CN104573705A - Clustering method for building laser scan point cloud data - Google Patents

Abstract

The invention discloses a clustering method for building laser scan point cloud data. The method comprises the following steps: converting the point cloud data into two-dimensional data to acquire a set X; appointing density threshold value MinPts; calculating the longest distance of MinPts objects with shortest distance from the point towards any point in the X; counting maximal and minimum value of the longest distance of all the points; classifying the difference value of the maximal and minimum value of the longest distance into n equal portions; making a circle by adopting the point generating the minimum value of the longest distance as the center and the gradual increasing value of the distance of the equal portions as the radius; calculating the number of points in each circle; adopting the difference value as a abscissa; adopting the numbers of the points in each circle as an ordinate; conducting fitting to form a curve; seeking points of inflexion of the curve; taking the number value of the abscissa corresponding to the points of inflexion as the value of sigma; building AQ-DBSCAN algorithm by adopting the MInPts and the sigma as conditions; conducting clustering on the points in the X to obtain the belonged cluster analysis of the points in the point cloud in the parts of the building.

Description

A clustering method for building laser scanning point cloud data

technical field

The invention relates to a clustering method for building laser scanning point cloud data.

Background technique

Ancient architecture is an important symbol of human civilization and a special carrier of cultural information. Therefore, inheriting and protecting ancient buildings is an unshirkable responsibility of contemporary people. However, due to the current economic development and urban construction, as well as the erosion of ancient buildings over time, many large ancient buildings have been deformed and damaged to varying degrees, and the protection of ancient buildings is facing a severe situation.

Three-dimensional color scanning technology has the characteristics of rapidity, non-contact, penetration, real-time, dynamic, initiative, high density, high precision, digitization, automation, etc. With the improvement of its measurement accuracy, scanning speed, spatial resolution, etc. With progress and price reduction, it has been more and more widely used in the protection of ancient buildings. It adopts non-contact measurement methods, which can go deep into complex environments and on-site scanning operations without damaging objects, and directly collect the three-dimensional data of various large, complex and irregular entities to the computer completely. In order to quickly reconstruct the 3D model of the scanned object. At the same time, the 3D laser point cloud data collected by it not only contains the spatial information of the target, but also records the reflection intensity information and color grayscale information of the target.

The 3D point cloud model of the building can be quickly obtained by using a 3D laser scanner, and each point in the point cloud data contains explicit 3D coordinates. Based on point cloud data, simple 3D browsing and roaming of buildings can be realized. However, on the one hand, the pure point cloud data structure has a huge amount of data, which is difficult to be applied to the efficient roaming and browsing of ancient buildings. On the other hand, it cannot provide higher-level analysis and research because it cannot provide semantic information. On this basis, it is necessary to further process the point cloud data, so as to extract and reconstruct the three-dimensional structure of the object, so as to realize the functions of display, analysis, measurement, simulation, simulation, monitoring, storage, and retrieval. At present, point cloud data processing requires human-computer interaction to a large extent, and the degree of automation is low.

Data segmentation is an important step in feature extraction and 3D modeling of point cloud data. Rabbani et al (2006) describe the point-by-point identification process of data segmentation into point cloud data, points with the same identification are considered to belong to the same surface or area, and those points with similar characteristics in a continuous area are segmented into a subset.

Current point cloud segmentation algorithms mainly include four categories: clustering-based segmentation, region-growing-based segmentation, model-fitting-based segmentation, and other hybrid segmentation algorithms.

General clustering algorithms generally include five types: hierarchical methods, partition methods, density-based methods, model-based methods, and grid-based methods.

Density-based clustering methods consider clusters to be high-density object regions separated by low-density regions in the data space, while data in sparse data regions are considered noise data. The algorithm sets a certain threshold, and as long as the density of the adjacent area of a certain point is greater than this threshold, the clustering can continue. Such methods can discover clusters of arbitrary shape and can filter "noisy" data. It can be found that clusters of arbitrary shapes are the biggest advantage of the algorithm, and its main disadvantage is that it is sensitive to user-defined density parameters, and different thresholds have a greater impact on the clustering results. Typical algorithms of this type of method include DBSCAN algorithm (Density-Based Spatial Clustering of Applications with Noise) and so on.

Contents of the invention

Because the clustering of point cloud data of ancient buildings on Gaussian spheres or other curves has various shapes, and there are a lot of noise points in them. Therefore, the density-based clustering method is more suitable for the clustering of this type of data. Aiming at the characteristics of ancient building point cloud data, the present invention proposes an improved AQ-DBSCAN algorithm based on the DBSCAN algorithm, and provides a clustering method for building laser scanning point cloud data using the AQ-DBSCAN algorithm. The AQ-DBSCAN algorithm of the present invention The DBSCAN algorithm solves the problem of automatic estimation of the neighborhood radius in the DBSCAN algorithm, and the problem of effectively selecting representative points, that is, reducing the amount of calculation when spreading in the cluster, so as to quickly perform density clustering processing.

The present invention applies the AQ-DBSCAN algorithm to the clustering of ancient building point cloud data to obtain which part of the building the point in the point cloud belongs to. For this reason, the technical solution provided by the present invention is:

A clustering method for building laser scanning point cloud data, comprising:

Step 1. Convert each point in the three-dimensional point cloud data obtained after laser scanning the building into two-dimensional data, and obtain a set X of two-dimensional data of the point cloud;

Step 2: Specify the density threshold minimum number of included points MinPts, for any point in the set X, calculate the farthest distance of the objects with the smallest number of included points MinPts closest to the point, and count the farthest distances of all points in the set X maximum and minimum values;

Step 3: Divide the difference m between the maximum value and the minimum value of the farthest distance into n equal parts, take the point in the point cloud that produces the minimum value of the farthest distance as the center, and use the equal distance m/n as the unit step by step The incremental value is the radius, make n circles, and calculate the number of points in each circle;

Step 4. Use the difference m as the abscissa, the equal distance m/n as the incremental unit interval of the abscissa, and the number N _m of points in each circle as the ordinate, draw a coordinate map, and fit the coordinate map Points on to form a drawn curve;

Step 5. Find the inflection point of the drawn curve, and use the value of the abscissa corresponding to the inflection point as the value of the neighborhood radius σ;

Step 6. Establish the AQ-DBSCAN algorithm based on the density threshold MinPts and the neighborhood radius σ, and use the AQ-DBSCAN algorithm to cluster the points in the set X to obtain which building the points in the point cloud belong to Cluster analysis of parts.

Preferably, in the clustering method of described building laser scanning point cloud data, described step 6 comprises:

6.1) Take the point that produces the minimum value of the farthest distance in the point cloud as the center of the circle, and use the neighborhood radius σ as the radius to draw the first-level circle. If the number of points in the first-level circle is less than MinPts, eliminate If the first-level circle is greater than MinPts; continue to 6.2)

6.2) With cσ where 0<c<1 is the radius, draw the first-level sub-circle. If the number of points in the annular area between the first-level circle and the first-level sub-circle is less than MinPts, select the annular area All the points in are used as the second-level circle center; if the number of points in the annular area is greater than or equal to MinPts, select MinPts points as the second-level circle center;

6.3) Draw at least one second-level circle with the selected second-level circle center as the center and the neighborhood radius σ as the radius. If the number of points in any second-level circle is less than MinPts, eliminate the second-level circle. If the secondary circle is greater than MinPts; then repeat steps 6.2) and 6.3), and draw circles step by step until the drawn circles are eliminated;

6.4) Gather the points in the circles of all levels into one class, and remove them from the point cloud, and repeat the operations from step 2 to step 6 for the points in the remaining point cloud.

Preferably, in the clustering method of the building laser scanning point cloud data, in the step 6.2), the method of selecting MinPts points as the second-level center of circle is:

In the circular area, first select the point farthest from the center of the first level as the first point, then select the point farthest from the first point as the second point, and then select the distance between the first point and the second point. The farthest point of the sum of the distances of the two points is taken as the third point, and follow this rule until the MinPts point is selected.

Preferably, in the clustering method of the building laser scanning point cloud data, the MinPts is 4.

Preferably, in the clustering method of building laser scanning point cloud data, the c is 3/4.

Preferably, in the clustering method of building laser scanning point cloud data, the inflection point is the point with the largest slope change.

Preferably, in the clustering method of the building laser scanning point cloud data, in the first step, each point in the three-dimensional point cloud data obtained after the laser scanning of the building is mapped to a Gaussian sphere so that into two-dimensional data.

The basic concept of DBSCAN algorithm is introduced below.

DBSCAN algorithm:

The density at a point x _i in a point set X can generally be defined as follows:

${\mathrm{<span\; class="notranslate">\; f</span>}}^{\mathrm{<span\; class="notranslate">\; D.</span>}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; )</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>$

Where U( _xi ) is the weight of point x _i ; Is the kernel function, the kernel function usually chooses a symmetric density function at the origin, such as Gaussian function, Epanechnikov kernel function, etc.; σ is the bandwidth of the kernel function, which actually represents an r-sphere neighborhood with σ as the radius. make

U(x _i )＝{1|x _i ∈X}

$\mathrm{<span\; class="notranslate">\; K</span>}<span\; class="notranslate">\; (</span>\frac{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\left\{\begin{array}{c}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; ,</span>\frac{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; 1</span>}\\ \mathrm{<span\; class="notranslate">\; 0</span>}<span\; class="notranslate">\; ,</span>\frac{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}{\mathrm{<span\; class="notranslate">\; \&sigma;</span>}}<span\; class="notranslate">\; ></span>\mathrm{<span\; class="notranslate">\; 1</span>}\end{array}\right.$

Then f ^D (x) is the definition of density in the DBSCAN algorithm.

Define the 1σ neighborhood: given a data set X, x ₀ is an object in the set X, then the dimensional hypersphere area with x ₀ as the center and σ as the radius is called the σ neighborhood σx ₀ of x ₀ , namely:

σ(x ₀ )={x∈X|D(x, x ₀ )≤σ}

Wherein, D(x, x ₀ ) represents the distance between x and x ₀ .

Definition 2 core point: For x ₀ ∈ X, given an integer MinPts, if the number of objects in σ(x ₀ ) satisfies |Nσ(x ₀ )|≥MinPts, then x ₀ is called (σ, MinPts) under the condition core point. Objects that fall within the σ neighborhood of a core point, but are not core points, are called boundary points.

Definition 3 Direct density reachability: Under the condition (σ, MinPts), if the objects x and x ₀ satisfy:

x∈σ(x ₀ );

|Nσ(x ₀ )|≥MinPts.

Then x is said to be directly density-reachable from x ₀ .

Definition 4 Density reachable: In a data set X, for an object sequence x ₀ , x ₁ , x ₂ , ..., x _n , if under the condition (σ, MinPts), _xi to _xi+1 are direct Density reachable (0≤i<n), then the object x ₀ to x _n is said to be density reachable.

Definition 5 Density connected: Given a data set X, under the condition of (σ, MinPts), if there are objects x ₀ , x _i and x _j such that x ₀ to x _i and x ₀ to x _j are density reachable, Then it is said that _xi and x _j are density connected under the condition of (σ, MinPts).

Definition 6 Clusters and noise: All density-connected objects form a cluster, and if an object is not in any cluster, then this object is called noise.

The algorithm steps of DBSCAN are as follows:

The DBSCAN algorithm needs to pre-specify the two parameters σ and MinPts, and establish clusters by iteratively searching for points directly accessible to the density.

The geometric shapes of ancient building components mainly include planes, cylinders and other types, which are mapped to Gaussian spheres into shapes such as points and lines. The DBSCAN algorithm can identify clusters of arbitrary shapes, and has strong applicability for the preliminary segmentation of Gaussian mapping point clouds of ancient buildings. Since the DBSCAN algorithm needs to perform neighborhood query and calculation for each point, its operation efficiency is low, and the time complexity of the DBSCAN algorithm is O(n ² ). For point clouds of ancient buildings with a large amount of data, the efficiency of the DBSCAN clustering algorithm is particularly low.

When the DBSCAN algorithm determines the (σ, MinPts) parameters, it needs to perform statistical calculations, draw a graph of the density and the number of points, and then determine the parameters through manual selection. Manual selection is obviously not suitable for automatic ancient building point cloud segmentation requirements.

The DBSCAN algorithm needs to judge whether each point in the neighborhood is a core point when a category grows, and it needs to calculate the number of points in the neighborhood of all points σ( _xi ). Although the MultyGrid-KD tree index can be established on the point cloud data after Gaussian mapping to speed up the judgment, the operation efficiency of the algorithm is still low.

The FDBSCAN algorithm proposed by Zhou Shuigeng et al. (2000) is an improved algorithm for DBSCAN. The FDBSCAN algorithm proposes to select a specified number of representative points (not all neighborhood points) in the neighborhood instead of category growth. The number of representative points is related to the dimension of the space, that is, the number of representative points in n-dimensional space is 2n, and the representative points here refer to the second-level dots, third-level dots, etc. .

As shown in Figure 2, the influence of the divergence of the four representative points in the two-dimensional space on the expansion effect of row clustering is demonstrated. When the divergence of the representative points is not good, there is a situation that a certain point and _xi density can be reached, but this point cannot be searched through the neighborhood search of the representative points. In Figure 2, point A points to x _i , the circle pointed to by 1 is the area range of σ( _xi ), point B is a representative point selected from σ( _xi ), that is, the center of the second-level circle, and the circle pointed to by 2 represents The point is the range of the neighborhood area of the second-level circle center, and point C is the representative point further selected from the neighborhood representing point B for cluster expansion, that is, the third-level circle center. In this figure, since the representative points are all concentrated in the upper area of the neighborhood, the expansion direction of the clustering is all upward, and the lower point objects cannot be classified into the same category.

In view of the above problems, on the one hand, it is necessary to merge these points and related subcategories formed later. On the other hand, in the selection of representative points, both the speed of the algorithm and the diffusion of representative points should be taken into account. Zhou Shuigeng et al. (2000) proposed two algorithms for selecting representative points, but they are not suitable for dense point cloud data on the Gaussian sphere in terms of algorithm efficiency and divergence of representative points.

Compared with DBSCAN and FDBSCAN algorithms, the improvement of AQ-DBSCAN algorithm is mainly reflected in two points: one is to provide automatic estimation of σ under the premise of given MinPts parameters, and the other is to realize a faster density clustering algorithm.

Compared with the FDBSCAN algorithm, the AQ-DBSCAN search algorithm not only sets the number of representative points to four according to the two-dimensional characteristics of the data on the Gaussian spherical surface, thereby reducing the amount of calculation, but also proposes a representative point cloud data suitable for dense point cloud data. The point selection algorithm improves the efficiency of representative point selection on the basis of ensuring the spread of representative points.

Description of drawings

Fig. 1 is in the present invention with difference _m as abscissa, with the number N of points in each circle as the plotted curve formed by ordinate fitting;

Figure 2 is a graph showing the influence of the divergence of the four representative points in the two-dimensional space on the effect of row clustering expansion in the FDBSCAN algorithm;

Fig. 3 (a) is the point cloud data figure in one embodiment of the present invention;

Figure 3(b) is a Gaussian map of the point cloud data corresponding to Figure 3(a);

Fig. 3 (c) is the N _m curve figure of automatic statistics in one embodiment of the present invention;

Fig. 3 (d) is the clustering effect figure on the Gaussian sphere that utilizes AQ-DBSCAN algorithm to obtain in one embodiment of the present invention;

Figure 3 (e) is a clustering effect diagram corresponding to a space surface in one embodiment of the present invention;

Fig. 3 (f) is the clustering effect figure on the Gaussian sphere that utilizes DBSCAN algorithm to obtain in one embodiment of the present invention;

Fig. 3 (g) utilizes AQ-DBSCAN algorithm in an embodiment of the present invention and then carries out the next step after the processing to the segmentation result figure of the column point cloud data outside the Hall of Supreme Harmony;

Fig. 4 (a) is the point cloud data figure in another embodiment in the present invention;

Figure 4(b) is a Gaussian map of the point cloud data corresponding to Figure 4(a);

Fig. 4 (c) is the N _m curve figure of automatic statistics in another embodiment of the present invention;

Fig. 4 (d) is the clustering effect figure on the Gaussian sphere that utilizes AQ-DBSCAN algorithm to obtain in another embodiment of the present invention;

Fig. 4 (e) is in another embodiment of the present invention, corresponds to the clustering effect diagram on the space surface;

Fig. 4 (f) is the clustering effect figure on the Gaussian sphere that utilizes DBSCAN algorithm to obtain in another embodiment of the present invention;

Fig. 4 (g) utilizes AQ-DBSCAN algorithm in another embodiment of the present invention to carry out the segmentation result diagram of the door beam point cloud data outside the Hall of Supreme Harmony after the next step of processing again;

Fig. 5 (a) is the point cloud data figure in another embodiment among the present invention;

Figure 5(b) is a Gaussian map of the point cloud data corresponding to Figure 5(a);

Fig. 5 (c) is the N _m curve figure of automatic statistics in another embodiment of the present invention;

Fig. 5 (d) is the clustering effect figure on the Gaussian sphere that utilizes AQ-DBSCAN algorithm to obtain in another embodiment of the present invention;

Fig. 5 (e) is a clustering effect diagram corresponding to the space surface in another embodiment of the present invention;

Fig. 5 (f) is the clustering effect figure on the Gaussian sphere that utilizes DBSCAN algorithm to obtain in another embodiment of the present invention;

Fig. 5 (g) is the segmentation result figure to the section point cloud data of Gate of Supreme Harmony after utilizing AQ-DBSCAN algorithm to carry out the processing of next step again in another embodiment of the present invention;

Fig. 6 is the time-consuming comparison diagram of the algorithm of DBSCAN and AQ-DBSCAN of the present invention;

Fig. 7 is a time-consuming comparison diagram of the algorithms of FDBSCAN and AQ-DBSCAN of the present invention.

Detailed ways

The present invention will be further described in detail below in conjunction with the accompanying drawings, so that those skilled in the art can implement it with reference to the description.

As shown in Figure 1, the present invention provides a kind of clustering method of building laser scanning point cloud data, comprising:

Step 1. Map each point in the 3D point cloud data obtained after laser scanning the building to a Gaussian sphere and convert it into 2D data, and obtain a set X of 2D data of the point cloud;

Step 2: Specify the density threshold minimum number of included points MinPts, for any point in the set X, calculate the farthest distance of the objects with the smallest number of included points MinPts closest to the point, and count the farthest distances of all points in the set X Maximum value and minimum value; in the division of two-dimensional data, MinPts is generally taken as 4.

Step 3: Divide the difference m between the maximum value and the minimum value of the farthest distance into n equal parts, take the point in the point cloud that produces the minimum value of the farthest distance as the center, and use the equal distance m/n as the unit step by step The incremental value is the radius, make n circles, and calculate the number of points in each circle;

Step 4. Use the difference m as the abscissa, the equal distance m/n as the incremental unit interval of the abscissa, and the number N _m of points in each circle as the ordinate, draw a coordinate map, and fit the coordinate map The points above form the drawn curve shown in Figure 1; in Figure 1, due to the difference in the farthest distance between the plane and the cylinder on the Gaussian sphere, the curve shows slight undulations in the process of extending upward .

Step 5. Find the inflection point of the drawn curve, the inflection point is the point with the largest slope change, and the value of the abscissa corresponding to the inflection point is taken as the value of the neighborhood radius σ. The method of finding the inflection point is: connect the starting point and the end point of the curve shown in Figure 1 to form a straight line L, calculate the distance H from each point on the curve to the straight line L, and count the point with the largest H as the inflection point.

The algorithm steps for the automatic statistics of the value of the neighborhood radius σ are as follows:

Step 6. Establish the AQ-DBSCAN algorithm based on the density threshold MinPts and the neighborhood radius σ, and use the AQ-DBSCAN algorithm to cluster the points in the set X to obtain which building the points in the point cloud belong to Cluster analysis of parts.

In this step six include:

6.1) Take the point that produces the minimum value of the farthest distance in the point cloud as the center of the circle, and use the neighborhood radius σ as the radius to draw the first-level circle. If the number of points in the first-level circle is less than MinPts, eliminate If the first-level circle is larger than MinPts; continue to 6.2), that is, select representative points to diverge points.

6.2) With cσ where 0<c<1 is the radius, draw the first-level sub-circle. If the number of points in the annular area between the first-level circle and the first-level sub-circle is less than MinPts, select the annular area All the points in are used as the second-level circle center; if the number of points in the annular area is greater than or equal to MinPts, select MinPts points as the second-level circle center. Among them, the method of selecting MinPts points as the second-level circle center is:

In the circular area, first select the point farthest from the center of the first level as the first point, then select the point farthest from the first point as the second point, and then select the distance between the first point and the second point. The farthest point of the sum of the distances of the two points is taken as the third point, and follow this rule until the MinPts point is selected.

That is, let x ₀ be an object in the point set X, let 0<c<σ, then

c(x ₀ )={x∈σ(x ₀ )|ε<D(χ, x ₀ )≤σ}

c(x ₀ ) is the candidate point set of the second-level circle center (ie representative point) in the neighborhood of x ₀ . c(x ₀ ) is a ring-shaped region centered on x ₀ and located between c and σ. Selecting the second-level circle center (representative point) in the peripheral area of σ(x ₀ ) is helpful to enhance the diffuseness of representatives and reduce the frequency of neighborhood queries during category expansion. The closer c is to σ, the fewer the number of candidate points and the higher the efficiency of searching for representative points, but at the same time it may lead to poor divergence of points. According to the characteristics of point cloud data of ancient buildings, considering the factors of diffusion and efficiency, the value of c in this algorithm is 3σ/4.

6.3) Draw at least one second-level circle with the selected second-level circle center as the center and the neighborhood radius σ as the radius. If the number of points in any second-level circle is less than MinPts, eliminate the second-level circle. If the secondary circle is greater than MinPts; then repeat steps 6.2) and 6.3), and draw circles step by step until the drawn circles are eliminated;

6.4) Gather the points in the circles of all levels into one class, and remove them from the point cloud, and repeat the operations from step 2 to step 6 for the points in the remaining point cloud.

In the clustering method of the described building laser scanning point cloud data, in the described step 6.2), the method of selecting MinPts points as the second level of circle center is:

In this circular area, first select the point farthest from the center of the first level as the first point P1, then select the point farthest from the first point as the second point, and then select the distance from the first point and The point with the farthest sum of the distances of the second point is taken as the third point, and follow this rule until the MinPts point is selected.

That is: suppose the existing representative point set is P, and the representative point to be searched is p _k , then

${\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; max</span>}}_{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; c</span>}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; x</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}<span\; class="notranslate">\; )</span>}\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; )</span>$

in

$\mathrm{<span\; class="notranslate">\; D.</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; min</span>}}_{{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; P</span>}}{<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; p</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; |</span><span\; class="notranslate">\; |</span>}_{\mathrm{<span\; class="notranslate">\; 2</span>}}<span\; class="notranslate">\; ,</span>$

The following is the specific description of the second-level circle center selection algorithm:

The complete algorithm of AQ-DBSCAN of the present invention is as follows:

Some examples of clustering of point clouds of buildings obtained by the method of the present invention are listed below. The data in the present invention comes from the point cloud acquired by the ground laser scanner of the Gate of Supreme Harmony in the Forbidden City and the Hall of Supreme Harmony.

In one embodiment, as shown in FIG. 3 , the present invention collects and clusters the column data outside the Hall of Supreme Harmony. Figure 3(a) is the point cloud data of the pillars outside the Hall of Supreme Harmony. The number of points sampled in this point cloud is 141999. Figure 3(b) is a Gaussian map corresponding to this column. Figure 3(c) is the N _m curve of the automatic statistics of the AQ-DBSCAN algorithm and the estimated σ parameter value (0.067831). DBSCAN and FDBSCAN algorithms mainly rely on manual judgment of the shape of the N _m curve and the change point of the extended trend in parameter estimation. It can be seen from Figure 3(c) that the AQ-DBSCAN algorithm gives the result that the σ value is in line with manual judgment. Figure 3(d) is the clustering effect diagram on the Gaussian sphere. Because the cylinders and some appendages overlap on the Gaussian sphere, they are clustered into one category, and these will be further divided through the analysis of the overlapping area in the future. Figure 3(e) is a clustering effect map corresponding to the space surface. Figure 3(f) is the clustering obtained on the Gaussian sphere using the existing DBSCAN algorithm. It can be seen from the Gaussian sphere that the clustering effects obtained by using the AQ-DBSCAN algorithm and the DBSCAN algorithm are completely consistent. Fig. 3 (g) is to utilize the method of the present invention to carry out the segmentation result diagram of the column point cloud outside the Hall of Supreme Harmony after the next step of processing again, as can be seen, after utilizing the method of the present invention, to the data of ancient building point cloud Segmentation can achieve better applicability.

In another embodiment, as shown in FIG. 4 , the present invention collects and clusters the data of the door beams outside the Hall of Supreme Harmony. Figure 4(a) is the point cloud data of the beams on the pillars outside the Hall of Supreme Harmony. The number of points sampled in this point cloud is 642984. Figure 4(b) is the corresponding Gaussian map. Figure 4(c) is the N _m curve of the automatic statistics of the AQ-DBSCAN algorithm and the estimated σ parameter value (0.011236). It can be seen from Fig. 3(c) that the AQ-DBSCAN algorithm gives the result that the σ value conforms to the manual judgment. Figure 4(d) is the clustering effect diagram of 4(a) on the Gaussian sphere, and Figure 4(e) is the clustering effect diagram corresponding to the space surface. It can be seen from Figure 4(e) that AQ- The DBSCAN clustering differentiates the front and sides of the beam. Figure 4(f) is the clustering obtained by using the existing DBSCAN algorithm on the Gaussian sphere. It can be seen from the Gaussian sphere that the clustering effects obtained by using the AQ-DBSCAN algorithm and the DBSCAN algorithm are completely consistent. Fig. 4 (g) is to utilize the method of the present invention to carry out the segmentation result figure of the door beam point cloud outside the Hall of Supreme Harmony after the next step of processing again, as can be seen, after utilizing the method of the present invention, to the ancient building point cloud Data segmentation can achieve better applicability.

In yet another embodiment, as shown in FIG. 5 , the present invention collects part of the data of the Gate of Supreme Harmony, including a column and its associated beams. Figure 5(a) includes the point cloud data of the column and the beam. The number of points sampled in this point cloud is 5771. Figure 5(b) is the corresponding Gaussian map. Figure 5(c) is the N _m curve of the automatic statistics of the AQ-DBSCAN algorithm and the estimated σ parameter value (0.014331). It can also be seen from the figure that the AQ-DBSCAN algorithm gives the σ value in line with the result of manual judgment. Figure 5(d) is a clustering effect diagram on the Gaussian sphere. Since the cylinder and the beam overlap on the Gaussian sphere, they are also clustered into one category. Figure 5(e) is a clustering effect map corresponding to the space surface. Figure 5(f) is the clustering obtained by using the existing DBSCAN algorithm on the Gaussian sphere. It can be seen from the Gaussian sphere that the clustering effects obtained by using the AQ-DBSCAN algorithm and the DBSCAN algorithm are completely consistent. Fig. 5 (g) is to utilize the method of the present invention to carry out the segmentation result figure of the point cloud of the door of Supreme Harmony part data after the next step of processing again, as can be seen, after utilizing the method of the present invention, to the ancient building point cloud Data segmentation can achieve better applicability.

Compared with the DBSCAN algorithm, the method of the present invention uses AQ-DBSCAN to expand the region, and the clustering speed is greatly improved under the premise of keeping the algorithm effect consistent.

The algorithm time consumption of the above three embodiments is shown in Table 1:

Table 1 Time-consuming of DBSCAN, FDBSCAN, AQ-DBSCAN clustering algorithms

In Figure 6 and Figure 7, the time-consuming curve is made with time as the abscissa and the number of processed points as the total coordinate. In Figure 6, curve A is the time-consuming of the DBSCAN algorithm, and curve B is the time-consuming of the AQ-DBSCAN algorithm , in Fig. 7, the curve C is the time consumption of the FDBSCAN algorithm, and the curve D is the time consumption of the AQ-DBSCAN algorithm. As shown in Figure 6 and Figure 7 and Table 1, it can be seen from the time-consuming comparison that the time-consuming of the AQ-DBSCAN algorithm is significantly less than that of the DBSCAN and FDBSCAN algorithms. The more points, the more obvious the acceleration of AQ-DBSCAN algorithm. For example, for 141999 points, DBSCAN takes 6 times as long as AQ-DBSCAN, and for 642984 points, it takes 65 times.

Due to the particularity of point clouds of ancient buildings, the general neighborhood segmentation method is often inapplicable when applied. In response to these problems, the present invention proposes a clustering method suitable for point clouds of ancient buildings. After the point cloud data is mapped onto a Gaussian sphere and transformed into two-dimensional data, the mapped point sets are clustered, and the core point set is screened out. , laying the foundation for later feature extraction and segmentation.

The invention makes a beneficial exploration for the division, storage and expression of all large buildings, and has important theoretical value and social significance.

Although the embodiment of the present invention has been disclosed as above, it is not limited to the use listed in the specification and implementation, it can be applied to various fields suitable for the present invention, and it can be easily understood by those skilled in the art Therefore, the invention is not limited to the specific details and examples shown and described herein without departing from the general concept defined by the claims and their equivalents.

Claims (7)

1. A clustering method of building laser scanning point cloud data, is characterized in that, comprises: Step 1. Convert each point in the three-dimensional point cloud data obtained after laser scanning the building into two-dimensional data, and obtain a set X of two-dimensional data of the point cloud; Step 2: Specify the density threshold minimum number of included points MinPts, for any point in the set X, calculate the farthest distance of the objects with the smallest number of included points MinPts closest to the point, and count the farthest distances of all points in the set X maximum and minimum values; Step 3: Divide the difference m between the maximum value and the minimum value of the farthest distance into n equal parts, take the point in the point cloud that produces the minimum value of the farthest distance as the center, and use the equal distance m/n as the unit step by step The incremental value is the radius, make n circles, and calculate the number of points in each circle; Step 4. Use the difference m as the abscissa, the equal distance m/n as the incremental unit interval of the abscissa, and the number of points in each circle as the ordinate, draw a coordinate map, and fit the coordinates on the map points to form a drawn curve; Step 5. Find the inflection point of the drawn curve, and use the value of the abscissa corresponding to the inflection point as the value of the neighborhood radius σ; Step 6: Establish the AQ-DBSCAN algorithm based on the density threshold MinPts and the neighborhood radius σ, and use the AQ-DBSCAN algorithm to cluster the points in the set X to obtain which building the points in the point cloud belong to Cluster analysis of parts. 2. the clustering method of building laser scanning point cloud data as claimed in claim 1, is characterized in that, described step 6 comprises: 6.1) Take the point that produces the minimum value of the farthest distance in the point cloud as the center of the circle, and use the neighborhood radius σ as the radius to draw the first-level circle. If the number of points in the first-level circle is less than MinPts, eliminate If the first-level circle is greater than MinPts; continue to 6.2) 6.2) With cσ where 0<c<1 is the radius, draw the first-level sub-circle. If the number of points in the annular area between the first-level circle and the first-level sub-circle is less than MinPts, select the annular area All the points in are used as the second-level circle center; if the number of points in the annular area is greater than or equal to MinPts, select MinPts points as the second-level circle center; 6.3) Draw at least one second-level circle with the selected second-level circle center as the center and the neighborhood radius σ as the radius. If the number of points in any second-level circle is less than MinPts, eliminate the first-level circle If the secondary circle is greater than MinPts; then repeat steps 6.2) and 6.3), and draw circles step by step until the drawn circles are eliminated; 6.4) Gather the points in the circles of all levels into one class, and remove them from the point cloud, and repeat steps 2 to 6 for the remaining points in the point cloud. 3. the clustering method of building laser scanning point cloud data as claimed in claim 2, it is characterized in that, in described step 6.2), the method of choosing MinPts points as the second level of circle center is: In the circular area, first select the point farthest from the center of the first level as the first point, then select the point farthest from the first point as the second point, and then select the distance between the first point and the second point. The farthest point of the sum of the distances of the two points is taken as the third point, and follow this rule until the MinPts point is selected. 4. the clustering method of building laser scanning point cloud data as claimed in claim 1 or 3, is characterized in that, described MinPts is 4. 5. the clustering method of building laser scanning point cloud data as claimed in claim 4, is characterized in that, described c is 3/4. 6. The clustering method of building laser scanning point cloud data as claimed in claim 1, is characterized in that, described inflection point is the point that slope changes the most. 7. the clustering method of the point cloud based on AQ-DBSCAN algorithm as claimed in claim 1, is characterized in that, in described step 1, each point in the three-dimensional point cloud data that building is carried out laser scanning and obtains Both are mapped to a Gaussian sphere to be transformed into two-dimensional data.