Structural plane recognition method based on random forest and dynamic DBSCAN algorithm

Abstract

The present invention discloses a structural plane recognition method based on a random forest and a dynamic DBSCAN algorithm, which solves the drawbacks that the current structural plane recognition operation is cumbersome and inconvenient, easily causes the loss of real information and has structural plane undulation errors. The key points of the technical solution are to collect the original point cloud data of the structural plane by using a three-dimensional laser scanner, calculate and extract the multi-scale spatial feature values of the point cloud, classify the point cloud based on the random forest classification algorithm, divide all the point clouds in the area into plane points and intersection line points, set the neighbor retrieval quantity for the original point cloud data, and determine the neighborhood range of each point cloud through the nearest neighbor algorithm. Finally, the point clouds of different structural planes are segmented by the dynamic DBSCAN algorithm to complete the recognition of the structural plane. The structural plane recognition method based on a random forest and a dynamic DBSCAN algorithm of the present invention does not require resampling and meshing, effectively retains the structural plane information, and is more accurate and convenient.

Description

Technical Field

[0001] The present invention relates to the technology for identifying rock mass structural planes, and particularly to a method for identifying structural planes based on a random forest and a dynamic DBSCAN algorithm. Background Art

[0002] In rock mechanics calculations and stability analyses, the characteristics and distribution patterns of rock mass structural planes have an important influence on the stress conduction of rock masses. Therefore, it is necessary to conduct investigations and statistics on the structural planes.

[0003] Since the traditional manual measurement method using a compass and a tape measure has extremely low efficiency, limited coverage, and high danger, it cannot comprehensively reflect the characteristics of rock mass structural planes. It is necessary to perform three-dimensional reconstruction of the rock mass through three-dimensional laser scanning. The collected three-dimensional point cloud is point cloud coordinate data without any identification. Therefore, it is necessary to segment the point clouds belonging to different structural planes. In the past, it was often necessary to construct evenly distributed point cloud data through resampling, grid the point cloud, and perform clustering analysis based on the grid attitude, which often caused the loss of real information and errors caused by the undulation of the structural plane. Summary of the Invention

[0004] The purpose of the present invention is to provide a method for identifying structural planes based on a random forest and a dynamic DBSCAN algorithm, which does not require resampling and meshing, effectively retains the structural plane information, and is more accurate and convenient.

[0005] The above technical purpose of the present invention is achieved through the following technical solutions:

[0006] A method for identifying structural planes based on a random forest and a dynamic DBSCAN algorithm includes the following steps:

[0007] Collect and obtain the original point cloud data of the structural plane by using a three-dimensional laser scanner, and calculate and extract the multi-scale spatial feature values of the point cloud;

[0008] Perform point cloud classification based on the random forest classification algorithm, and divide all the point clouds in the area into plane points and intersection line points;

[0009] Set the neighbor search quantity for the original point cloud data, and determine the neighborhood range of each point cloud through the nearest neighbor algorithm;

[0010] Segment the point clouds of different structural planes through the dynamic DBSCAN algorithm to complete the identification of the structural planes.

[0011] In summary, the present invention has the following beneficial effects:

[0012] By combining the random forest with the dynamic DBSCAN algorithm, it is possible to directly identify the structural planes using the original data without resampling and meshing conditions, maximizing the retention of the surface features of the structural planes and having good application prospects in rock mass structure analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a schematic flow diagram of the present method;

[0014] Figure 2 It is the effect diagram of the present method in the example;

[0015] Figure 3 It is the data diagram of a single structural plane segmented from the point cloud data in the example. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0016] The present invention will be further described in detail below with reference to the accompanying drawings.

[0017] According to one or more embodiments, a method for identifying structural planes based on the random forest and dynamic DBSCAN algorithms is disclosed. As Figure 1 shown, it includes the following steps:

[0018] S1. Use a three-dimensional laser scanner to collect the original point cloud data of the structural plane, and calculate and extract the multi-scale spatial feature values of the point cloud;

[0019] S2. Based on the random forest classification algorithm, classify the point cloud, and divide all the point clouds in the area into plane points and intersection line points;

[0020] S3. Set the number of neighboring points retrieved for the original point cloud data, and determine the neighborhood range of each point cloud through the nearest neighbor algorithm;

[0021] S4. Segment the point clouds of different structural planes through the dynamic DBSCAN algorithm (dynamic clustering algorithm) to complete the identification of the structural planes.

[0022] The extraction of the multi-scale spatial feature values of the point cloud in step S1 is specifically as follows:

[0023] By calculating the distance between the nearest neighbor points of the point cloud, determine the distribution characteristics of the nearest neighbor points, delete the outlier points according to the distribution characteristics, and determine the minimum calculation space scale, denoted as r0;

[0024] For each point in the point cloud data, extract the subspace point cloud data (D1, D2, D3... D n within the neighborhood range of R1, R2, R3... R n ), calculate its covariance matrix (C1, C2, C3... C n ), and obtain the eigenvalues of each matrix arranged from largest to smallest after normalization, forming 3×n eigenvalues.

[0025] In S2, the random forest classification algorithm constructs a training set, assigns the training samples to two categories, i.e., structural surface points and boundary points, through manual labeling, trains the random forest classifier until convergence, and then inputs the calculated eigenvalues into the classifier for classification, thereby obtaining the categories of the point cloud.

[0026] In S3, let the original point cloud dataset be A, the number of point clouds be N, and the set number of neighboring points retrieved be ne. Through the nearest neighbor algorithm, determine that the retrieval radii of each point cloud are (ρ1, ρ2, ρ 3……… ρ N )

[0027] In S4, the segmentation of the point clouds of different structural surfaces by the dynamic dbscan algorithm is specifically as follows:

[0028] S41. Through point cloud classification, delete the point cloud categories of the boundary points in the original point cloud dataset A to obtain the set A1.

[0029] S42. Arbitrarily select a point p i in the set A1 as the starting point for retrieval, and search for its neighboring points within the retrieval radius ρ i ; if the number of neighboring points is equal to ne, then consider the point p i as an internal point of the structural surface, put the point p i into the single structural surface set J, and put its neighboring points into the set I of points to be retrieved; if it is less than ne, then consider the point p i as a boundary point of the structural surface and do not perform any operation, and randomly select a point again.

[0030] S43. Arbitrarily select a point, p j , in the set I, calculate its neighboring points within the retrieval radius ρ j ; if the number of neighboring points is equal to ne, then consider the point as an internal point of the structural surface, put the point into the single structural surface set J, and put its neighboring points into the set I of points to be retrieved; if the number of neighboring points is less than ne, then consider the point as a boundary point of the structural surface, put the point into the single structural surface set J, and its neighboring points are not recorded in the set to be retrieved.

[0031] S44. Repeat step S43 to calculate and judge the points in the set I until all points are traversed and there is no growth, then the points in the set J are considered to be points of the same structural surface.

[0032] S45. Repeat step S42 to select and judge the points in the set A1 until all points in A1 are traversed, and complete the segmentation of the point cloud.

[0033] For clarity, an example is given below:

[0034] For example Figure 2As shown, by setting two intersecting planes indoors, the original point cloud data is obtained using a three-dimensional laser scanner as shown in (a) of Figure 2 , and the eigenvalues in its multi-scale space are calculated. The classification between structural surface points, i.e., plane points and boundary points, is achieved through a random forest classifier, as shown in (b) of Figure 2 . Subsequently, the segmentation between different plane points is achieved through the dbscan algorithm, and the segmentation result is obtained as shown in (c) of Figure 2 . As shown in Figure 3 , it is a single structural surface segmented from the point cloud data, which is verified through practical applications. It can be seen that the segmentation of the surface points of the rock mass structure can be completed even under the conditions of uneven structural surfaces and local occlusion.

[0035] This specific embodiment is only an explanation of the present invention, and it is not a limitation of the present invention. Those skilled in the art can make modifications to this embodiment without creative contributions according to needs after reading this specification, but as long as it is within the scope of the claims of the present invention, it is protected by the patent law.

Claims

1. A structural plane recognition method based on the random forest and dynamic DBSCAN algorithms, characterized in that It includes the following steps: Use a three-dimensional laser scanner to collect the original point cloud data of the structural plane, and calculate and extract the multi-scale spatial eigenvalues of the point cloud. Specifically, the calculation and extraction of the multi-scale spatial eigenvalues of the point cloud are as follows: Calculate the distance between the nearest neighbor points of the point cloud, determine the distribution characteristics of the nearest neighbor points, delete the outlier points according to the distribution characteristics, and determine the minimum calculation spatial scale, denoted as r0; Extract the subspace point cloud data (D1, D2, D3... Dn) within the R1, R2, R3... Rn neighborhood ranges for each point in the point cloud data, calculate its covariance matrix (C1, C2, C3... Cn), and obtain the eigenvalues arranged from largest to smallest after normalization for each matrix, forming 3×n eigenvalues; Based on the random forest classification algorithm, perform point cloud classification to divide all the point clouds in the area into plane points and boundary line points; Set the neighbor retrieval amount for the original point cloud data, and determine the neighborhood range of each point cloud through the nearest neighbor algorithm; Use the dynamic dbscan algorithm to segment the point clouds of different structural planes to complete the recognition of the structural planes.

2. The structural plane recognition method based on the random forest and dynamic DBSCAN algorithm according to claim 1, characterized in that, The specific process of using the random forest classification algorithm for point cloud classification is as follows: Construct a training set, assign the training samples to two categories, i.e., structural plane points and boundary line points, through manual marking, and train the random forest classifier until convergence; Input the calculated eigenvalues into the classifier for classification to obtain the category of the point cloud.

3. The structural plane recognition method based on the random forest and dynamic DBSCAN algorithms according to claim 1 is characterized in that: Let The original point cloud dataset is A, the number of point clouds is N, the set number of neighboring points to be retrieved is ne, and through the nearest neighbor algorithm, the retrieval radii of each point cloud are determined to be (ρ1, ρ2, ρ 3……… ρ N ) when the number of neighboring points is ne.

4. The structural plane recognition method based on the random forest and dynamic DBSCAN algorithms according to claim 3, characterized in that The specific process of using the dynamic dbscan algorithm to segment the point clouds of different structural planes is as follows: Through point cloud classification, delete the point cloud category of the boundary line points in the original point cloud data set A to obtain set A1; Select an arbitrary point p in set A1 i As the retrieval starting point, search for adjacent points within its retrieval radius ρ in A1 i ; If the number of adjacent points is equal to ne, then the point p is considered i to be an internal point of the structural plane, and the point p i is put into the single structural plane set J, and its adjacent points are put into the set I of points to be retrieved; if it is less than ne, then the point p is considered i to be a boundary point of the structural plane and no operation is performed, and a point is randomly selected again; Randomly select a point, p, in the set I j , and calculate its search radius ρ j For the adjacent points in, if the number of adjacent points is equal to ne, then this point is considered an internal point of the structural plane. Put this point into the single structural plane set J, and put its adjacent points into the set I of points to be searched; if the number of adjacent points is less than ne, then this point is considered a boundary point of the structural plane. Put this point into the single structural plane set J, and its adjacent points are not recorded in the set to be searched; Repeat the calculation and judgment of the points in set I until all points are traversed and there is no growth. Then, the points in set J are considered to be points of the same structural plane; Repeat the selection and judgment of the points in set A1 until all the points in A1 are traversed to complete the segmentation of the point cloud.

Images