Point cloud clustering method based on normal vector information

Abstract

The invention discloses a point cloud clustering method based on normal vector information, and belongs to the technical field of point cloud segmentation. Comprising the following steps: acquiring point cloud data; calculating a normal vector of each point in the point cloud data; performing preliminary clustering on the point cloud data based on a DBSCAN algorithm according to the normal vector; points which are not classified into any cluster of the preliminary cluster are added into a noise point set; carrying out plane and curved surface distinguishing on a cluster set subjected to preliminary clustering based on a normal vector included angle; further plane distinguishing is conducted on the obtained plane cluster set according to the projection distance in the main normal vector direction, and a final plane cluster set is obtained; and completing point cloud clustering based on the final plane cluster set, the curved surface cluster set and the noise point set. By combining the normal vector information of the point cloud and the DBSCAN clustering algorithm, the clustering and classification problems of the point cloud data are effectively solved, and a more efficient and stable point cloud processing method is provided especially for a complex three-dimensional object surface.

Description

Technical Field

[0001] The present invention relates to the field of point cloud segmentation technology in three-dimensional point cloud data processing technology, and in particular to a point cloud clustering method based on normal vector information. Background Art

[0002] As a high-precision data expression for describing the geometric shape of an object surface, 3D point cloud data has been widely used in autonomous driving, building information modeling, industrial inspection, reverse engineering, etc. However, the characteristics of point cloud data, including sparsity, irregularity, and noise interference, make segmentation and clustering of point cloud data a technical difficulty.

[0003] Point cloud clustering is the process of dividing point cloud data into clusters with similar attributes or semantics. Existing point cloud clustering can be mainly divided into the following categories:

[0004] Density-based methods: such as DBSCAN (Density-Based Spatial Clustering of Applications with Noise), which divides point clouds into clusters with higher density by defining density thresholds and neighborhood radius, can effectively overcome the influence of noise, and is adaptive for finding clusters of arbitrary shapes and sizes. However, the traditional DBSCAN algorithm defines neighborhoods based on Euclidean distance and cannot use the normal vector information of point clouds, which makes it difficult to effectively process areas that are geometrically similar but spatially adjacent, such as adjacent parallel planes or complex surfaces.

[0005] Methods based on geometric features: such as region growing method. By gradually expanding the neighborhood of the point, the region is recursively grown based on whether the geometric features of adjacent points (such as the normal vector angle or curvature difference) meet the set threshold. This method is suitable for identifying simple geometric shapes (such as planes and surfaces), but when the point cloud contains multiple adjacent regions or parallel planes, it is prone to mis-segmentation problems and is more sensitive to noise points.

[0006] Methods based on deep learning, such as PointNet and PointNet++, can achieve efficient segmentation and clustering through point cloud feature learning. However, these methods rely on a large amount of labeled data, have high training costs, and are difficult to adapt to small samples or real-time processing scenarios. Summary of the invention

[0007] In view of the problems existing in the prior art, the present invention takes the normal vector as the core feature, proposes a definition of normal distance in the normal vector space, and introduces improvements based on DBSCAN through this distance to realize the utilization of geometric information in the point cloud, and further distinguish between planes and curved surfaces to achieve effective recognition of different types of surfaces.

[0008] In order to achieve the above object, the present invention proposes a point cloud clustering method based on normal vector information, comprising the following steps:

[0009] (1) Obtain point cloud data;

[0010] (2) calculating the normal vector of each point in the point cloud data;

[0011] (3) performing preliminary clustering of the point cloud data based on the normal vectors based on the DBSCAN algorithm;

[0012] (4) For points that cannot be classified into any cluster in the preliminary clustering, they are marked as noise points and added to the noise point set;

[0013] (5) The cluster sets of preliminary clustering are distinguished between planes and curved surfaces based on the normal vector angles to obtain plane cluster sets and curved surface cluster sets;

[0014] (6) further distinguishing the obtained plane cluster set according to the projection distance in the direction of the main normal vector to obtain a final plane cluster set;

[0015] (7) Complete point cloud clustering based on the final plane cluster set, surface cluster set and noise point set.

[0016] Furthermore, the step (2) is specifically as follows:

[0017] (2.1) For each point p in the point cloud data i , take the K points with the closest Euclidean distance to form the first neighborhood S i , where: K is the preset value;

[0018] (2.2) Calculate the first neighborhood S i Center And the covariance matrix C:

[0019]

[0020] Where: p i =(x i ,y i , z i ), i=1, ..., N;

[0021] (2.3) Obtain the eigenvalues ​​and eigenvectors of the covariance matrix C, and denote the eigenvalue with the smallest absolute value as λ min ,λ min The corresponding eigenvector is denoted as v min ;

[0022] (2.4) Get each point p i The normal vector at ;

[0023]

[0024] Where: n i For point p i The normal vector of .

[0025] Furthermore, the step (3) is specifically as follows:

[0026] (3.1) The normal vector of each point in the point cloud data is defined as follows:

[0027] (3.1.1) Point p i With p j The normal distance θ between ij :

[0028] θ ij =cos -1 (|n i ·n j |)

[0029] Where: n i 、n j For point p i 、p j The normal vector of

[0030] (3.1.2) Based on the normal distance, define the distance belonging to point p i The second neighbor

[0031] Where: θ is the predefined normal distance threshold;

[0032] (3.1.3) For each point p i , search its second neighborhood The number of points in the second neighborhood is set to MinPts. If If the number of points contained is greater than MinPts, pi is marked as a core point;

[0033] (3.1.4) If p i is the core point, and p j Located in p i The second neighborhood of j From p i Density directly; for point p i and p k , if there exists a series of points p i , p j , ..., p k , so that any adjacent point satisfies the density direct relationship, then p k From p i Density can reach;

[0034] (3.2) Initialize all points in the point cloud data as “unvisited” and initialize the cluster set Initialize the core point set

[0035] (3.3) According to the method described in step (3.1), find all the core points in the point cloud data and add them to the core point set Ω;

[0036] (3.4) Select an “unvisited” core point p from the core point set Ω i , p i And all from p i The density-reachable points are taken as a new cluster C k Add to cluster set C and add C k All points in are marked as "visited";

[0037] (3.5) Repeat step (3.4) until all points in the core point set Ω have been marked as "visited", completing the preliminary clustering cluster set C = {C 1 , C 2 , ..., C n}.

[0038] Furthermore, the step (5) is specifically as follows:

[0039] (5.1) Calculate each cluster C in the cluster set of the preliminary clustering k The average of all pairs of normal vector angles in :

[0040]

[0041] Among them: point p i 、p j is the cluster set C k The point in For point p i 、p j The corresponding normal vector angle; n i 、n j For point p i 、p j Normal vector; M is the cluster set C k The number of combinations of any two points in

[0042] (5.2) Calculate the cluster set C k The variance of the angles between all pairs of normal vectors in :

[0043]

[0044] (5.3) Setting the threshold like Then the current cluster C k is classified as a surface cluster; if Then the current cluster C k Classify into flat clusters;

[0045] (5.4) The set of all plane clusters is denoted as the plane cluster set, and the set of all surface clusters is denoted as the surface cluster set.

[0046] Furthermore, the step (6) is specifically as follows:

[0047] (6.1) For each cluster S in the set of plane clusters k , calculate the principal normal vector:

[0048]

[0049] Where: s is the cluster S k The number of midpoints, point p j For cluster S k The point in j For point p j The normal vector of

[0050] (6.2) For cluster S k , calculate the projection length of each point to the direction of its principal normal vector, and get the cluster S k The projection length set of all points D = {d 1 , d 2 , ..., d s};

[0051]

[0052] Where: d i For point p i The length of the projection to the direction of its principal normal vector is a scalar;

[0053] (6.3) Select the minimum value d in the projection length set min and the maximum value d max , the interval [d min , d max ] is evenly divided into q small intervals with left closing and right opening, q is a given variable; count the number of points in each small interval, recorded as n 1 , n 2 , ..., n q , generate a histogram;

[0054] (6.4) Compare the number of points in any two adjacent small intervals. If n j <n j+1 , then let the label m j is 1; if n j =nj+1 , then let the label m j is 0; if n j >n j+1 , then let the label m j is -1; where j = 1, ..., q-1, m q =m q-1 ;

[0055] (6.5) All intervals corresponding to the mark 0 are merged with the next interval into one interval, whose mark is the mark of the next interval, to form a new histogram; all intervals with consecutive equal marks in the new histogram are merged into one interval, and the original marks are retained to obtain the final histogram;

[0056] (6.6) All the intervals of the final histogram are judged from small to large, and the adjacent intervals marked as 1 and -1 are merged. The points in the intervals are classified into the same plane cluster. The points in the unmerged intervals are each regarded as a separate plane cluster, and the final plane cluster set F = {F 1 , F 2 , ..., F m}, m is the number of plane clusters contained in the final plane cluster set.

[0057] Furthermore, any of the plane clusters F m The plane equation is:

[0058] The plane cluster F m The coordinates of each point are q i =(x i ,y i , z i ), i = 1, ..., n; the principal normal vector is

[0059] Calculate the geometric center point p of the plane cluster c =(x c ,y c , z c );

[0060]

[0061] Where: N is the plane cluster F m the number of midpoints;

[0062] Construct the equation of a plane:

[0063] a(xx c )+b(yy c )+c(zz c )=0

[0064] After expansion, we get:

[0065] ax+by+cz+d=0

[0066] Among them: a, b, c are parameters, d = -(ax c +by c +cz c ).

[0067] Furthermore, the surface cluster set can also be further distinguished based on the k-means method.

[0068] Beneficial effects of the present invention:

[0069] 1. The present invention takes the normal vector as the core feature, proposes the definition of normal distance in the normal vector space, and introduces improvements based on DBSCAN through the distance to realize the utilization of geometric information in the point cloud.

[0070] 2. Efficiently distinguish different planes among parallel planes. In the initial clustering, the consistency of the normal vectors of the points is used to separate the parallel planes from other geometric regions. In the further refined clustering, the distribution of the projection distance from the points to the main normal vector direction is analyzed to identify different planes among multiple parallel planes. This method does not need to rely on deep learning models, has high computational efficiency, and is suitable for a variety of application requirements in point cloud scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 Schematic diagram of the overall process of the point cloud clustering method based on normal vector information according to an embodiment of the present invention. DETAILED DESCRIPTION

[0072] The present invention is further explained below in conjunction with the accompanying drawings and embodiments.

[0073] like Figure 1 As shown, the embodiment of the present invention proposes a point cloud clustering method based on normal vector information, comprising the following steps:

[0074] S101. Obtain point cloud data.

[0075] Get the original point cloud data, P = {p i =(x i ,y i , z i )|i=1,...,N}. Where: (x i ,y i , z i ) is the point cloud coordinate, and N is the number of point cloud data.

[0076] S102: Calculate the normal vector of each point in the point cloud data.

[0077] (1) For each point p in the point cloud data i , take the K points with the closest Euclidean distance to form the first neighborhood] i Wherein: K is a preset value, and in the embodiment of the present invention, K=50.

[0078] (2) Calculate the first neighborhood S i Center And the covariance matrix C:

[0079]

[0080]

[0081] Where: C is a 3*3 covariance matrix.

[0082] (3) Obtain the eigenvalues ​​and eigenvectors of the covariance matrix C, and denote the eigenvalue with the smallest absolute value as λ min ,λ min The corresponding eigenvector is denoted as v min .

[0083] (4) Get each point p i The curvature and normal vector at ;

[0084] ρ i =λ min

[0085]

[0086] Where: i is the curvature, n i is the normal vector.

[0087] S103 , performing preliminary clustering of the point cloud data based on the normal vectors based on the DBSCAN algorithm.

[0088] (1) The concepts used by the DBSCAN clustering algorithm are defined based on the normal vector of each point in the point cloud data:

[0089] (1.1) Point p i With p j The normal distance θ between ij :

[0090] θ ij =cos -1 (|n i ·n j |)

[0091] Where: n i 、n j For point p i 、p j The normal vector of .

[0092] (1.2) Based on the normal distance, define the distance belonging to point p i The second neighbor

[0093] Where: θ is a predefined normal distance threshold, which is a very small value. In the embodiment of the present invention, ε is set θ =5°, indicating the size of the neighborhood area on the unit sphere.

[0094] (1.3) For each point p i , search its second neighborhood The number of points in the second neighborhood is set to MinPts. If If the number of points contained is greater than MinPts, then p i Mark as core point.

[0095] In the embodiment of the present invention, MinPts=10 is set.

[0096] (1.4) If p i is the core point, and p j Located in p i The second neighborhood of j From p i Density directly; for point p i and p k , if there exists a series of points p i , p j , ..., p k , so that any adjacent point satisfies the density direct relationship, then p k From p i The density can be reached; for point p j and p k , if there is a core point p i , so that p j and p k All from p i The density can reach, then it is called p j and p k Density connected.

[0097] (2) Initialize all points in the point cloud data as “unvisited” and initialize the cluster set Initialize the core point set

[0098] (3) According to the method described in step (1), find all the core points in the point cloud data and add them to the core point set Ω.

[0099] (4) Select an “unvisited” core point p from the core point set Ωi , p i And all from p i The density-reachable points are taken as a new cluster C k Add to cluster set C and add C k All points in are marked as visited.

[0100] (5) Repeat step (4) until all points in the core point set Ω have been marked as "visited", completing the preliminary clustering cluster set C = {C 1 , C 2 , ..., C n}.

[0101] S104. For points that cannot be classified into any cluster of the preliminary clustering, mark them as noise points and add them to the noise point set.

[0102] S105 , distinguishing planes and curved surfaces from the initially clustered cluster sets based on the normal vector angles, to obtain plane cluster sets and curved surface cluster sets.

[0103] (1) Calculate each cluster C in the preliminary clustering cluster set k The average of all pairs of normal vector angles in :

[0104]

[0105] Among them: point p i 、p j is the cluster set C k The point in For point p i 、p j The corresponding normal vector angle; n i 、n j For point p i 、p j Normal vector; M is the cluster set C k The number of combinations of any two points in .

[0106] (2) Calculate the cluster set C k The variance of the angles between all pairs of normal vectors in :

[0107]

[0108] (3) Setting the threshold like Then the current cluster C k is classified as a surface cluster; if Then the current cluster C k Classified as a flat cluster.

[0109] (4) The set of all plane clusters is recorded as the plane cluster set, and the set of all surface clusters is recorded as the surface cluster set.

[0110] S106 , further distinguishing the obtained plane cluster set according to the projection distance in the direction of the main normal vector to obtain a final plane cluster set.

[0111] (1) For each cluster S in the set of planar clusters k , calculate the principal normal vector:

[0112]

[0113] Where: s is the cluster S k The number of midpoints, point p j For cluster S k The point in j For point p j The normal vector of .

[0114] (2) For cluster S k , calculate the projection length of each point to the direction of its principal normal vector, and get the cluster S k The projection length set of all points D = {d 1 , d 2 , ..., d s}.

[0115]

[0116] Where: d i For point p i The length of the projection onto its principal normal direction, a scalar.

[0117] (3) Select the minimum value d in the projection length set min and the maximum value d max , the interval [d min , d max ] is evenly divided into q small intervals with left closing and right opening, q is a given variable; count the number of points in each small interval, recorded as n 1 , n 2 , ..., n q , generate a histogram.

[0118] (4) Compare the number of points in any two adjacent small intervals. If n j <n j+1 , then let the label m j is 1; if n j =n j+1 , then let the label m j is 0; if n j >n j+1 , then let the label mj is -1; where j = 1, ..., q-1, m q =m q-1 .

[0119] (5) All intervals corresponding to the mark 0 are merged with the next interval into one interval, and its mark is the mark of the next interval, forming a new histogram; all intervals with consecutive equal marks in the new histogram are merged into one interval, and the original marks are retained to obtain the final histogram;

[0120] (7) All the intervals of the final histogram are judged from small to large, and the adjacent intervals marked as 1 and -1 are merged. The points in the interval are classified into the same plane cluster. The points in the unmerged intervals are each regarded as a separate plane cluster, and the final plane cluster set F = {F 1 , F 2 , ..., F m}, m is the number of intervals contained in the final histogram.

[0121] The final histogram is judged for peaks and valleys. If the mark changes from 1 to -1, the interval marked as 1 is the peak, and if the mark changes from -1 to 1, the interval marked as -1 is the valley. Adjacent peaks and valleys are merged into one interval as the same plane cluster, and individual peaks or valleys that are not merged are regarded as a plane cluster.

[0122] For any of the plane families F m The plane equation is:

[0123] Planar cluster F m The coordinates of each point are q i =(x i ,y i , z i ), i = 1, ..., n; the principal normal vector is

[0124] Calculate the geometric center point p of the plane cluster c =(x c ,y c , z c ):

[0125]

[0126] Where: N is the plane cluster F m the number of midpoints;

[0127] Construct the equation of a plane:

[0128] a(xx c )+b(yy c)+c(zz c )=0

[0129] After expansion, we get:

[0130] ax+by+cz+d=0

[0131] Among them: a, b, c are parameters, d = -(ax c +by c +cz c ).

[0132] S107 , completing point cloud clustering based on the final plane cluster set, surface cluster set and noise point set.

[0133] The surface cluster set can also be further distinguished based on the k-means method.

[0134] The above is only an embodiment of the present application and is not intended to limit the protection scope of the present application. Any modification, equivalent replacement and improvement made within the spirit and scope of the present application are included in the protection scope of the present application.

Claims

1. A point cloud clustering method based on normal vector information, characterized in that: The steps include: (1) Obtain point cloud data; (2) calculating the normal vector of each point in the point cloud data; (3) performing preliminary clustering of the point cloud data based on the normal vectors based on the DBSCAN algorithm; (4) For points that cannot be classified into any cluster in the preliminary clustering, they are marked as noise points and added to the noise point set; (5) The cluster sets of preliminary clustering are distinguished between planes and curved surfaces based on the normal vector angles to obtain plane cluster sets and curved surface cluster sets; (6) further distinguishing the obtained plane cluster set according to the projection distance in the direction of the main normal vector to obtain a final plane cluster set; (7) Complete point cloud clustering based on the final plane cluster set, surface cluster set and noise point set.

2. The point cloud clustering method based on normal vector information according to claim 1, characterized in that: The step (2) is specifically: (2.1) For each point p in the point cloud data i , take the K points with the closest Euclidean distance to form the first neighborhood S i , where: K is the preset value; (2.2) Calculate the first neighborhood S i Center And the covariance matrix C: Where: p i =(x i ,y i , z i ), i=1, ..., N; (2.3) Obtain the eigenvalues ​​and eigenvectors of the covariance matrix C, and denote the eigenvalue with the smallest absolute value as λ min ,λ min The corresponding eigenvector is denoted as v min ; (2.4) Get each point p i The normal vector at ; Where: n i For point p i The normal vector of .

3. The point cloud clustering method based on normal vector information according to claim 1, characterized in that: The step (3) is specifically: (3.1) The normal vector of each point in the point cloud data is defined as follows: (3.1.1) Point p i With p j The normal distance θ between ij : i ij =cos -1 (|n i ·n j |) Where: n i 、n j For point p i 、p j The normal vector of (3.1.2) Based on the normal distance, define the distance belonging to point p i The second neighbor Where: θ is the predefined normal distance threshold; (3.1.3) For each point p i , search its second neighborhood The number of points in the second neighborhood is set to MinPts. If If the number of points contained is greater than MinPts, then p i Mark as core point; (3.1.4) If p i is the core point, and p j Located in p i The second neighborhood of j From p i Density directly; for point p i and p k , if there exists a series of points p i , p j , ..., p k , so that any adjacent point satisfies the density direct relationship, then p k From p i Density can reach; (3.2) Initialize all points in the point cloud data as "unvisited" and initialize the cluster set Initialize the core point set (3.3) According to the method described in step (3.1), find all the core points in the point cloud data and add them to the core point set Ω; (3.4) Select an "unvisited" core point p from the core point set Ω i , p i And all from p i The density-reachable points are taken as a new cluster C k Add to cluster set C and add C k All points in are marked as "visited"; (3.5) Repeat step (3.4) until all points in the core point set Ω have been marked as "visited", completing the preliminary clustering cluster set C = {C1, C2, ..., C n }.

4. The point cloud clustering method based on normal vector information according to claim 1, characterized in that: The step (5) is specifically as follows: (5.1) Calculate each cluster C in the cluster set of the preliminary clustering k The average of all pairs of normal vector angles in : Among them: point p i 、p j is the cluster set C k The point in For point p i 、p j The corresponding normal vector angle; n i 、n j For point p i 、p j Normal vector; M is the cluster set C k The number of combinations of any two points in (5.2) Calculate the cluster set C k The variance of the angles between all pairs of normal vectors in : (5.3) Setting the threshold like Then the current cluster C k is classified as a surface cluster; if Then the current cluster C k Classify into flat clusters; (5.4) The set of all plane clusters is denoted as the plane cluster set, and the set of all surface clusters is denoted as the surface cluster set.

5. The point cloud clustering method based on normal vector information according to claim 1, characterized in that: The step (6) is specifically as follows: (6.1) For each cluster S in the set of plane clusters k , calculate the principal normal vector: Where: s is the cluster S k The number of midpoints, point p j For cluster S k Points in, n j For point p j The normal vector of (6.2) For cluster S k , calculate the projection length of each point to the direction of its principal normal vector, and get the cluster S k The projection length set of all points D = {d1, d2, ..., d s }; Where: d i For point p i The length of the projection to the direction of its principal normal vector is a scalar; (6.3) Select the minimum value d in the projection length set min and the maximum value d max , the interval [d min , d max ] is evenly divided into q small intervals with left closing and right opening, q is a given variable; the number of points in each small interval is counted, recorded as n1, n2, ..., n q , generate a histogram; (6.4) Compare the number of points in any two adjacent small intervals. If n j <n j+1 , then let the label m j is 1; if n j =n j+1 , then let the label m j is 0; if n j >n j+1 , then let the label m j is -1; where j = 1, ..., q-1, m q =m q-1 ; (6.5) All intervals corresponding to the mark 0 are merged with the next interval into one interval, whose mark is the mark of the next interval, to form a new histogram; all intervals with consecutive equal marks in the new histogram are merged into one interval, and the original marks are retained to obtain the final histogram; (6.6) All the intervals of the final histogram are judged from small to large, and the adjacent intervals marked as 1 and -1 are merged. The points in the interval are classified into the same plane cluster. The points in the unmerged intervals are each regarded as a plane cluster separately, and the final plane cluster set F = {F1, F2, ..., F m }, m is the number of plane clusters contained in the final plane cluster set.

6. The point cloud clustering method based on normal vector information according to claim 5, characterized in that: Any of the plane families F m The plane equation is: The plane cluster F m The coordinates of each point are q i =(x i ,y i , z i ), i=1,…,n; The principal normal vector is Calculate the geometric center point p of the plane cluster c =(x c ,y c , z c ); Where: n is the plane cluster F m the number of midpoints; Construct the equation of a plane: a(x-x c )+b(y-y c )+c(z-z c )=0 After expansion, we get: ax+by+cz+d=0 Among them: a, b, c are parameters, d = -(ax c +by c +cz c ).

7. The point cloud clustering method based on normal vector information according to claim 1, characterized in that: The surface cluster set can also be further distinguished based on the k-means method.

Images