to a certain size, so as to subsequently count the difference of each point elevation in the grid. The method comprises the following specific steps:

step 1: and projecting the original point cloud data to the xoy plane along the z-axis direction.

Step 2: after projection, firstly setting the side length s of the square grid, and then determining the row number R and the column number C of the grid according to the point cloud plane distribution.

In the formula, x _max 、y _max 、x _min 、y _min Respectively the maximum value and the minimum value of x and y in the projection point cloud plane distribution, and ceil () is an upward rounding symbol.

Step 3: and (3) establishing indexes for the points, and calculating the grid serial number of the point by using the formula (2).

Wherein (x) _i ,y _i ) Is the plane coordinate of a certain point, (R) _i ,C _i ) The row number and column number of the grid where the point is located.

2. Coarse screening to filter out ground point cloud. In the point cloud data, the elevation value of ground points is generally smaller, buildings are larger than other ground objects, and the number of ground object points gradually decreases with the increase of the elevation value. Based on the characteristic, the invention filters ground points according to the elevation distribution condition of the point cloud data. The method comprises the following specific steps:

step 1: to overcome the effect of terrain fluctuation on elevation distribution, a small-range grid is assumedThe topographic relief inside the point cloud data is negligible, and the elevation value H of the lowest point O of all the point cloud data is used _o And sequentially calculating the elevation difference delta h from the lowest point to the point O in each grid as a reference, and subtracting the delta h from the elevation value of each point in the grid to obtain an elevation distribution statistical value, wherein a specific formula is shown as a formula (3). Because the method is easy to remove the point cloud of the roof surface as the low ground point, the invention adds the limitation condition of the roof surface, namely, the grid points with smaller grid height difference and the lowest elevation in the grid larger than the elevation of all the points on average are reserved.

In the formula, H _imin Is the elevation of the lowest point in the ith grid,

for the elevation value of the j point in the ith grid,

and g is the total number of grids, and n is the total number of point clouds in the ith grid.

Step 2: calculating the elevation distribution statistic of each point cloud according to the formula (3) to obtain the maximum elevation z _max And minimum elevation z _min (ii) a Will maximum elevation z _max And minimum elevation z _min Between them are divided into M equal parts, the distance between every two equal parts is z _Δ Counting the elevation value in z _Δ The number N of points in each interval is increased progressively; the elevation value z horizontal coordinate and the point number N are used as the vertical coordinate to draw a corresponding frequency histogram, which is an elevation-frequency curve, as shown in fig. 2 (a).

Step 3: connecting the starting point and the end point of the elevation-frequency curve to obtain a straight line L, calculating the distance H from each point on the elevation-frequency curve to the straight line L, counting to obtain the point with the maximum H, and taking the elevation value of the point as an elevation threshold value z for screening ground point cloud and non-ground point cloud as shown in figure 2(b) _t This point is actually an inflection point on the elevation-frequency curve.

Step 4: according to elevation threshold z _t Traversing all grids, removing the elevation statistic value less than the elevation threshold value z _t Removing short substrate points to obtain the point cloud after coarse screening.

And step two, clustering the point cloud after coarse screening by adopting an LPD-DBSCAN algorithm to obtain a clustering result.

Points on the same ground object are adjacent in space, and different ground objects are connected through the ground. Because a plurality of ground objects (such as trees) which are missed to be screened exist after the coarse screening, and the connection of the ground does not exist among various ground objects after the coarse screening to present an independent state, the point cloud after the coarse screening can be clustered by using a clustering method based on the local point cloud density in the step so as to assist in further extraction.

1. And solving the optimal neighborhood point number of each point.

In the step, an optimal neighborhood calculation mode based on characteristic values is adopted to adaptively select the number of the proper neighborhood points for each point in the subsequent self-adaptive selection of parameters (neighborhood search radius Eps) in the clustering process and solve the dimension characteristics of each point, but the search range of the mode needs manual setting (generally set to be 10, 100), the optimal neighborhood points can not be found in a smaller search range, and the calculation cost is increased in a larger search range. In order to enlarge the neighborhood search range and reduce the calculation cost, the invention adopts a segmentation mode to traverse different ranges by adopting different step lengths on the search range based on the statistical result of the optimal neighborhood of point clouds of different categories, and the specific steps are as follows:

step 1: setting traversal range [ k ] of number k of neighborhood points _min ,k _max ]The number of initial neighborhood points is k _min . Wherein, step k is traversed _Δ A range of 1 is [ k ] _min ,k ₁ ]∪[k ₂ ,k _max ]，k _Δ A range of 2 is (k) ₁ ,k ₂ ). In this example, k _min ＝10，k _max ＝200，k ₁ ＝50，k ₂ ＝180。

Step 2: calculating three corresponding to the point p under the k neighborhood according to a principal component analysis methodA characteristic value lambda ₁ 、λ ₂ 、λ ₃ (λ ₁ ＞λ ₂ ＞λ ₃ ). Step 3: calculating Shannon entropy value E of point p in k neighborhood according to characteristic value obtained by Step2 _λ The calculation formula is as follows:

E _λ ＝-L ₁ ln(L ₁ )-L ₂ ln(L ₂ )-L ₃ ln(L ₃ )

step 4: continuously searching the number k of the neighborhood points which is k + k _Δ Repeating Sep 2-Step 3 to obtain corresponding entropy E under each neighborhood point number _λ Until k equals k _max 。

Step 5: comparing entropy values E corresponding to the number k of the points p in different neighborhoods _λ The number k of the neighborhood points corresponding to the minimum entropy value is the optimal neighborhood k of the point p _opt 。

Step 6: and (4) obtaining the optimal number of neighborhood points of each point according to the steps of Sep 1-Step 5.

2. And determining the neighborhood search radius Eps of each point in the LPD-DBSCAN clustering algorithm.

The DBSCAN algorithm is proposed by Ester Martin equal to 1996, and is a typical density clustering algorithm. The algorithm is not influenced by the size and the shape of the data, and the data can be divided into different categories according to the density of the data. The core idea is that two parameters of a neighborhood search radius Eps and a neighborhood minimum point number MinPts of a data point are preset for a point set X, and a point cluster is established by iteratively searching a core point satisfying the formula (6). As shown in fig. 3, there is a certain point set Y, point a is one point in the point set, where MinPts is 4, and there are 4 points in the Eps neighborhood of point a, point a is referred to as a core point, and then iterative determination is performed on other points in the search neighborhood, that is, point B, D is marked as the core point, point C, E is marked as boundary points (the number of points in the Eps neighborhood is less than 4), and then iterative determination is performed on point B, D until the Eps neighborhood only contains boundary points.

In the formula, Eps (x) ₀ ) Is a point x ₀ Set of neighborhood points of (2), D (x, x) ₀ ) Is point x and core point x ₀ The distance of (a) to (b),

is a core point x ₀ The number of neighborhood points.

The point cloud data has non-uniformity, and the local point cloud density is different, so that the point density of the ground object close to the scanner is usually larger than that of the ground object far away. Therefore, when the DBSCAN algorithm is directly used for clustering, a phenomenon that a plurality of clustering segmentation blocks appear in a region with low density often exists, and the subsequent building facade extraction result is influenced. Aiming at the problem, the method is based on a local point cloud density (LPD) calculation mode, and an optimal neighborhood point number calculation method is combined to calculate the LPD of each point, so that the LPD is used as a neighborhood search radius of subsequent DBSCAN clustering.

In the formula, LPD _i For local point cloud density, k is the optimal number of neighborhood points of the current point, d _kimax The farthest distance (hereinafter referred to as the farthest distance of the current point) in the distances between the current point and the k neighboring points in the optimal number of the neighboring points.

The process of specifically determining the optimal neighborhood search radius Eps of each point is as follows:

step 1: solving the optimal neighborhood point number of each point obtained according to the method in the step1 and obtaining the farthest distance d by the optimal neighborhood point number of each point _max The local point cloud density LPD of each point is obtained by formula (7) _i 。

Step 2: according to all pointsLocal Point cloud Density, determining the median LPD of the local Point cloud Density for all points _m The local point cloud density is greater than the median LPD of the local point cloud densities of all the points _m The point clouds of (2) are classified into high-density point clouds, and the point clouds less than or equal to are classified into low-density point clouds, so that all the point drawings are classified into two types.

Step 3: for each point in the point cloud class with small density, determining the median LPD of the local point cloud density of all the points in the point cloud class _sm Selecting the median LPD of the local point cloud density _sm The farthest distance of the corresponding point is used as the neighborhood search radius for carrying out DBSCAN clustering on each point in the point cloud; similarly, for each point in the high-density point cloud, the median LPD of the local point cloud density of all the points in the point cloud is determined _bm Selecting the median LPD of the local point cloud density _bm And the farthest distance of the corresponding point is used as the neighborhood search radius for carrying out DBSCAN clustering on each point in the point cloud. The selection diagram of the neighborhood search radius Eps when the LPD is low is shown in FIG. 4(a), and the selection diagram of the neighborhood search radius Eps when the LPD is high is shown in FIG. 4 (b).

3. And (3) clustering the point cloud data screened in the first step by using the neighborhood search radius of each point determined in the step (2) and adopting a DBSCAN clustering algorithm, thereby obtaining a plurality of clustering blocks.

And step three, further screening the clustering result of the step two to screen out other ground objects (such as trees and electric poles) in the scene, so that accurate building facade point cloud extraction is realized.

After clustering with the LPD-DBSCAN algorithm, there are a few outliers. In this step, outliers and small block regions (cluster blocks with a total number less than T, in this embodiment, T is 20) are further clustered. Although many other ground objects (such as trees and electric poles) still exist in the scene, most single buildings after clustering are presented as a category, and at the moment, semantic rules can be well utilized for fine extraction. After the dimension characteristics of each point are calculated on the basis of the optimal neighborhood, the building vertical face is finely extracted on the basis of the characteristics that the ratio of the building linear-planar points is high and the size of the cluster block is large, and the semantic rule selection condition is shown as a formula (9).

In the formula, n _linearity 、n _planarity 、n _scattering Respectively the number of linear, planar and spherical points, delta, in the clustering block calculated based on the dimension characteristics _n Is a linear-planar point proportion threshold value h in a clustering block _hight 、l _width Height difference and maximum width of clustering block, h _Δ 、l _Δ Respectively a height difference threshold and a maximum width threshold.

The method is applied in the following specific examples to illustrate the effectiveness of the method of the invention.

1) And (4) test data. The experiments were performed on a computer with Intel (R) core (TM) i7-108750H CPU @2.30GHz RAM 32 GB. The public data sets of the point clouds obtained by two different platforms are selected to verify the effectiveness of the method, and the data sets are displayed as shown in fig. 5(a) and fig. 5 (b). Data 1 is part of Data in an open source Data set Robotic 3D Scan retrieval-12, the Data is collected by a Riegl VZ-400 ground laser scanner at a certain place, and 6 types of ground objects such as buildings, the ground, trees, automobiles and the like are contained in a scene, as shown in FIG. 5 (a); data 2 is from the Paris-rue-Madame open source dataset, which is collected by L3D2 mobile laser scanner on rue Madame street in Paris, france, and the scene contains 26 types of ground objects such as buildings, ground, vegetation, cars, etc., as shown in fig. 5 (b). It should be noted that, the original point cloud data volume is large, and the down-sampling processing is performed during the test, so as to improve the calculation efficiency.

2) And (4) evaluating the index. In order to quantitatively evaluate the performance of the method, three indexes of Correctness (Correctness), Completeness (Completeness) and F1 score are adopted to evaluate the building facade extraction result, and the specific calculation mode is shown as formula (10).

In the formula, tp (true positive) represents the number of building facade point clouds correctly extracted by the method, fp (false positive) represents the number of building facade point clouds erroneously extracted by the method, and fn (false negative) represents the number of building facade point clouds not extracted by the method.

3) And setting parameters. The method is used for extracting the building facade from test data, and related parameters mainly comprise grid side length s, difference part number M of elevation values, low ground screening elevation H, LPD-DBSCAN clustering time neighborhood minimum point number MinPts and linear-planar point proportion threshold delta _n Cluster block height difference threshold h _Δ And a maximum width threshold value l _Δ Seven parameters. The s value is too small, ground points are not easy to screen, and corresponding errors are increased if the s value is too large, and the s value is mainly selected from 1-2 m; h is set to be 3 m; the size of the clustering block is set to be 3m multiplied by 3 m; other parameter settings are shown in table 1.

TABLE 1 Experimental parameter settings

4) And comparing the building facade extraction results.

And (4) comparing clustering methods. In order to verify the effectiveness of the LPD-DBSCAN algorithm provided by the invention, the building point cloud Data in the test Data 1 is taken as an example, and the DBSCAN algorithm and the LPD-DBSCAN algorithm provided by the invention are used for clustering. The clustering results are respectively shown in fig. 6(a) and fig. 6(b), and it can be found that the algorithm of the present invention has a significant advantage in the scene with uneven density.

And secondly, comparing the extraction methods of the facades of different buildings. To better evaluate the effectiveness of the method of the invention, three exemplary methods were selected for comparison with the method of the invention. Method 1: firstly, a Multi-Rule Region Growing (MRG) method based on dimension features is used for roughly classifying scene point clouds, then corresponding Growing criteria are set for segmentation, and then semantic rules (segmentation block size h) of building vertical faces are combined _p ×l _p Normal vector direction range theta _p Plane fitting residual T _p ) Extracting a building facade. The parameter setting conditions in the fine extraction stage are shown in table 2. Method 2: in order to compare the effect of the point cloud classification method on the building facade extraction, the test selects an outdoor point cloud classification tool in Trimble Realworks which can be suitable for different scenes to classify test data, so that the building facade extraction is realized. A Method 3: the extraction process (hereinafter referred to as the EPSB process) is mainly divided into two steps: removing ground points according to the coordinate in the z direction and the included angle between the normal vector and the z direction; and (4) carrying out Euclidean distance clustering on the non-ground points, and extracting the building facade by combining the planar point proportion and the maximum height difference of the clustering blocks. Wherein the maximum relief h of the terrain involved _t Normal vector angle threshold n _t Euclidean clustering distance threshold L _t Planar dot proportion threshold K _α And the maximum height difference Δ H are shown in table 3.

The extraction results of two different scenes by the method and three comparison methods are shown in table 4, the Data 1 and Data 2 building elevation truth results are respectively shown in fig. 7(a) and 7(b), the extraction effect and the local detail effect of extracting Data 1 by applying the four methods are respectively shown in fig. 8(a), 9(a), 10(a) and 11(a), and the extraction effect and the local detail effect of extracting Data 2 by applying the four methods are respectively shown in fig. 8(b), 9(b), 10(b) and 11(b), wherein a black frame is a local amplification area, and a black circle is an area with missing method or obvious false extraction.

TABLE 2 Method 1 Fine extraction stage parameters

TABLE 3 Method 3 Process parameters

TABLE 4 comparison of the results of different extraction methods

As can be seen from fig. 8(a) to 10(b), the missing and mis-extracted regions exist in all of the three comparison methods. The reason for analyzing the MRG method is mainly that, in the case of a complex scene, there are many small segmented blocks, which makes it difficult to set parameters and makes it very easy to delete small building segmented blocks by mistake when the method subsequently uses semantic rules to extract buildings, and in addition, there are segment groove portions (as shown by black circles in fig. 8 (b)) in the Data 2, which are disordered in the normal vector direction of ground points, and therefore cannot be merged with other ground points when the ground point region grows. For the Realworks method, there are many missing regions (as shown by black circles in fig. 9), and the reason for analyzing the missing regions is mainly that the classification features selected by the method are limited and are not suitable for building regions with poor quality and complex shapes. For the EPSB method, the reason for analyzing the method is mainly that the proportion of the planar point calculated according to the dimension characteristics is different in different building point clouds, so that the threshold value is difficult to select when building extraction is performed after clustering, and a missing extraction area (as shown in black circle in fig. 10 (a)) is very easy to occur, in addition, the method also has the similar problem with the MRG method in Data 2, the EPSB method cannot effectively filter the ground points in the area (as shown in black circle in fig. 10 (b)) when the ground points are separated by using the normal vector, so that the building point clouds on both sides are clustered into one clustering block, and a good extraction result is difficult to obtain. Comparing fig. 7(a), fig. 7(b) and fig. 11(a), fig. 11(b), it can be seen that, in sharp contrast to the other 3 methods, the building extraction result of the method of the present invention is closer to the building truth value. The main reasons are that: in the ground point removing stage, only the elevation characteristic is utilized, and other geometric characteristics are not relied on, so that the influence of the unevenness of the terrain is avoided, and the applicability is stronger; the neighborhood search radius can be adaptively selected for clustering in the clustering stage, and the problem that the subsequent fine extraction is influenced by small clustering blocks caused by uneven local area density can be solved.

According to the parameter settings of the different methods in table 1-table 3 and the extraction results of the point cloud of the building obtained by the different methods in table 4 and fig. 8(a) -fig. 11(b), in combination with the above analysis of the reasons for the missing and mis-extracted areas of the different methods, it can be found that:

the method is easier to set parameters, and the extraction result does not depend on the accurate selection of the parameters. In the MRG method, parameters such as a main direction, a normal vector and the like need to be set in a region growing stage, a segmentation result depends on the selection of the parameters, and in a building extraction stage, only the size and the direction of the normal vector of a segmentation block are set, so that tree points are easy to be extracted by mistake, and therefore a better extraction result can be obtained by further increasing constraint conditions (such as plane fitting residual errors); in the stage of removing the ground points, the EPSB method needs to continuously try a maximum relief threshold value according to relief, and in the stage of extracting the building, the proportion of the planar points needs to be controlled to prevent the mistaken screening of the building and the mistaken extraction of other ground objects.

Secondly, under the condition of complex scenes, the division blocks of the MRG algorithm are small, so that subsequent building extraction is inconvenient; the classification-based method in Realworks software is easy to miss extraction due to the limitation of classification characteristics, and the method has better effect of extracting the vehicle-mounted point cloud data from the building; although the EPSB method has better extraction result than the former two methods, the phenomena of extraction leakage and extraction error are easy to occur. Compared with other methods, the building point cloud extraction method has the advantages of good overall result, integrity, correctness and F1 score.

In conclusion, the invention provides a building facade extraction method combining a virtual grid and an LPD-DBSCAN algorithm by taking ground/vehicle-mounted LiDAR point cloud data as a research object based on the practical requirement of reconstructing a three-dimensional model of a city at present. The data of two different scenes are selected for testing, and the result shows that compared with a comparison method, the method has higher building facade extraction precision and can better extract the building facades with uneven point cloud density distribution and different shape complexity degrees. In addition, the method is simple in principle, does not need to set more parameters and select more complex building semantic rules, has strong adaptability, can provide early guarantee for subsequent building detail feature extraction (such as windows and the like), geometric feature extraction (such as line and plane features and the like) and building modeling, and can provide reliable building facade information for applications such as building reconstruction and city fine management.

Claims

1. A building facade point cloud extraction method is characterized by comprising the following steps:

4) clustering point cloud data by adopting a DBSCAN clustering method according to the neighborhood search radius of each point;

2. The method for extracting the point cloud of the facade of the building according to claim 1, wherein the optimal number of the neighborhood points of a certain point is determined by adopting the following method in the step 1):

setting the number k of initial neighborhood points of the point p ₀ (ii) a Taking the number of the initial neighborhood points as the number of the current neighborhood points;

E _λ ＝-L ₁ ln(L ₁ )-L ₂ ln(L ₂ )-L ₃ ln(L ₃ )

3. The building facade point cloud extraction method of claim 2, wherein the traversal range is [ k ] _min ,k _max ]The traversal step length is as follows: in [ k ] _min ,k ₁ ]∪[k ₂ ,k _max ]The traversal step length in the range is k _△1 In (k) ₁ ,k ₂ ) The traversal step length in the range is k _△2 ，k _△1 ＜k _△2 。

4. The building facade point cloud extraction method of claim 3, wherein k is _△1 ＝1，k _△2 ＝2。

5. The method for extracting point cloud from a facade of a building according to claim 1, wherein before determining the optimal number of neighborhood points of each point in the point cloud data in step 1), the obtained point cloud data is further screened to remove short feature points.

6. The building facade point cloud extraction method according to claim 5, wherein the screening method comprises the following steps:

projecting the point cloud data to a corresponding two-dimensional grid plane along the z-axis direction, dividing the grid according to a set size, and determining the grid where each point is located;

determining the elevation values of all points, finding the elevation value corresponding to the lowest point in all points, calculating the elevation difference value from the lowest point in each grid to the lowest point in all points, and subtracting the elevation difference value corresponding to the grid from the elevation value of each point in the grid to obtain the result of the elevation distribution statistical value of each point;

dividing an interval between the maximum value and the minimum value of the elevation distribution statistical value into M equal-interval intervals, wherein M is larger than 1, counting the number of points in each equal-interval, drawing a frequency histogram by taking the equal-interval intervals as horizontal coordinates and the number of points as vertical coordinates, connecting a starting point and an end point of a curve in the frequency histogram to obtain a straight line, calculating the distance between each point on the curve and the straight line, and counting the elevation value of a point corresponding to the maximum distance value to be used as an elevation threshold value for screening ground points and non-ground points;

and screening out points with the elevation value smaller than the elevation threshold value.

7. The method for extracting the point cloud of the facade of the building according to claim 1, wherein the point cloud of the facade of the building extracted in the step 5) meets the following three conditions:

h _hight ≥h _Δ

l _width ≥l _Δ

in the formula, n _linearity 、n _planarity 、n _scattering Respectively the number of linear, planar and spherical points in the clustering block, delta _n Is a linear-planar point proportion threshold value h in a clustering block _hight 、l _width Respectively the height difference and the maximum width of the clustering block, h _Δ 、l _Δ Respectively a height difference threshold and a maximum width threshold.

8. The building facade point cloud extraction method according to claim 1, wherein N is 2, and the division is performed according to the following method: and determining the median of the local point cloud densities of all the points, and dividing the points with the local point cloud densities larger than the median of the local point cloud densities of all the points into high-density point clouds, or else, dividing the points into low-density point clouds.

9. The method of claim 6, wherein points are retained when the filtering is performed, where the elevation difference value in the grid is smaller than a predetermined threshold and the lowest elevation in the grid is greater than the average elevation of all points.