Multi-Type Structural Plane Hierarchical Recognition Method Based on Point Cloud Data

Abstract

The present invention provides a multi-type structural plane hierarchical recognition method based on point cloud data. Step 1: Perform three-dimensional laser scanning on the slope rock mass to obtain the spatial coordinate information of the point cloud of the slope rock mass to be processed at the observation site. Step 2: Complete the voxel downsampling of the original data based on KDtree nearest neighbor search to obtain the preliminarily processed point cloud. Step 3: Use the random Hough transform space to calculate the normal vector of the primary processed point cloud data, and use PCA principal component analysis to calculate the corresponding curvature feature to obtain the secondary processed point cloud. Step 4: Using the normal vector difference and curvature threshold as limiting conditions, use the region growing algorithm to cluster and group the secondary processed point cloud to obtain the clustering plane, and use RGB random coloring to obtain the recognition result of the bedding fault type structural plane. Step 5: Use the Markov random field point neighborhood feature prediction method to calculate the surface features of the primary processed point cloud, and use HSV coloring to distinguish to obtain the recognition result of the joint fracture type structural plane.

Description

Technical Field

[0001] The present invention relates to the technical field of using 3D laser scanning to identify and extract rock mass structural planes, and particularly to an identification and extraction technology for two different types of structural planes, namely bedding planes and joints, of reservoir bank rock masses. Background Art

[0002] The rock mass structural plane is one of the important indicators for evaluating the quality of rock masses. The quality of rock masses depends on the internal factors that constitute the structural characteristics of the rock masses, and the rock mass structural plane is one of the main controlling factors. Therefore, how to accurately and quickly obtain the morphological characteristics of different types of structural planes is of great significance for rock engineering construction.

[0003] The extraction of structural plane morphological information is inseparable from the accurate identification of structural planes. There are many methods for identifying rock mass structural planes, including image recognition, video recognition, and 3D laser scanning technology. Among them, 3D laser scanning technology is the most widely used and representative in recent years. This technology obtains structural plane information by secondary processing of the point cloud information of the rock mass structural plane. Every step from data preprocessing to structural plane recognition and then to subsequent information extraction is very important. The present invention mainly focuses on the research of data preprocessing and structural plane recognition.

[0004] Patent CN 111553292 A proposes a point cloud recognition based on triangular meshes; Patent CN 106896213 B proposes a point cloud recognition and extraction method, which uses the least squares method with fixed weights to calculate the point cloud normal vector and performs region growing based on the normal angle to identify the structural plane; Patent CN 112529844 A is based on the spherical K-means clustering method of the normal vector cosine of the random Hough space transformation to identify the structural plane; the above methods mention very little about the data preprocessing method, do not consider the influence of preprocessing on the subsequent data processing, and the least squares method with fixed weights still has the commonality of this method, that is, it has a smoothing effect on the sharpness of the data and does not perform well at the sharp intersections. The selection of the value of K in K-means clustering depends very much on experience and is very sensitive to the initial value and outliers. Compared with region growing, the use of feature classification has greater limitations and higher difficulty. Therefore, there are the following deficiencies in the secondary processing of the recognition of rock mass structural plane point cloud data at present:

[0005] The preprocessing of structural plane point clouds is mentioned very little, and the rationality of data reduction and the necessity of improving the speed and accuracy for subsequent calculations are not considered; Structural planes with large differences in appearance such as bedding planes and faults and structural planes with small differences such as joint fissures are structural planes of different scales, and it is difficult to identify multiple types with the same method without distinction; The method has strong artificial experience limitations and high usage difficulty. Summary of the Invention

[0006] In view of the above problems, the object of the present invention is to propose a multi-type structural plane hierarchical recognition method based on point cloud data. This method uses voxel downsampling for nearest neighbor search, which is a data reduction method that can significantly reduce the data volume under the condition of ensuring data integrity, reduce the number of abnormal points while retaining data features, and speed up the subsequent data processing speed.

[0007] In order to achieve the above technical features, the object of the present invention is realized as follows: A multi-type structural plane hierarchical recognition method based on point cloud data, comprising the following steps:

[0008] Step 1: Perform three-dimensional laser scanning on the slope rock mass to obtain the spatial coordinate information of the slope rock mass point cloud to be processed at the observation site;

[0009] Step 2: Complete voxel downsampling of the original data based on KDtree nearest neighbor search to obtain preliminarily processed point cloud;

[0010] Step 3: Use random Hough transform space to calculate the normal vector of the preliminarily processed point cloud data, and use PCA principal component analysis to calculate the corresponding curvature feature to obtain the secondarily processed point cloud;

[0011] Step 4: Using the normal vector difference and curvature threshold as limiting conditions, use the region growing algorithm to cluster and group the secondarily processed point cloud to obtain a clustering plane, and use RGB random coloring to obtain the recognition result of bedding fault type structural planes;

[0012] Step 5: Use the Markov random field point neighborhood feature prediction method to calculate the surface features of the preliminarily processed point cloud, and use HSV coloring to distinguish to obtain the recognition result of joint fracture type structural planes.

[0013] The specific processing process of Step 2 is as follows:

[0014] Step 2.1: Establish a KDtree tree data structure for the original point cloud data, count the total number N of the original point cloud, and set the lengths of this space in the x, y, and z directions to be L x , L y , L z ;

[0015] Step 2.2: Find the minimum distance i min of the point cloud. Taking the minimum distance as the unit length, divide the point cloud space into n = l × w × h small cubes, and set the initial unit cube size l = L x × i min , w = L y × i min , h = L z × i min ;

[0016] Step 2.3: Calculate the centroid point of each non-empty voxel as P centroid (x centroid ,y centroid ,z centroid ), and construct the downsampled data point set P c ;

[0017] Step 2.4: Use KDtree to traverse the centroid points P centroid ,and set K = 1 to perform the nearest neighbor search centered on P centroid ,and replace it with the nearest neighbor point to P centroid to form a new data set P i ,forming the sampled point cloud, and the total number of points in the point cloud is N1;

[0018] Step 2.5: Judge After testing, when the number of points exceeds 100,000 in data downsampling ,the situation of local feature loss will occur, and it is necessary to adjust and judge according to the actual data volume and detailed situation.

[0019] The specific processing process of the said Step 3 is as follows:

[0020] Step 3.1: Perform random Hough transform space to calculate the normal vector of the point cloud data and form a normal vector set;

[0021] Step 3.2: Use PCA principal component analysis to calculate the curvature feature. This curvature is not the curvature in the mathematical sense, but approximates the curvature information based on the idea of surface variation, and is to solve the eigenvalues of the covariance matrix formed by the data. The specific process is as follows:

[0022] Step 3.2.1: According to the KDtree data structure, set the value of the number of search neighborhood points K, and set the initial calculation point P i ;

[0023] Step 3.2.2: Construct a spatial point coordinate set P i around the neighborhood K of P k ;

[0024] Step 3.2.3: Calculate the covariance matrix of this data set P k and construct a covariance tensor;

[0025] Step 3.2.4: Calculate the eigenvalues based on the covariance tensor and sort them as λ 1i ≥λ 2i ≥λ 3i ≥0;

[0026] Step 3.2.5: Calculate the surface curvature of P i :

[0027]

[0028] Step 3.2.6: Loop through Steps 3.2.2 to 3.2.4 to calculate the curvature features of all points and form a curvature number set.

[0029] The specific processing procedure of Step 4 is as follows:

[0030] Step 4.1: Import the secondary point cloud data P i ={X i , Y i , Z i}, and form a data set N nor ={N xi , N yi , N zi , σ i} using the normal vector and surface curvature. Sort the point cloud according to the size of σ i and find σ i min and add it to the seed point set;

[0031] Step 4.2: Set the neighborhood K. Neighborhood points are added to the growing point set. For each seed point, calculate the difference in the normal angles of its neighborhood points. If the difference is less than the set threshold θ th , then this point is seriously considered and proceeds to the next judgment in Step 4.3;

[0032] Step 4.3: If the difference in the normal angles θ th test is passed, then perform the curvature threshold cth test. If it is less than the curvature threshold cth, then this point belongs to the current plane of this seed point;

[0033] Step 4.4: Set the minimum number of plane points C min , the maximum number of plane points C max , and remove the points that have passed the two tests from the growing point set;

[0034] Step 4.5: Repeat Steps 4.1 to 4.4, and use RGB random coloring to mark different colors for all clustered planes;

[0035] Step 4.6: Keep calculating until the number of planes that can be generated from the remaining points is less than C min then stop the calculation;

[0036] Step 4.7: Identify different structural planes through RGB color differentiation.

[0037] The specific processing procedure of Step 5 is as follows:

[0038] Step 5.1: According to the KDtree data structure, set the neighborhood range using the point cloud spacing at preprocessing, and set the initial calculation point P i ;

[0039] Step 5.2: Around P i Construct a spatial point coordinate dataset P for the neighborhood K k , calculate the covariance matrix of this dataset, and construct a covariance tensor:

[0040]

[0041] Step 5.3: Calculate the eigenvalues according to the covariance tensor and sort them as L1≥L2≥L3≥0, and define the eigenvectors e1, e2, e3 corresponding to the eigenvalues;

[0042] Step 5.4: Calculate the surface feature parameters of the center point P i , construct the geometric feature set of each point, and use a binary classifier to predict the contour score P(c i |X i ) of each point P i ), and the parameters that can best reflect the surface features are as follows:

[0043]

[0044]

[0045]

[0046]

[0047] Verticality = 1 - |<[0, 0, 1], e3>| (7)

[0048] Step 5.5: Set a voxel grid, score the geometric feature set of each voxel point using an adaptive threshold, and find the seed point P(c i = 0|X i ) with a high contour score, and perform a secondary judgment f;

[0049] Step 5.6: Use a multi-spanning tree to construct the neighborhood of the seed point, judge the correctness of its being a high contour score point by the distance to the neighborhood points, etc., and retain the point with the highest score in the feature set candidate set after passing the e and f detections;

[0050] Step 5.7: According to the statistics of the feature set scores and spatial position distributions, use a Markov random field to select continuous optimal feature points from the feature set candidate set to obtain a feature contour model;

[0051] Step 5.8: Loop through steps 5.1 - 5.7 until the point P i(i＝N) calculation is completed;

[0052] Step 5.9: Use HSV coloring to segment and identify surface features.

[0053] The present invention has the following beneficial effects:

[0054] 1. By using voxel downsampling with proximity search, the present invention provides a data reduction method that can significantly reduce the amount of data without distorting the data, reduce the number of outliers while preserving the data features, and speed up the subsequent data processing.

[0055] 2. For the original data, the present invention uses region growing with random Hough transform normal vector difference and curvature feature constraints, introduces a statistical "voting" algorithm, adds curvature constraint conditions, improves the recognition accuracy of region growing for sharp intersections between rock structural planes, and improves the normal vector calculation effect, such as J1 and J2 in specific examples.

[0056] 3. The present invention uses the Markov random field point neighborhood feature prediction method to identify surface joint fracture structural planes with small normal vector differences, increasing the comprehensiveness of structural plane identification.

[0057] 4. The present invention uses two methods to identify the same data to obtain multiple types of structural planes, improving the data utilization rate, simplifying the identification process of multiple types of structural planes, and improving the identification effect of three-dimensional point cloud data in the case of coexistence of multiple types of structural planes. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present invention will be further described below with reference to the drawings and embodiments.

[0059] Figure 1 It is a flow chart of the identification method of the present invention.

[0060] Figure 2 It is a diagram of the experimental simulation specimen in Embodiment 2 of the present invention.

[0061] Figure 3 It is the identification of bedding structural planes by region growing in Embodiment 2 of the present invention.

[0062] Figure 4 It is the identification of surface joint fractures by surface features in Embodiment 2 of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0063] The embodiments of the present invention will be further described below with reference to the drawings.

[0064] Embodiment 1:

[0065] A multi-type structural plane hierarchical identification method based on point cloud data includes the following steps:

[0066] Step 1: Perform three-dimensional laser scanning on the slope rock mass to obtain the spatial coordinate information of the point cloud of the slope rock mass to be processed at the observation site;

[0067] Step 2: Complete the voxel downsampling of the original data based on KDtree nearest neighbor search to obtain the preliminarily processed point cloud:

[0068] Step 2.1: Establish a KDtr ee tree data structure for the original point cloud data, count the total number N of the original point cloud, and set the lengths of this space in the x, y, and z directions to be L x , L y , L z ;

[0069] Step 2.2: Find the minimum point cloud spacing i min , and taking the minimum distance as the unit length, divide the point cloud space into n = l × w × h small cubes. Set the initial unit cube size l = L x ×i min , w = L y ×i min , h = L z ×i min ;

[0070] Step 2.3: Calculate the centroid point of each non-empty voxel as P centroid (x centroid , y centroid , z centroid ), and construct the downsampled data point set P c with all the centroid points;

[0071] Step 2.4: Traverse the centroid point P centroid using KDtree, and set K = 1 to perform the nearest neighbor search centered on P centroid , and replace it with the nearest neighbor point to P centroid to form a new data set P i , forming the sampled point cloud, and the total number of the point cloud is N1;

[0072] Step 2.5: Judge After testing, it is found that when the number of points exceeds during data downsampling for 100,000 points, local feature loss will occur. It is necessary to make adjustment judgments according to the actual data volume and detailed situation.

[0073] Among them, the KDtree nearest neighbor search is to establish a KDtree data structure for data point partitioning, realize the fast search and efficient iteration of data, and based on this function, search for the element with the smallest neighborhood distance of the target point; the voxel downsampling is to divide the data space into grids and reduce the data volume based on the data weight distribution within the grids; the K is the number of neighbors of the target point, and K = 1 represents the nearest neighbor search, that is, the data point with the closest distance.

[0074] Step 3: Calculate the normal vector of the primary processed point cloud data using the random Hough transform space, and calculate the corresponding curvature feature using PCA principal component analysis to obtain the secondary processed point cloud;

[0075] Step 3.1: Perform random Hough transform space calculation on the point cloud data to calculate the normal vector, and form a normal vector number set;

[0076] Step 3.2: Use PCA principal component analysis to calculate the curvature feature. This curvature is not the curvature in the mathematical sense, but approximates the curvature information based on the idea of surface variation, which is to solve the eigenvalues of the covariance matrix formed by the data. The specific process is as follows:

[0077] Step 3.2.1: According to the KDtree data structure, set the value of the number of search domain points K, and set the initial calculation point P i ;

[0078] Step 3.2.2: Build a spatial point coordinate number set P i around the neighborhood K of P k ;

[0079] Step 3.2.3: Calculate the covariance matrix of this data set P k and construct a covariance tensor;

[0080] Step 3.2.4: Calculate the eigenvalues based on the covariance tensor and sort them as λ 1i ≥ λ 2i ≥ λ 3i ≥ 0;

[0081] Step 3.2.5: Calculate the surface curvature of Pi:

[0082]

[0083] where σ i represents the curvature feature of any point data, λ 1i , λ 2i , λ 3i represents the eigenvalues of any point data, numbered from large to small.

[0084] Step 3.2.6: Loop through steps 3.2.2 to 3.2.4 to calculate the curvature features of all points and form a curvature number set.

[0085] Among them, the random Hough transform space is based on the corresponding relationship between two coordinate spaces, mapping a straight line or curve in the Cartesian coordinate system to a point in the Hough space, converting shape detection into single-point data statistics, so as to implement a statistical voting mechanism to improve the normal vector numerical calculation;

[0086] Step 4: Using the normal vector difference and the curvature threshold as limiting conditions, the region growing algorithm is used to cluster and group the secondary processed point cloud to obtain the clustering plane, and RGB random coloring is adopted to obtain the recognition result of the bedding fault type structural plane;

[0087] Step 4.1: Import the secondary point cloud data P i ={X i , Y i , Z i}, and use the normal vector and the surface curvature to form the data set N nor ={N xi , N yi , N zi , σ i}, sort the point cloud according to the size of σ i , find σ i min and add it to the seed point set;

[0088] Step 4.2: Set the neighborhood K, and the neighborhood points are added to the growing point set. For each seed point, calculate the normal angle difference of its neighborhood points. If the difference is less than the set threshold θ th , then this point is considered key and enters the next judgment in Step 4.3;

[0089] Step 4.3: If the normal angle difference θ th test is passed, then the curvature threshold cth test is performed. If it is less than the curvature threshold cth, then this point belongs to the current plane of this seed point;

[0090] Step 4.4: Set the minimum number of plane points C min , the maximum number of plane points C max , and remove the points that have passed the two tests from the growing point set;

[0091] Step 4.5: Repeat Step 4.1 to Step 4.4, and use RGB random coloring to mark different colors for all clustered planes;

[0092] Step 4.6: Keep calculating until the number of planes that can be generated from the remaining points is less than C min then stop the calculation;

[0093] Step 4.7: Identify different structural planes through RGB color differences.

[0094] Among them, the RGB random coloring is based on the combination of three colors, namely R, G, and B (red, green, and blue). The intensity range of each color is [0, 255], that is, R ∈ [0, 255], G ∈ [0, 255], B ∈ [0, 255]. Different intensity combinations of RGB are assigned to each point according to whether it belongs to the same seed point plane to achieve random coloring of different planes; σ i minis the minimum value of the curvature feature calculated in Step 3.2.5.

[0095] Step 5: Use the Markov random field point neighborhood feature prediction method to calculate the surface features of the primary processed point cloud, and use HSV coloring to distinguish, to obtain the recognition result of the joint fracture type structural plane:

[0096] Step 5.1: According to the KDtree data structure, set the neighborhood range using the point cloud spacing at the time of preprocessing, and set the initial calculation point P i ;

[0097] Step 5.2: Around P i Construct the spatial point coordinate set P k for the neighborhood K, calculate the covariance matrix of this data set, and construct the covariance tensor:

[0098]

[0099] where K is the number of neighborhood points of P i and P k is the neighborhood point set, and is the average neighborhood point set;

[0100] Step 5.3: Calculate the eigenvalues according to the covariance tensor and sort them as L1≥L2≥L3≥0, and define the eigenvectors e1, e2, e3 corresponding to the eigenvalues;

[0101] Step 5.4: Calculate the surface feature parameters of the center point P i , construct the geometric feature set of each point, and use a binary classifier to predict the contour score P(c i |X i ) of each point P i ), where the parameters that can best reflect the surface features are as follows:

[0102]

[0103]

[0104]

[0105]

[0106] Verticality = 1 - |<[0, 0, 1], e3>| (7)

[0107] where: L1, L2, L3 are the eigenvalues of each point, numbered from largest to smallest, e1, e2, e3 are the eigenvectors corresponding to the eigenvalues, and P(c i |X i ) is the Bayesian probability formula, and X i is the sample data.

[0108] Step 5.5: Set up a voxel grid, score the geometric feature sets of each voxel point using an adaptive threshold, and find the seed point P(c i = 0|X i ), and perform a secondary judgment f;

[0109] Step 5.6: Use a multi-spanning tree to construct the neighborhood of the seed point, judge the correctness of its being a high contour score point by the distance to the neighborhood points, etc., and retain the point with the highest score in the feature set candidate set after passing the e and f detections;

[0110] Step 5.7: According to the statistics of the feature set scores and spatial position distributions, use the Markov random field to select continuous optimal feature points from the feature set candidate set to obtain a feature contour model;

[0111] Step 5.8: Loop through steps 5.1 - 5.7 until the point P i(i＝N) Calculation is completed;

[0112] Step 5.9: Use HSV coloring to segment and identify surface features.

[0113] Among them, the Markov random field, namely MRF, is a probability prediction model. By the prerequisite of different scores within the feature set and different spatial distributions of its points, calculate the probabilities of different point paths, and based on different probability distributions, realize the function of extracting the path where the probability is the highest, that is, the optimal high feature contour point; the HSV coloring is a color model composed of three elements: H hue, S saturation, and V value, where H ∈ [0°, 360°], different angles represent different colors, S ∈ [0, 100], different values represent different color saturations, and V ∈ [0, 100], different values represent different color brightnesses, and joint fissures are highlighted through different intensity combinations; the adaptive threshold, namely (NMS), is a necessary post-processing process in object detection. The purpose is to eliminate redundant data. Compare the data with the highest score in step 5.4 as the target with other remaining data for prediction box comparison. If it does not exceed the threshold, it is retained. Based on this screening principle, accurate high contour score points are obtained.

[0114] Example 2:

[0115] In this example, the applicant uses concrete test blocks to simulate the rock mass structure and conducts three-dimensional laser scanning and identification on the test blocks. Specifically, refer to the test block diagram in the experiment simulation in Figure 2 for the test block diagram in the experiment simulation. Figure 3 is for regional growth recognition of bedding structural planes (normal vector difference recognition). Figure 4 is for surface feature recognition of surface joint fissures.

[0116] Figure 2J1 to J18 are different surfaces of different rock samples. Among them, through actual measurement, the normal vectors of J9, J17, and J18 are close, and the normal vectors of J11 and J10 are close. Due to the angle problem, it can be seen that in Figure 3 J1 to 18 structural planes are respectively identified. Among them, through color type identification, J9, J17, and J18 are of the same color, and J11 and J10 are of the same color. Figure 4 Identification is carried out for the joint fissures with a small difference in normal vectors and a relatively serious damage degree in the same plane among L1 to L8.

Claims

1. A multi-type structural plane hierarchical recognition method based on point cloud data, comprising the following steps: Step 1: Perform three-dimensional laser scanning on the slope rock mass to obtain the spatial coordinate information of the slope rock mass point cloud to be processed at the observation site; Step 2: Complete the voxel downsampling of the original data based on KDtree neighbor search to obtain the preliminarily processed point cloud; Step 3: Use the random Hough transform space to calculate the normal vector of the preliminarily processed point cloud data, and use PCA principal component analysis to calculate the corresponding curvature feature to obtain the secondarily processed point cloud; Step 4: Using the normal vector difference and curvature threshold as limiting conditions, use the region growing algorithm to cluster and group the secondarily processed point cloud to obtain the clustering plane, and use RGB random coloring to obtain the recognition result of the bedding fault type structural plane; Step 5: Use the Markov random field point neighborhood feature prediction method to calculate the surface feature of the preliminarily processed point cloud, and use HSV coloring to distinguish to obtain the recognition result of the joint fracture type structural plane; The specific processing process of the said Step 2 is: Step 2.1: Establish a KDtree data structure for the original point cloud data, and count the total number of points in the original point cloud N , and set the lengths of the point cloud space in x , y , z to be L x , L y , L z ; Step 2.2: Find the minimum spacing of the point cloud i min , divide the point cloud space into n = l × w × h small cubes, and set the initial unit cube size l = L x × i min , w = L y × i min , h = L z × i min ; Step 2.3: Calculate the centroid point of each non-empty voxel quality as , and construct a downsampled data point set with all centroid points P c ; Step 2.4: Use KDtree to traverse the centroid points , and set K = 1 to perform nearest neighbor search centered on , and replace it with the nearest neighbor point to to form a new dataset P i , forming the sampled point cloud. The total number of points in the point cloud is N 1; Step 2.5: Determine , after testing 100,000 points, local feature loss will occur when the data is downsampled by more than . It is necessary to make adjustment judgments according to the actual data volume and detailed circumstances The specific processing process of the said Step 3 is: Step 3.1: Perform random Hough transform space to calculate the normal vector of the point cloud data to form a normal vector set; Step 3.2: Use PCA principal component analysis to calculate the curvature feature. This curvature is not the curvature in the mathematical sense, but approximates the curvature information based on the idea of surface variation, and is to solve the eigenvalues of the covariance matrix formed by the data. The specific process is: Step 3.2.1: Set the number of points in the search neighborhood according to the KDtree data structure K Set the initial calculation point P i ; Step 3.2.2: Around P i neighborhood K Construct a spatial point coordinate data set P k ; Step 3.2.3: Calculate the covariance matrix of this data set P k and construct a covariance tensor; Step 3.2.4: Calculate the eigenvalues based on the covariance tensor and sort them λ 1i ≥ λ 2i ≥ λ 3i ≥ 0; Step 3.2.5: Calculate P i Surface curvature: （1） Step 3.2.6: Loop through Step 3.2.2 to Step 3.2.4 to calculate the curvature features of all points and form a curvature set.

2. The multi-type structural plane hierarchical recognition method based on point cloud data according to claim 1, characterized in that The specific processing process of the said Step 4 is: Step 4.1: Import the secondary point cloud data , and construct a data set using the normal vector and surface curvature , according to the size, sort the point cloud, and find and add it to the seed point set; Step 4.2: Set the neighborhood K , and the neighborhood points are added to the growing point set. For each seed point, calculate the difference in the normal angles of its neighborhood points. If the difference is less than the set threshold θ th , then this seed point is given priority and proceeds to the next judgment in Step 4.3; Step 4.3: If the normal angle difference passes the θ th inspection, then perform curvature threshold cth inspection. If it is less than the curvature threshold cth , then this seed point belongs to the current plane of this seed point; Step 4.4: Set the minimum number of points in a plane C min , the maximum number of points in a plane C max , and remove the points that pass the two tests from the growth point set; Step 4.5: Repeat Step 4.1 to Step 4.4, and use RGB random coloring to mark different colors for all clustered planes; Step 4.6: Keep calculating until the number of planes that can be generated from the remaining points is less than C min then stop the calculation; Step 4.7: Identify different structural planes through RGB color distinction.

3. The multi-type structural plane hierarchical recognition method based on point cloud data according to claim 1, characterized in that The specific processing process of the said Step 5 is: Step 5.1: According to the KDtree data structure, set the neighborhood range using the point cloud spacing at the preprocessing time, and set the initial calculation point P i ; Step 5.2: Around P i neighborhood K Construct a spatial point coordinate data set P k , calculate the covariance matrix of this data set, and construct a covariance tensor: （2） Step 5.3: Calculate the eigenvalues based on the covariance tensor and sort them L 1≥ L 2≥ L 3≥0, and define the eigenvectors corresponding to the eigenvalues e 1, e 2, e 3; Step 5.4: Calculate the center point P i Surface feature parameters, construct the geometric feature set of each point, and use a binary classifier to predict the contour score of each point P i where the parameters that can best reflect the surface features are as follows: ​ （3） （4） （5） （6） （7） Step 5.5: Set up a voxel grid, score the geometric feature sets of each voxel point using an adaptive threshold, and find seed points with high contour scores , and perform a secondary judgment f ; Step 5.6: Use multiple spanning trees to construct the neighborhood of the seed points, and judge the correctness of their being high contour score points by the distance to the neighborhood points. Through e , f After detection, retain the point with the highest score in the feature set candidate set; Step 5.7: According to the feature set score and spatial position distribution statistics, use the Markov random field to select continuous optimal feature points from the feature set candidate set to obtain the feature contour model; Step 5.8: Loop through steps 5.1 - 5.7 until the calculation of P i（i=N） is completed; Step 5.9: Use HSV coloring to segment the surface features for recognition.

Images