CN107292276B - Vehicle-mounted point cloud clustering method and system - Google Patents

Vehicle-mounted point cloud clustering method and system Download PDF

Info

Publication number
CN107292276B
CN107292276B CN201710510689.6A CN201710510689A CN107292276B CN 107292276 B CN107292276 B CN 107292276B CN 201710510689 A CN201710510689 A CN 201710510689A CN 107292276 B CN107292276 B CN 107292276B
Authority
CN
China
Prior art keywords
hyper
point cloud
voxel
point
ground
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Expired - Fee Related
Application number
CN201710510689.6A
Other languages
Chinese (zh)
Other versions
CN107292276A (en
Inventor
刘亚文
张颖
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Wuhan University WHU
Original Assignee
Wuhan University WHU
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Wuhan University WHU filed Critical Wuhan University WHU
Priority to CN201710510689.6A priority Critical patent/CN107292276B/en
Publication of CN107292276A publication Critical patent/CN107292276A/en
Application granted granted Critical
Publication of CN107292276B publication Critical patent/CN107292276B/en
Expired - Fee Related legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00—Scenes; Scene-specific elements
    • G06V20/50—Context or environment of the image
    • G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
    • G06V20/588—Recognition of the road, e.g. of lane markings; Recognition of the vehicle driving pattern in relation to the road
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/23—Clustering techniques

Abstract

The invention provides a vehicle-mounted point cloud clustering method and a system, which comprises point cloud denoising and filtering, wherein scattered noise points of the point cloud are removed, point cloud filtering processing is carried out, and ground points and non-ground points are distinguished; segmenting the point cloud to generate a hyper-voxel, wherein the hyper-voxel is generated by segmenting non-ground points by adopting a density-based spatial clustering algorithm, so that the condition that different ground objects are not adhered is ensured; and (3) carrying out point cloud clustering with spatial context association, wherein the point cloud clustering with spatial context association comprises the steps of analyzing the characteristics of the hyper-voxels and spatial context association, and integrating the multi-factor weight to increase the hyper-voxels areas so as to finish point cloud clustering. The method solves the problem of over-segmentation or insufficient segmentation of the vehicle-mounted point cloud cluster, meets the requirement of rapidly acquiring three-dimensional space information, and has important market value.

Description

Vehicle-mounted point cloud clustering method and system
Technical Field
The invention relates to the technical field of vehicle-mounted laser scanning data processing, in particular to a vehicle-mounted point cloud clustering method and system combining spatial context association.
Background
Point cloud data acquired by a Vehicle-mounted Laser Scanning system (VLS) has the characteristics of high density, high precision, quick acquisition and the like, and is a technical means for quickly acquiring three-dimensional space information at the present front edge. At present, scholars at home and abroad carry out a great deal of research on the application aspect of point cloud data, wherein vehicle-mounted point cloud clustering is an important component of VLS data processing and information extraction and is a precondition and a key link for realizing automatic identification of ground objects.
The current common point cloud clustering methods can be divided into methods such as division clustering, hierarchical clustering, grid clustering and density clustering. The dividing and clustering method depends on the determination of the initial clustering number, the proximity matrix calculation in the hierarchical clustering method is complex, the VLS point cloud data volume is large, the calculation is slow, the grid clustering algorithm is suitable for the data clustering with relative regular shapes and strong directionality, and is not suitable for the point cloud data presenting the cluster shape in the class. The density clustering method is to identify clusters according to the density of objects, divide areas with high enough density into clusters, and find clusters with any shapes in a noisy spatial data set. However, in the case that the ground features of the city street view are dense, the ground features distributed in the vertical direction are abundant, and the adhesion between the ground features is serious, only density clustering of the spatial distance between points is considered, so that the serious under-segmentation or over-segmentation condition is caused.
Disclosure of Invention
The invention mainly solves the technical problems in the prior art and provides a technical scheme of vehicle-mounted point cloud clustering by taking a hyper-voxel as an object and combining spatial context association.
The technical problem of the invention is mainly solved by the following technical scheme:
a vehicle-mounted point cloud clustering method comprises the following steps,
step 1, point cloud denoising and filtering, including removing scattered noise points from the point cloud, and carrying out point cloud filtering processing to distinguish ground points and non-ground points;
step 2, segmenting point clouds to generate hyper-voxels, wherein non-ground points are segmented by adopting a density-based spatial clustering algorithm to generate the hyper-voxels, so that the condition that different ground objects are not adhered is ensured;
step 3, spatial context-associated point cloud clustering, which comprises the steps of analyzing the characteristics of the hyper-voxels and spatial context association, and integrating multi-factor weights to increase the hyper-voxels regions to finish point cloud clustering,
step 3.1, calculating the characteristics of each hyper-voxel, including dimension information VD of the hyper-voxel, three-dimensional coordinates V (X, Y, Z) of a central point and bounding box information (Vmax, Vmin);
the dimensionality information VD of the hyper-voxel is calculated in a mode that the dimensionality of each point in the hyper-voxel is calculated, three-dimensional, two-dimensional and one-dimensional point numbers are counted, and the dimensionality with the largest point number is assigned to the hyper-voxel;
the three-dimensional coordinates V (X, Y, Z) of the central point are obtained by averaging the three-dimensional coordinates of all points in the hyper-voxels;
bounding box information (Vmax, Vmin) is obtained from a cuboid bounding box of the hyper-voxel;
step 3.2, according to a preset initial value Rs of the search radius, calculating spatial correlation characteristics of the superpixels, including dimensional consistency of the superpixels and plane distance D between center points of the superpixelscHeight difference H between center points of voxels of supervoxelcAnd the voxel bounding boxes intersect or are connected in the direction vertical to the ground;
step 3.3, combining the spatial context association of the hyper voxels to carry out region growing clustering, wherein the growing condition W is the weight of the comprehensive multifactor among the hyper voxels, and the calculation is as follows
Figure GDA0002107466590000021
Wherein, CuFor the spatial context correlation factor of hyper-voxels, wuAs the contribution degree of each factor to the clustering, u is 1,2,3, 4;
C1for dimension correlation factor, describe the same property of the same ground physical dimension, when the two superpixel dimension information are the same, C1The value is 1, otherwise, the value is 0;
C2describing geometric constraint of ground object space plane distribution for the plane position association factor, and when the distance Dc between two hyper-voxel planes is less than a preset threshold value, C2The value is 1, otherwise, the value is 0;
C3describing geometric constraint of ground object height distribution for a height correlation factor, and when the height difference Hc of two hyper-voxels is smaller than a preset threshold value, C3The value is 1, otherwise, the value is 0;
C4describing the geometric constraint of the vertical distribution of the ground object space for the position correlation factor in the vertical direction, and when two hyper-voxels intersect or are connected in the direction vertical to the ground, C4The value is 1, otherwise 0.
Moreover, the calculation of the dimensions of the points within the hyper-voxel is carried out as follows,
for any point p, a neighborhood point set is set as the nearest K points q contained in a sphere by taking the point p as the center of the spherekSet of (2), set center of gravity
Figure GDA0002107466590000022
As follows below, the following description will be given,
Figure GDA0002107466590000023
the matrices M and C are calculated as follows,
Figure GDA0002107466590000031
Figure GDA0002107466590000032
performing characteristic decomposition on the matrix C to obtain sorted characteristic values of lambda respectively1,λ2,λ3，
Calculating the parameter a1D,a2DAnd a3DAs follows, the one-dimensional, two-dimensional and three-dimensional information of the point p is described separately,
Figure GDA0002107466590000033
where μ is a regularization coefficient, and μ ═ σ is defined1，
Figure GDA0002107466590000034
t is 1,2,3, if adDAt the maximum, the point p dimension is d.
Further, assuming that the OZ direction of the coordinate system O-XYZ is the vertical ground direction, the intersection or the connection of the voxel bounding boxes in the vertical ground direction means that the rectangles projected on the XOY plane by the voxel bounding boxes are mutually included, and the rectangles projected on the XOZ plane are intersected or connected.
And, DcAnd HcThe calculation is as follows,
Figure GDA0002107466590000035
HC＝|Z-Z′|
wherein, (X, Y, Z) and (X ', Y ', Z ') are respectively three-dimensional coordinates of the center points of adjacent superpixels;
and, wuThe value of (A) is determined by a correlation factor C which is preferentially related to the spatial position of the ground feature2、C3And C4Followed by a dimension correlation factor C1。
The invention also provides a vehicle-mounted point cloud clustering system, which comprises the following modules,
the first module is used for point cloud denoising and filtering, including removing scattered noise points from the point cloud, and carrying out point cloud filtering processing to distinguish ground points and non-ground points;
the second module is used for segmenting the point cloud to generate a hyper-voxel, and comprises a step of segmenting non-ground points by adopting a density-based spatial clustering algorithm to generate the hyper-voxel so as to ensure that different ground objects are not adhered;
the third module is used for point cloud clustering of spatial context association, including analyzing the characteristics of the hyper-voxels and spatial context association, and performing regional growth of the hyper-voxels by synthesizing multi-factor weights to finish the point cloud clustering, and comprises the following units, namely a first unit, a second unit, a third module and a third module, wherein the first unit is used for calculating the characteristics of each hyper-voxel, including the dimension information VD of the hyper-voxel, the three-dimensional coordinates V (X, Y, Z) of a central point and bounding box information (Vmax, Vmin);
the dimensionality information VD of the hyper-voxel is calculated in a mode that the dimensionality of each point in the hyper-voxel is calculated, three-dimensional, two-dimensional and one-dimensional point numbers are counted, and the dimensionality with the largest point number is assigned to the hyper-voxel;
the three-dimensional coordinates V (X, Y, Z) of the central point are obtained by averaging the three-dimensional coordinates of all points in the hyper-voxels;
bounding box information (Vmax, Vmin) is obtained from a cuboid bounding box of the hyper-voxel;
a second unit for calculating spatial correlation characteristics of the superpixel according to a preset initial value Rs of the search radius, including the dimension consistency of the superpixel and the plane distance D between the center points of the superpixelcHeight difference H between center points of voxels of supervoxelcAnd the voxel bounding boxes intersect or are connected in the direction vertical to the ground;
a third unit, for performing region growing clustering by combining the spatial context association of the hyper-voxels, wherein the growing condition W is the weight of the comprehensive multifactor among the hyper-voxels, and the calculation is as follows
Wherein, CuFor the spatial context correlation factor of hyper-voxels, wuAs the contribution degree of each factor to the clustering, u is 1,2,3, 4;
C1for dimension correlation factor, describe the same property of the same ground-object dimension when two superbodiesThe information of the element dimension is the same, C1The value is 1, otherwise, the value is 0;
C2describing geometric constraint of ground object space plane distribution for the plane position association factor, and when the distance Dc between two hyper-voxel planes is less than a preset threshold value, C2The value is 1, otherwise, the value is 0;
C3describing geometric constraint of ground object height distribution for a height correlation factor, and when the height difference Hc of two hyper-voxels is smaller than a preset threshold value, C3The value is 1, otherwise, the value is 0;
C4describing the geometric constraint of the vertical distribution of the ground object space for the position correlation factor in the vertical direction, and when two hyper-voxels intersect or are connected in the direction vertical to the ground, C4The value is 1, otherwise 0.
Moreover, the calculation of the dimensions of the points within the hyper-voxel is carried out as follows,
for any point p, a neighborhood point set is set as the nearest K points q contained in a sphere by taking the point p as the center of the spherekSet of (2), set center of gravity
Figure GDA0002107466590000042
As follows below, the following description will be given,
Figure GDA0002107466590000051
the matrices M and C are calculated as follows,
Figure GDA0002107466590000052
Figure GDA0002107466590000053
performing characteristic decomposition on the matrix C to obtain sorted characteristic values of lambda respectively1,λ2,λ3，
Calculating the parameter a1D,a2DAnd a3DAs follows, the one-dimensional, two-dimensional and three-dimensional information of the point p is described separately,
Figure GDA0002107466590000054
where μ is a regularization coefficient, and μ ═ σ is defined1，
Figure GDA0002107466590000055
t is 1,2,3, if adDAt the maximum, the point p dimension is d.
Further, assuming that the OZ direction of the coordinate system O-XYZ is the vertical ground direction, the intersection or the connection of the voxel bounding boxes in the vertical ground direction means that the rectangles projected on the XOY plane by the voxel bounding boxes are mutually included, and the rectangles projected on the XOZ plane are intersected or connected.
And, DcAnd HcThe calculation is as follows,
Figure GDA0002107466590000056
HC＝|Z-Z′|
wherein, (X, Y, Z) and (X ', Y ', Z ') are respectively three-dimensional coordinates of the center points of adjacent superpixels;
and, wuThe value of (A) is determined by a correlation factor C which is preferentially related to the spatial position of the ground feature2、C3And C4Followed by a dimension correlation factor C1。
According to the vehicle-mounted point cloud clustering method, the vehicle-mounted point cloud is segmented through a density-based spatial clustering algorithm (DBSCAN algorithm) to generate the superpixel, and on the basis, vehicle-mounted point cloud clustering is performed by combining spatial context correlation of the superpixel. According to the technical scheme, the point cloud clustering result is effectively improved, the number of clustering clusters is greatly reduced while the accuracy of point cloud clustering is ensured, good basic data are provided for subsequent unsupervised classification or supervised classification, the problem of over-segmentation or under-segmentation of vehicle-mounted point cloud clustering is solved, the requirement for rapidly acquiring three-dimensional space information is met, and the method has an important market value.
Drawings
FIG. 1 is a schematic diagram of a voxel bounding box intersecting or connected in a vertical direction in an embodiment of the invention.
FIG. 2 is a diagram illustrating region growing clustering in conjunction with spatial context association according to an embodiment of the present invention.
Detailed Description
The technical scheme of the invention is further specifically described in the following by combining the drawings and the embodiment.
The invention provides a vehicle-mounted point cloud clustering method based on spatial context association. The method comprises the steps of utilizing a density-based spatial clustering algorithm DBSCAN to divide point clouds to form voxels, and carrying out vehicle-mounted point cloud clustering by analyzing the characteristics of the voxels and the spatial context correlation among the voxels and integrating multi-factor weights. In the clustering process, the spatial distance of the points is considered, and the spatial context correlation of the neighborhood points is combined, so that the problem of over-segmentation or insufficient segmentation of the vehicle-mounted point cloud clustering is greatly improved.
The data of the embodiment is urban street view point cloud data obtained by a Reigle vux-1 ground laser scanner, and the point cloud number is 2180726. The sampling interval of the data is 0.01 degrees, and the data mainly comprises objects such as building surfaces, power lines, telegraph poles, trees, street lamp traffic sign posts and the like.
The embodiment provides a vehicle-mounted point cloud clustering method combined with spatial context association, which comprises the following steps:
step 1, point cloud denoising and filtering. Scattered noise points of the point cloud are removed, point cloud filtering processing is carried out, and ground points and non-ground points are distinguished.
The point cloud denoising and filtering in the step 1 comprises the following steps:
step 1.1, establishing KD-tree space index of the point cloud, calculating the density of any point, and removing isolated points and scattered noise points according to the density value of the point;
step 1.2, filtering the point cloud by using a ground elevation reference value: for the point cloud obtained in the step 1.1 after denoising, calculating the average elevation of the point with the normal vector vertical to the ground and the elevation value meeting the preset threshold value
Figure GDA0002107466590000061
And standard deviation of height ε willAnd setting the average elevation as a ground elevation reference value, taking the standard deviation of the elevation as a tolerance threshold, judging the point cloud after denoising, and separating out ground points and non-ground points. In specific implementation, the elevation threshold value can be obtained by adding a preset parameter value to the minimum point cloud elevation value, for example 2 meters.
As shown in formula (1), any point pi and the average elevation
Figure GDA0002107466590000062
If the difference value of (A) is less than the standard deviation epsilon of the height difference, the point is a ground point, otherwise, the point is a non-ground point.
Figure GDA0002107466590000071
In an embodiment, step 1 is implemented as follows:
when the dot density is calculated point by point, the radius is 1.5m, the density threshold value is 10, and when the dot density is less than 10, the dot is marked as an isolated noise dot.
And calculating normal vectors of each point after denoising, searching for any point K as 100 adjacent points, and performing least square fitting on planes of the adjacent points to obtain the normal vectors of the points. And judging whether the normal vector of the point is vertical to the ground or not (in specific implementation, an angle deviation range can be preset, approximate vertical is allowed), and the elevation value is less than 13.03 meters (the minimum elevation value of the point cloud is 11.03 meters plus 2 meters), if so, reserving, and repeating the steps until all the points are judged. The average elevation of the ground points judged in the previous step is counted
Figure GDA0002107466590000072
11.76m and a standard deviation of height ε of 0.4 m in terms of average elevationAnd judging all points by adopting a formula (1) according to the sum-height standard deviation epsilon, and dividing non-ground points and ground points.
And 2, segmenting the point cloud to generate the hyper-voxels. And for the non-ground points, adopting a density-based spatial clustering algorithm DBSCAN to segment, generating the hyper-voxels, and ensuring that different ground objects do not have adhesion, namely, mixed hyper-voxels do not exist.
The step 2 of segmenting the point cloud and generating the hyper-voxels comprises the following steps:
step 2.1, analyzing the point cloud density and setting point neighborhood searching range Eps; segmenting the point cloud according to the elevation values, and giving the minimum sample number Minpts corresponding to the point cloud in different elevation ranges;
step 2.2, segmenting the point cloud by using a DBSCAN algorithm:
(1) marking the initial state of all points in the point cloud as unmarked, and establishing a point queue to be processed.
(2) And taking any unmarked point p in the point cloud, obtaining the corresponding minimum sample number Minpts according to the elevation value of the point p, and calculating the neighborhood density value pValue of the point p in the neighborhood search range Eps.
If pValue is less than Minpts, mark point p as a noise point.
If the pValue is greater than or equal to Minpts, marking the point p and the neighborhood points thereof as a new point cluster C, and adding all the neighborhood points of the point p into the point queue to be processed.
(3) And (3) calculating the neighborhood density value qValue for any unprocessed point q in the point queue to be processed, and if the qValue is greater than or equal to Minpts, marking the unmarked point in the neighborhood point of q as the point cluster C and adding the point cluster C into the point queue to be processed.
(4) And (3) sequentially carrying out operation on the points in the point queue to be processed until all the points are processed, and emptying the point queue to be processed.
(5) Repeating (2) - (4) until all points are marked, completing the segmentation.
And 2.3, performing point number constraint on the point clusters obtained by segmentation, if the point number is too small, marking the points in the point clusters as noise points, and enabling the rest point clusters meeting the constraint to become hyper-voxels. In specific implementation, a corresponding point threshold value can be preset, and if the point number of a certain point cluster is smaller than the threshold value, the point is marked as a noise point.
In the embodiment, a DBSCAN algorithm is used for voxel segmentation, the search neighborhood radius Eps is 0.6m, when the elevation value of a point is between 14.34m and 22.34m, MinPts is 450, otherwise, MinPts is 50, and 212 voxels are obtained after segmentation.
And 3, clustering the point clouds related to the spatial context. Analyzing the characteristics and spatial context association of the hyper-voxels, and synthesizing the multi-factor weight to increase the hyper-voxels regions to complete point cloud clustering.
The step 3 of combining spatial context association clustering comprises the following steps:
and 3.1, calculating the characteristics of each hyper-voxel, including dimension information VD of the hyper-voxel, three-dimensional coordinates V (X, Y, Z) of the center point and bounding box information (Vmax, Vmin).
The embodiment utilizes PCA (principal component analysis) algorithm to calculate the dimension of each point in the superpixel, counts three-dimensional, two-dimensional and one-dimensional points, and assigns the dimension with the maximum points to the superpixel.
For any point p, the neighborhood point set is set as the nearest K points q contained in the sphere by taking the point p as the center of the spherekSet of (2), set center of gravity
Figure GDA0002107466590000081
Comprises the following steps:
Figure GDA0002107466590000082
the calculation matrices M and C are shown in equations (3) and (4):
Figure GDA0002107466590000084
performing characteristic decomposition on the matrix C to obtain sorted characteristic values of lambda respectively1,λ2,λ3(0＜λ3≤λ2≤λ1)，
Calculating the parameter a according to equation (5)1D,a2DAnd a3DOne-dimensional, two-dimensional and three-dimensional information of the point p is described, respectively.
Figure GDA0002107466590000091
Where μ is a regularization coefficient, and μ ═ σ is defined1，
Figure GDA0002107466590000092
t is 1,2, 3. If a isdDAt the maximum, the point p has dimension d, d being 1,2, 3.
The three-dimensional coordinates V (X, Y, Z) of the center point are averaged from the three-dimensional coordinates of all points in the superpixel.
Wherein, (Xn, Yn, Zn) is the coordinate of any point in the superpixel, and N is the number of points contained in the superpixel.
The bounding box information (Vmax, Vmin) is the maximum minimum X, Y and Z value of the cuboid bounding box of the hyper-voxel, Vmax (X1, Y1, Z1) and Vmin (X2, Y2, Z2) are:
Figure GDA0002107466590000094
Figure GDA0002107466590000095
step 3.2, according to a preset initial value Rs of the search radius, calculating spatial correlation characteristics of the superpixels, including dimensional consistency of the superpixels and plane distance D between center points of the superpixelscHeight difference H between center points of voxels of supervoxelcAnd the hyper-voxel bounding boxes intersect or meet in the vertical ground direction (OZ direction). DcAnd HcAccording to equation (8)
Figure GDA0002107466590000096
Wherein, (X, Y, Z) and (X ', Y ', Z ') are three-dimensional coordinates of the center points of adjacent superpixels respectively.
The intersection or junction of the hyper-voxel bounding boxes in the direction perpendicular to the ground (set as the OZ direction of the coordinate system O-XYZ) means that the rectangles projected on the XOY plane of the hyper-voxel bounding boxes are mutually contained and the rectangles projected on the XOZ plane are intersected or connected, see fig. 1.
The spatial correlation characteristics selected by the invention are respectively used for describing main context correlation existing in the spatial distribution of the urban street view ground objects. For city street view features, the dimension information of the same or the same feature is the same, and the same or the same feature meets certain geometric constraints on the spatial relationship, such as being located at the same height and being distributed horizontally or vertically. If the power lines are in linear distribution and the adjacent points are basically at the same height, the telegraph poles are in linear vertical distribution, and the trees are the combination of three-dimensional points and two-dimensional points, so that the vertical distribution characteristic is met. After the ground objects of the street view are over-segmented, the obtained multiple superpixels still meet the spatial distribution geometric constraint of the ground objects. Wherein the consistency of the superpixel dimension describes the correlation of the superpixel dimension and the plane distance D between the center points of the superpixelcHeight difference H between center points of voxels of supervoxelcAnd the voxel bounding boxes intersect or meet in the vertical ground direction (OZ direction) to describe the geometrical constraint of voxel space distribution.
Step 3.3, performing region growing clustering by combining spatial context association of the hyper voxels, wherein the growing condition W is a weight of comprehensive multifactor among the hyper voxels, and the formula (9) shows:
Figure GDA0002107466590000101
wherein, CuFor the spatial context correlation factor of hyper-voxels, wuAnd u is 1,2,3 and 4, which is the contribution degree of each factor to the clustering.
Wherein, C1For dimension correlation factor, describe the same property of the same ground physical dimension, when the two superpixel dimension information are the same, C1The value is 1, otherwise, the value is 0; c2Describing geometric constraint of ground object space plane distribution for the plane position association factor, and when the distance Dc between two hyper-voxel planes is less than a preset threshold value, C2The value is 1, otherwise, the value is 0; c3Describing the height of the ground object for the height correlation factorDistribution geometric constraint, when the height difference Hc of two hyper-voxels is less than a preset threshold value, C3The value is 1, otherwise, the value is 0; c4Describing the geometric constraint of the vertical distribution of the ground object space for the position correlation factor in the vertical direction, and when two hyper-voxels intersect or are connected in the direction vertical to the ground, C4The value is 1, otherwise 0.
Figure GDA0002107466590000102
Figure GDA0002107466590000103
Figure GDA0002107466590000104
Figure GDA0002107466590000105
wuThe value of (A) is determined by a correlation factor C which is preferentially related to the spatial position of the ground feature2、C3And C4Followed by a dimension correlation factor C1Therefore W is2、W3And W4Respectively take the values of 0.3 and W1The value is 0.1.
When the growth condition W satisfies the threshold, the super voxel is grown until the clustering is completed, see fig. 2.
In specific implementation, the values of the threshold of the growth condition W, the plane distance threshold D and the height difference threshold H can be preset according to the distribution density of the actual clustered ground objects. In the embodiment, after the characteristics of the hyper-voxels are calculated, the computation of the hyper-voxel clustering weight is performed. And (3) setting the initial value Rs of the search radius to be 2m, calculating the spatial correlation characteristics among the superpixels, and increasing when the increasing condition W of the superpixels is more than or equal to 2. The plane distance threshold D between the center points of the hyper-voxels, the height difference threshold H between the center points of the hyper-voxels and the threshold of the intersection or connection of the hyper-voxel bounding boxes in the vertical direction are all 0.1 m.
In order to better illustrate the clustering effect of the invention, the number of the main four types of ground object clustering clusters in the experimental results is counted and compared with the DBSCAN algorithm and the artificial clustering respectively, and the results are shown in the table I. The comparison shows that the clustering result of the method is closer to the artificial clustering result, and the number of the clustering clusters is greatly reduced compared with the DBSCAN algorithm.
TABLE-Cluster result comparison (Unit: Cluster)
Type of ground feature Feature 1 Ground object 2 Ground object 3 Ground object 4
DBSCAN 53 12 44 103
The method of the invention 14 5 29 20
Artificial clustering 6 4 30 14
Therefore, the invention has the following advantages: in the vehicle-mounted point cloud clustering, the correlation of the space distance between points is considered, the data is organized by the superpixel, the space context correlation of the points is mined, and the vehicle-mounted point cloud clustering effect is effectively improved.
In specific implementation, the method provided by the invention can realize automatic operation flow based on software technology, and can also realize a corresponding system in a modularized mode.
The embodiment of the invention also provides a vehicle-mounted point cloud clustering system which comprises the following modules,
the first module is used for point cloud denoising and filtering, including removing scattered noise points from the point cloud, and carrying out point cloud filtering processing to distinguish ground points and non-ground points;
the second module is used for segmenting the point cloud to generate a hyper-voxel, and comprises a step of segmenting non-ground points by adopting a density-based spatial clustering algorithm to generate the hyper-voxel so as to ensure that different ground objects are not adhered;
the third module is used for point cloud clustering of spatial context association, including analyzing the characteristics of the hyper-voxels and spatial context association, and performing regional growth of the hyper-voxels by synthesizing multi-factor weights to finish the point cloud clustering, and comprises the following units, namely a first unit, a second unit, a third module and a third module, wherein the first unit is used for calculating the characteristics of each hyper-voxel, including the dimension information VD of the hyper-voxel, the three-dimensional coordinates V (X, Y, Z) of a central point and bounding box information (Vmax, Vmin);
the dimensionality information VD of the hyper-voxel is calculated in a mode that the dimensionality of each point in the hyper-voxel is calculated, three-dimensional, two-dimensional and one-dimensional point numbers are counted, and the dimensionality with the largest point number is assigned to the hyper-voxel;
the three-dimensional coordinates V (X, Y, Z) of the central point are obtained by averaging the three-dimensional coordinates of all points in the hyper-voxels;
bounding box information (Vmax, Vmin) is obtained from a cuboid bounding box of the hyper-voxel;
a second unit for calculating spatial correlation characteristics of the superpixel according to a preset initial value Rs of the search radius, including the dimension consistency of the superpixel and the plane distance D between the center points of the superpixelcHeight difference H between center points of voxels of supervoxelcAnd the voxel bounding boxes intersect or are connected in the direction vertical to the ground;
a third unit, for performing region growing clustering by combining the spatial context association of the hyper-voxels, wherein the growing condition W is the weight of the comprehensive multifactor among the hyper-voxels, and the calculation is as follows
Figure GDA0002107466590000121
Wherein, CuFor the spatial context correlation factor of hyper-voxels, wuAs the contribution degree of each factor to the clustering, u is 1,2,3, 4;
C1for dimension correlation factor, describe the same property of the same ground physical dimension, when the two superpixel dimension information are the same, C1The value is 1, otherwise, the value is 0;
C2describing geometric constraint of ground object space plane distribution for the plane position association factor, and when the distance Dc between two hyper-voxel planes is less than a preset threshold value, C2The value is 1, otherwise, the value is 0;
C3describing geometric constraint of ground object height distribution for a height correlation factor, and when the height difference Hc of two hyper-voxels is smaller than a preset threshold value, C3The value is 1, otherwise, the value is 0;
C4describing the geometric constraint of the vertical distribution of the ground object space for the position correlation factor in the vertical direction, and when two hyper-voxels intersect or are connected in the direction vertical to the ground, C4The value is 1, otherwise 0.
The specific implementation of each module can refer to the corresponding step, and the detailed description of the invention is omitted.
The specific embodiments described herein are merely illustrative of the spirit of the invention. Various modifications or additions may be made to the described embodiments or alternatives may be employed by those skilled in the art without departing from the spirit or ambit of the invention as defined in the appended claims.

Claims (10)

1. A vehicle-mounted point cloud clustering method is characterized by comprising the following steps: comprises the following steps of (a) carrying out,
step 1, point cloud denoising and filtering, including removing scattered noise points from the point cloud, and carrying out point cloud filtering processing to distinguish ground points and non-ground points;
step 2, segmenting point clouds to generate hyper-voxels, wherein non-ground points are segmented by adopting a density-based spatial clustering algorithm to generate the hyper-voxels, so that the condition that different ground objects are not adhered is ensured;
step 3, spatial context-associated point cloud clustering, which comprises the steps of analyzing the characteristics of the hyper-voxels and spatial context association, and integrating multi-factor weights to increase the hyper-voxels regions to finish point cloud clustering,
step 3.1, calculating the characteristics of each hyper-voxel, including dimension information VD of the hyper-voxel, three-dimensional coordinates V (X, Y, Z) of a central point and bounding box information (Vmax, Vmin);
the dimensionality information VD of the hyper-voxel is calculated in a mode that the dimensionality of each point in the hyper-voxel is calculated, three-dimensional, two-dimensional and one-dimensional point numbers are counted, and the dimensionality with the largest point number is assigned to the hyper-voxel;
the three-dimensional coordinates V (X, Y, Z) of the central point are obtained by averaging the three-dimensional coordinates of all points in the hyper-voxels;
bounding box information (Vmax, Vmin) is obtained from a cuboid bounding box of the hyper-voxel;
step 3.2, according to a preset initial value Rs of the search radius, calculating spatial correlation characteristics of the superpixels, including dimensional consistency of the superpixels and plane distance D between center points of the superpixelscHeight difference H between center points of voxels of supervoxelcAnd the voxel bounding boxes intersect or are connected in the direction vertical to the ground;
step 3.3, combining the spatial context association of the hyper voxels to carry out region growing clustering, wherein the growing condition W is the weight of the comprehensive multifactor among the hyper voxels, and the calculation is as follows
Wherein, CuFor the spatial context correlation factor of hyper-voxels, wuAs the contribution degree of each factor to the clustering, u is 1,2,3, 4;
C1for dimension correlation factor, describe the same property of the same ground physical dimension, when the two superpixel dimension information are the same, C1The value is 1, otherwise, the value is 0;
C2describing geometric constraint of ground object space plane distribution for the plane position association factor, and when the distance Dc between two hyper-voxel planes is less than a preset threshold value, C2The value is 1, otherwise, the value is 0;
C3describing geometric constraint of ground object height distribution for a height correlation factor, and when the height difference Hc of two hyper-voxels is smaller than a preset threshold value, C3The value is 1, otherwise, the value is 0;
C4describing the geometric constraint of the vertical distribution of the ground object space for the position correlation factor in the vertical direction, and when two hyper-voxels intersect or are connected in the direction vertical to the ground, C4The value is 1, otherwise 0.
2. The vehicle-mounted point cloud clustering method according to claim 1, characterized in that: the calculation of the dimensions of the points within the hyper-voxel is carried out as follows,
for any point p, a neighborhood point set is set as the nearest K points q contained in a sphere by taking the point p as the center of the spherekSet of (2), set center of gravity
Figure FDA0002118263920000021
As follows below, the following description will be given,
Figure FDA0002118263920000022
the matrices M and C are calculated as follows,
Figure FDA0002118263920000023
Figure FDA0002118263920000024
for matrixC, performing characteristic decomposition to obtain sorted characteristic values of lambda respectively1,λ2,λ3，
Calculating the parameter a1D,a2DAnd a3DAs follows, the one-dimensional, two-dimensional and three-dimensional information of the point p is described separately,
Figure FDA0002118263920000025
where μ is a regularization coefficient, and μ ═ σ is defined1，
Figure FDA0002118263920000026
t is 1,2,3, if adDAt the maximum, the point p dimension is d.
3. The vehicle-mounted point cloud clustering method according to claim 1, characterized in that: and the OZ direction of the coordinate system O-XYZ is taken as the vertical ground direction, the intersection or connection of the hyper-voxel bounding boxes in the vertical ground direction is represented as the intersection or connection of rectangles projected by the hyper-voxel bounding boxes on the XOY plane, and the rectangles projected on the XOZ plane are intersected or connected.
4. The vehicle-mounted point cloud clustering method according to claim 1, characterized in that: dcAnd HcThe calculation is as follows,
Figure FDA0002118263920000027
HC＝|Z-Z′|
wherein, (X, Y, Z) and (X ', Y ', Z ') are three-dimensional coordinates of the center points of adjacent superpixels respectively.
5. The vehicle-mounted point cloud clustering method according to claim 1,2,3 or 4, characterized in that: w is auThe value of (A) is as follows2、W3And W4Is 0.3 greater than W1The value of (a).
6. The vehicle-mounted point cloud clustering system is characterized in that: comprises the following modules which are used for realizing the functions of the system,
the first module is used for point cloud denoising and filtering, including removing scattered noise points from the point cloud, and carrying out point cloud filtering processing to distinguish ground points and non-ground points;
the second module is used for segmenting the point cloud to generate a hyper-voxel, and comprises a step of segmenting non-ground points by adopting a density-based spatial clustering algorithm to generate the hyper-voxel so as to ensure that different ground objects are not adhered;
the third module is used for spatial context-associated point cloud clustering, comprises the following units of analyzing the characteristics of the hyper-voxels and spatial context association, carrying out regional growth of the hyper-voxels by integrating multi-factor weights and finishing the point cloud clustering,
a first unit for calculating the characteristics of each hyper-voxel, including dimension information VD of the hyper-voxel, three-dimensional coordinates V (X, Y, Z) of the center point and bounding box information (Vmax, Vmin);
the dimensionality information VD of the hyper-voxel is calculated in a mode that the dimensionality of each point in the hyper-voxel is calculated, three-dimensional, two-dimensional and one-dimensional point numbers are counted, and the dimensionality with the largest point number is assigned to the hyper-voxel;
the three-dimensional coordinates V (X, Y, Z) of the central point are obtained by averaging the three-dimensional coordinates of all points in the hyper-voxels;
bounding box information (Vmax, Vmin) is obtained from a cuboid bounding box of the hyper-voxel;
a second unit for calculating spatial correlation characteristics of the superpixel according to a preset initial value Rs of the search radius, including the dimension consistency of the superpixel and the plane distance D between the center points of the superpixelcHeight difference H between center points of voxels of supervoxelcAnd the voxel bounding boxes intersect or are connected in the direction vertical to the ground;
a third unit, for performing region growing clustering by combining the spatial context association of the hyper-voxels, wherein the growing condition W is the weight of the comprehensive multifactor among the hyper-voxels, and the calculation is as follows
Figure FDA0002118263920000031
Wherein, CuFor the spatial context correlation factor of hyper-voxels, wuAs the contribution degree of each factor to the clustering, u is 1,2,3, 4;
C1for dimension correlation factor, describe the same property of the same ground physical dimension, when the two superpixel dimension information are the same, C1The value is 1, otherwise, the value is 0;
C2describing geometric constraint of ground object space plane distribution for the plane position association factor, and when the distance Dc between two hyper-voxel planes is less than a preset threshold value, C2The value is 1, otherwise, the value is 0;
C3describing geometric constraint of ground object height distribution for a height correlation factor, and when the height difference Hc of two hyper-voxels is smaller than a preset threshold value, C3The value is 1, otherwise, the value is 0;
C4describing the geometric constraint of the vertical distribution of the ground object space for the position correlation factor in the vertical direction, and when two hyper-voxels intersect or are connected in the direction vertical to the ground, C4The value is 1, otherwise 0.
7. The vehicle-mounted point cloud clustering system according to claim 6, wherein: the calculation of the dimensions of the points within the hyper-voxel is carried out as follows,
for any point p, a neighborhood point set is set as the nearest K points q contained in a sphere by taking the point p as the center of the spherekSet of (2), set center of gravity
Figure FDA0002118263920000041
As follows below, the following description will be given,
Figure FDA0002118263920000042
the matrices M and C are calculated as follows,
Figure FDA0002118263920000043
Figure FDA0002118263920000044
performing characteristic decomposition on the matrix C to obtain sorted characteristic values of lambda respectively1,λ2,λ3，
Calculating the parameter a1D,a2DAnd a3DAs follows, the one-dimensional, two-dimensional and three-dimensional information of the point p is described separately,
Figure FDA0002118263920000045
where μ is a regularization coefficient, and μ ═ σ is defined1，
Figure FDA0002118263920000046
t is 1,2,3, if adDAt the maximum, the point p dimension is d.
8. The vehicle-mounted point cloud clustering system according to claim 6, wherein: and the OZ direction of the coordinate system O-XYZ is taken as the vertical ground direction, the intersection or connection of the hyper-voxel bounding boxes in the vertical ground direction is represented as the intersection or connection of rectangles projected by the hyper-voxel bounding boxes on the XOY plane, and the rectangles projected on the XOZ plane are intersected or connected.
9. The vehicle-mounted point cloud clustering system according to claim 6, wherein: dcAnd HcThe calculation is as follows,
Figure FDA0002118263920000047
HC＝|Z-Z′|
wherein, (X, Y, Z) and (X ', Y ', Z ') are three-dimensional coordinates of the center points of adjacent superpixels respectively.
10. The vehicle-mounted point cloud clustering system according to claim 6, 7, 8 or 9, wherein: w is auThe value of (A) is as follows2、W3And W4Is 0.3 greater than W1The value of (a).
CN201710510689.6A 2017-06-28 2017-06-28 Vehicle-mounted point cloud clustering method and system Expired - Fee Related CN107292276B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN201710510689.6A CN107292276B (en) 2017-06-28 2017-06-28 Vehicle-mounted point cloud clustering method and system

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN201710510689.6A CN107292276B (en) 2017-06-28 2017-06-28 Vehicle-mounted point cloud clustering method and system

Publications (2)

Publication Number Publication Date
CN107292276A CN107292276A (en) 2017-10-24
CN107292276B true CN107292276B (en) 2020-01-07

Family

ID=60098877

Family Applications (1)

Application Number Title Priority Date Filing Date
CN201710510689.6A Expired - Fee Related CN107292276B (en) 2017-06-28 2017-06-28 Vehicle-mounted point cloud clustering method and system

Country Status (1)

Country Link
CN (1) CN107292276B (en)

Families Citing this family (18)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110349158A (en) * 2018-04-04 2019-10-18 北京京东尚科信息技术有限公司 A kind of method and apparatus handling point cloud data
WO2019237319A1 (en) * 2018-06-15 2019-12-19 Bayerische Motoren Werke Aktiengesellschaft Incremental segmentation of point cloud
CN109633686B (en) * 2018-11-22 2021-01-19 浙江中车电车有限公司 Method and system for detecting ground obstacle based on laser radar
CN109934120B (en) * 2019-02-20 2021-04-23 东华理工大学 Step-by-step point cloud noise removing method based on space density and clustering
CN110060288A (en) * 2019-03-15 2019-07-26 华为技术有限公司 Generation method, device and the storage medium of point cloud characteristic pattern
CN110046661A (en) * 2019-04-10 2019-07-23 武汉大学 A kind of vehicle-mounted cloud clustering method cutting algorithm based on contextual feature and figure
CN110275153B (en) * 2019-07-05 2021-04-27 上海大学 Water surface target detection and tracking method based on laser radar
CN110673158B (en) * 2019-08-28 2022-10-04 国网福建省电力有限公司 Laser scanning point cloud-based three-dimensional space anomaly detection method for power production scene
WO2021035618A1 (en) * 2019-08-29 2021-03-04 深圳市大疆创新科技有限公司 Point cloud segmentation method and system, and movable platform
CN111145182B (en) * 2019-12-30 2022-05-27 芜湖哈特机器人产业技术研究院有限公司 Visual positioning three-dimensional point cloud segmentation method
CN113496160B (en) * 2020-03-20 2023-07-11 百度在线网络技术（北京）有限公司 Three-dimensional object detection method, three-dimensional object detection device, electronic equipment and storage medium
CN111723866A (en) * 2020-06-19 2020-09-29 新石器慧通（北京）科技有限公司 Point cloud clustering method and device, unmanned vehicle and readable storage medium
CN111860321B (en) * 2020-07-20 2023-12-22 浙江光珀智能科技有限公司 Obstacle recognition method and system
CN112070769B (en) * 2020-09-18 2022-06-03 福州大学 Layered point cloud segmentation method based on DBSCAN
CN112529044B (en) * 2020-11-20 2022-06-28 西南交通大学 Method for extracting and classifying railway contact network based on vehicle-mounted LiDAR
CN112462347B (en) * 2020-12-28 2023-08-15 长沙理工大学 Laser radar point cloud rapid classification filtering algorithm based on density clustering
CN116071694B (en) * 2023-03-07 2023-06-23 浙江华是科技股份有限公司 Ship detection method, device and computer readable storage medium
CN116309566B (en) * 2023-05-17 2023-09-12 深圳大学 Point cloud-based adhesive artificial rod-shaped object monomer extraction method and related equipment

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101702200A (en) * 2009-11-03 2010-05-05 武汉大学 Automatic classification method of airborne laser radar point cloud data
CN105701448A (en) * 2015-12-31 2016-06-22 湖南拓视觉信息技术有限公司 Three-dimensional face point cloud nose tip detection method and data processing device using the same
CN106599915A (en) * 2016-12-08 2017-04-26 立得空间信息技术股份有限公司 Vehicle-mounted laser point cloud classification method

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2532948B (en) * 2014-12-02 2021-04-14 Vivo Mobile Communication Co Ltd Object Recognition in a 3D scene

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101702200A (en) * 2009-11-03 2010-05-05 武汉大学 Automatic classification method of airborne laser radar point cloud data
CN105701448A (en) * 2015-12-31 2016-06-22 湖南拓视觉信息技术有限公司 Three-dimensional face point cloud nose tip detection method and data processing device using the same
CN106599915A (en) * 2016-12-08 2017-04-26 立得空间信息技术股份有限公司 Vehicle-mounted laser point cloud classification method

Also Published As

Publication number Publication date
CN107292276A (en) 2017-10-24

Similar Documents

Publication Publication Date Title
CN107292276B (en) Vehicle-mounted point cloud clustering method and system
CN111932688B (en) Indoor plane element extraction method, system and equipment based on three-dimensional point cloud
Lu et al. Fast 3D line segment detection from unorganized point cloud
CN106022381B (en) Automatic extraction method of street lamp pole based on vehicle-mounted laser scanning point cloud
Lu et al. Pairwise linkage for point cloud segmentation
CN112070769B (en) Layered point cloud segmentation method based on DBSCAN
CN108510516A (en) A kind of the three-dimensional line segment extracting method and system of dispersion point cloud
CN111915730A (en) Method and system for automatically generating indoor three-dimensional model from point cloud in consideration of semantics
CN110349260B (en) Automatic pavement marking extraction method and device
CN111986322B (en) Point cloud indoor scene layout reconstruction method based on structural analysis
CN109215112B (en) Method for marking single-side point cloud model
CN113920134B (en) Slope ground point cloud segmentation method and system based on multi-line laser radar
CN109493344A (en) A kind of semantic segmentation method of large-scale city three-dimensional scenic
CN114359876B (en) Vehicle target identification method and storage medium
CN115205690B (en) Method and device for extracting street tree in monomer mode based on MLS point cloud data
CN113920360A (en) Road point cloud rod extraction and multi-scale identification method
CN113484875A (en) Laser radar point cloud target hierarchical identification method based on mixed Gaussian ordering
CN116258857A (en) Outdoor tree-oriented laser point cloud segmentation and extraction method
Tian et al. Robust segmentation of building planar features from unorganized point cloud
CN112200248B (en) Point cloud semantic segmentation method, system and storage medium based on DBSCAN clustering under urban road environment
CN111861946B (en) Adaptive multi-scale vehicle-mounted laser radar dense point cloud data filtering method
Sun et al. Automated segmentation of LiDAR point clouds for building rooftop extraction
CN116071530B (en) Building roof voxelized segmentation method based on airborne laser point cloud
Hao et al. Automatic building extraction from terrestrial laser scanning data
CN115170950A (en) Outdoor scene building extraction method based on multi-feature constraint

Legal Events

Date Code Title Description
PB01 Publication
PB01 Publication
SE01 Entry into force of request for substantive examination
SE01 Entry into force of request for substantive examination
GR01 Patent grant
GR01 Patent grant
CF01 Termination of patent right due to non-payment of annual fee
CF01 Termination of patent right due to non-payment of annual fee

Granted publication date: 20200107

Termination date: 20200628