CN112070787A - Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory - Google Patents

Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory Download PDF

Info

Publication number
CN112070787A
CN112070787A CN202010794927.2A CN202010794927A CN112070787A CN 112070787 A CN112070787 A CN 112070787A CN 202010794927 A CN202010794927 A CN 202010794927A CN 112070787 A CN112070787 A CN 112070787A
Authority
CN
China
Prior art keywords
plane
hyper
points
subset
voxels
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.)
Granted
Application number
CN202010794927.2A
Other languages
Chinese (zh)
Other versions
CN112070787B (en
Inventor
张永军
祝宪章
刘欣怡
万一
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 CN202010794927.2A priority Critical patent/CN112070787B/en
Publication of CN112070787A publication Critical patent/CN112070787A/en
Application granted granted Critical
Publication of CN112070787B publication Critical patent/CN112070787B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/136—Segmentation; Edge detection involving thresholding
    • G—PHYSICS
    • G01—MEASURING; TESTING
    • G01S—RADIO DIRECTION-FINDING; RADIO NAVIGATION; DETERMINING DISTANCE OR VELOCITY BY USE OF RADIO WAVES; LOCATING OR PRESENCE-DETECTING BY USE OF THE REFLECTION OR RERADIATION OF RADIO WAVES; ANALOGOUS ARRANGEMENTS USING OTHER WAVES
    • G01S17/00—Systems using the reflection or reradiation of electromagnetic waves other than radio waves, e.g. lidar systems
    • G01S17/88—Lidar systems specially adapted for specific applications
    • G01S17/89—Lidar systems specially adapted for specific applications for mapping or imaging
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/187—Segmentation; Edge detection involving region growing; involving region merging; involving connected component labelling
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10028—Range image; Depth image; 3D point clouds

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Theoretical Computer Science (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Electromagnetism (AREA)
  • Radar, Positioning & Navigation (AREA)
  • Remote Sensing (AREA)
  • Image Generation (AREA)
  • Image Processing (AREA)

Abstract

The invention provides an aviation three-dimensional point cloud plane segmentation method based on an opponent inference theory, which comprises the steps of firstly segmenting an original input point cloud into multi-scale initial plane voxels and independent points by utilizing TBBP (tunnel boring bar) voxel segmentation methods with different resolution parameters and voxel significance geometric characteristics; then, aiming at each multi-scale initial plane superpixel, separating an initial strict plane superpixel and an independent point from each multi-scale initial plane superpixel according to a contradictory reasoning theory; and searching strict plane superpixels with the smallest distance between the tangent planes in the neighborhood range of each independent point, judging whether the tangent planes are combined or not according to a distance threshold, and further taking the remaining independent points as new input point clouds to repeat the previous steps to generate more strict plane superpixels. And finally, taking strict plane hyper-voxels as basic units, and performing region growing according to tangent plane distances and normal vector angles among the hyper-voxels to generate a complete plane set.

Description

Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory
Technical Field
The invention belongs to the fields of surveying and mapping science and technology, and relates to an aviation three-dimensional point cloud plane segmentation technology based on an opposite reasoning theory.
Background
The three-dimensional point cloud obtained by an airborne laser radar system and a multi-view stereo vision (MVS) technology is widely applied to the aspects of urban three-dimensional modeling, land monitoring, cultural protection, geographical registration and the like. The plane feature is one of the most widely used three-dimensional features, and is widely applied to building information model reconstruction (BIM), simultaneous localization and mapping (SLAM), target extraction and recognition, point cloud data compression, and the like. However, the point cloud obtained by scanning is generally disordered and contains various errors and noises, and how to separate an effective plane structure from discrete points is a key problem affecting the subsequent application of the point cloud.
Three-dimensional plane segmentation is a process of clustering three-dimensional points that are spatially continuous and belong to the same plane into groups or structures. The existing three-dimensional plane segmentation method based on geometric features can be roughly divided into three categories: the first is based on region growing, selecting one or more seed points from a point set according to the curvature of each point, finding out the neighboring points of each seed point and expanding the growing region according to predefined similarity features (such as normal vector and curvature), which can generally maintain the boundaries of planes, but the segmentation quality depends on the selection of seed points. The second method is based on model fitting, and is to fit local or global points mathematically through a preset plane equation, which is easily limited by the point cloud quality and parameter setting. The third method is based on feature clustering, and the neighbors are grouped according to the similarity of geometric attributes such as normal vector features, Euclidean distance and density, and the method is sensitive to noise and abnormal values and is also influenced by neighborhood definition. In addition to these three types of methods, some methods that have proven effective in two-dimensional image geometric primitive extraction are also applied to three-dimensional point cloud plane segmentation, such as scan line analysis and global energy optimization. The scanning line analysis-based method combines and groups scanning contours through geometric constraint and a specific similarity criterion, has the characteristics of quickness and stability, but the segmentation result depends on selection of a priority direction, and the application range is limited to structured point cloud. The method based on energy optimization firstly generates an initial plane set by using a method based on geometry, and then minimizes the energy function by constructing the energy function and refining the plane label of points, which can effectively inhibit the influence of noise points and outliers in some cases, but is easy to generate a locally optimal situation due to the initial segmentation result.
Disclosure of Invention
The invention aims to provide an aviation three-dimensional point cloud plane segmentation method based on an opponent reasoning theory, and solves the problems that the noise level is high, the point cloud with a complex structure is difficult to extract an accurate plane, and the plane parameter precision is low. The invention introduces the opposite reasoning theory and finds out the optimal plane subset by calculating the disorder metric values of different plane subsets. Compared with other methods in the prior art, the method adopts the opposite reasoning idea and the multi-scale hyper-voxel segmentation strategy which are not sensitive to parameters, the generated plane integrity and recall rate are very high, the corresponding plane parameters also have extremely high precision and strong robustness, and the method can be used for dealing with various complex and diverse aviation three-dimensional point clouds including an airborne sparse LiDAR point cloud, an airborne dense LiDAR point cloud and an aviation image dense matching point cloud. Therefore, the invention has important practical value and wide application prospect.
In order to achieve the aim, the technical scheme provided by the invention is an aviation three-dimensional point cloud plane segmentation method based on an opposite reasoning theory, which comprises the following steps of:
step 1, data preparation. The input data which can be processed by the method comprises the following steps: 1) airborne LiDAR point clouds; 2) and obtaining aerial image dense matching point clouds through an aerial image and a dense matching algorithm.
Step 2, segmenting the original input point cloud into multi-scale initial plane hyper-voxels and independent points by using TBBP hyper-voxel segmentation methods with different resolution parameters and hyper-voxel significance geometric features;
step 3, extracting the optimal plane subset of each multi-scale initial plane superpixel according to an opponent reasoning theory, thereby separating initial strict plane superpixel and independent points from each multi-scale initial plane superpixel;
step 4, searching strict plane hyper-voxels with the smallest distance of the tangent planes in the neighborhood range of each independent point, judging whether to combine the strict plane hyper-voxels according to a distance threshold value, and further taking the remaining independent points as new input point clouds to repeat the steps 1 to 3 to generate more strict plane hyper-voxels;
and 5, taking strict plane hyper-voxels as basic units, and performing region growing according to tangent plane distances and normal vector angles among the hyper-voxels to generate a complete plane set.
In the above three-dimensional point cloud plane segmentation method based on the opponent inference theory, in step 2, the specific method of multi-scale hyper-voxel segmentation is as follows:
and 2.1, carrying out hyper-voxel segmentation. And performing hyper-voxel segmentation on the input point cloud by adopting a TBBP (heated target boundary segmentation for 3D point clusters) algorithm to obtain a hyper-voxel set.
And 2.2, classifying the hyper-voxels. The generated hyper-voxel sets are classified into planar hyper-voxels and non-planar hyper-voxels according to the significance geometrical characteristics by adopting the method of Dong et al (2018). For a hyper-voxel SV ═ p containing arbitrarily n points1，...，pnOf its covariance matrix M3×3The calculation formula is as follows:
Figure BDA0002625184870000021
according to M3×3Characteristic value λ of1，λ2，λ3；(λ1≥λ2≥λ3) And a feature vector e1，e2，e3The significance characteristics g of the SV can be further calculated1，g2，g3And curvature fsAs follows:
Figure BDA0002625184870000031
whether a hyper-voxel SV is classified as a planar hyper-voxel or a non-planar hyper-voxel is determined according to the following formula:
Figure BDA0002625184870000032
wherein T isThe curvature threshold value is generally 0.05.
And 2.3, inputting the points classified into the non-planar hyper-voxels into the next layer of scale, and continuously repeating the step 2.1 and the step 2.2 until the scale is reduced to the minimum value or the number of the non-planar hyper-voxels is less than a certain specified number, wherein the minimum number of the points is in a direct proportion relation with the average point distance of the input point cloud.
In the above three-dimensional point cloud plane segmentation method based on the opponent inference theory, in step 3, the optimal three-dimensional plane subset extraction method is as follows:
and 3.1, calculating the geometric consistency measure. For a random three-dimensional point set S comprising n points, a virtual hypothesis H is first defined0It is used to indicate that there is no meaningful planar structure in S. The method comprises the following specific steps:
step 3.1.1, for a particular candidate plane PcCalculating an arbitrary point p in SiAt PcGeometric consistency measure of (2):
Figure BDA0002625184870000033
wherein d (p)i，Pc) To represent a point piTo PcThe distance tolerance value τ is used to indicate that there is a greater likelihood that the point P belongs to the planar model P when d (P, P) ≦ τ.
Step 3.1.2, denote S as belonging to an arbitrary subset of planes of S and satisfying
Figure BDA0002625184870000034
s is at PcThe geometric consistency measure of (a) is the maximum of the geometric consistency measures of the points within the s subset:
Figure BDA0002625184870000035
step 3.1.3, the global geometric consistency measure of S is defined as being at a certain PcObtained as above
Figure BDA0002625184870000036
Minimum value:
Figure BDA0002625184870000037
step 3.1.4, when there is a candidate plane model P derived from a subset of three pointscSuch that a subset s of k points satisfies:
Figure BDA0002625184870000038
then the geometric consistency of the subset s is said to reach the α level (or is said to satisfy the α -rigid condition), where α is a positive number. Under the null hypothesis H0The geometric consistency of the subset s satisfying the α -rigid condition is as follows:
Figure BDA0002625184870000041
wherein
Figure BDA0002625184870000042
Is shown in the imaginary hypothesis H0Lower part
Figure BDA0002625184870000043
A probability measure of (d).
And 3.2, calculating the disorder measure. The disorder metric value is obtained by calculating the NFA value of the subset of planes. Null hypothesis H as defined above0Next, the NFA value for the subset s that satisfies α -rigid is defined as:
Figure BDA0002625184870000044
wherein N issRepresenting the total number of subsets S that may exist in S. For any subset
Figure BDA0002625184870000045
The measure of disorder of s is obtained by computing (α, n, k):
Figure BDA0002625184870000046
and 3.3, obtaining the optimal plane subset. Number of points contained in the optimal subset and corresponding thereto
Figure BDA0002625184870000047
Is defined as:
Figure BDA0002625184870000048
optimal plane subset
Figure BDA0002625184870000049
Corresponding plane model
Figure BDA00026251848700000410
And the number of points contained
Figure BDA00026251848700000411
Is defined as:
Figure BDA00026251848700000412
the optimal plane subset separated from the multi-scale plane superpixel is the initial strict plane superpixel.
In the above three-dimensional point cloud plane segmentation method based on the opponent inference theory, in step 4, the independent point reclassification method is as follows:
step 4.1, for all points within a certain initial strict plane hyper-voxel, pass through the k-d treeAnd finding all neighborhood independent points of the neighborhood search radius gamma in the independent point set, deleting repeated neighborhood points, and taking the rest points as a candidate point set omega. Let an arbitrary point p in ΩΩPlanar model to initial strict planar hyper-voxels
Figure BDA00026251848700000413
Is expressed as
Figure BDA00026251848700000414
When the condition is satisfied
Figure BDA00026251848700000415
Then point p will be pointed outΩReclassification into corresponding strict plane hyper-voxels, TdTypically set to tau/2.
Step 4.2, the remaining independent point set is taken as a new input point set and steps 1 to 4.1 are performed until no new strict plane hyper-voxels are generated.
Compared with the prior art, the technology of the invention has the following advantages and beneficial effects:
1) the input point cloud is divided into initial plane voxels with a plurality of scales through a plurality of groups of TBBP hyper-voxel segmentation and significance characteristic operators, so that the number of basic processing units is greatly reduced while the point cloud contour information is retained to a large extent, and the efficiency of the method is greatly improved compared with that of a method completely based on point units;
2) the plane segmentation problem is converted into the minimum disorder measure extraction problem through an optimal plane subset extraction method based on the opponent inference theory, and therefore the optimal plane subset is extracted from the initial plane superpixel and serves as a strict plane superpixel. The method can effectively obtain accurate plane subsets and corresponding high-precision plane parameters thereof, so that the method can be suitable for various complex and diversified aviation three-dimensional point clouds without being limited by the interference of the problems of large noise of the point clouds, more outliers, complex structures and the like.
3) The region growing method using strict plane hyper-voxels as the only basic growing unit can ensure the integrity of the finally generated plane, so that the method can quickly and effectively acquire the three-dimensional plane set with high precision, high integrity and high recall rate.
Drawings
FIG. 1 is a flow chart of the present invention.
Fig. 2 is a comparison of the results of processing an aerial three-dimensional point cloud by the method of the present invention according to an embodiment of the present invention with the conventional RANSAC-based method, (a) is an input point cloud, (b) is the results of processing the dense aerial LiDAR point cloud by the conventional RANSAC-based method, and includes a complete view (left) and two partial enlarged views (middle and right), (c) is the results of multi-scale voxel segmentation, (d) is the results of strict planar voxel extraction, (e) is the results of reclassification of independent points, (f) is the results of planar segmentation on the dense aerial LiDAR point cloud by the method of the present invention after region growing, and includes a complete view (left) and two partial enlarged views (middle and right), (g) is the results of planar segmentation on the sparse aerial LiDAR point cloud by the method, and includes a complete view (left) and two partial enlarged views (middle and right).
Detailed Description
The following further describes the specific technical scheme of the invention according to the attached drawings and the embodiment.
The invention aims to provide an aviation three-dimensional point cloud plane segmentation method based on an opposite reasoning theory. The method introduces a contradictory reasoning theory and finds out the optimal plane subset by calculating the disorder metric values of different plane subsets. Compared with the prior art, the method has the advantages that the adopted opponent inference idea and the multi-scale hyper-voxel segmentation strategy are insensitive to parameters, the generated plane integrity and recall rate are very high, the used plane parameters also have extremely high precision and strong robustness, and the method can be used for dealing with various complex and diverse aviation three-dimensional point clouds including an airborne sparse LiDAR point cloud, an airborne dense LiDAR point cloud and an aviation image dense matching point cloud. The specific implementation method provided by the embodiment comprises the following steps:
step 1, data preparation. The input data which can be processed by the method comprises the following steps: 1) airborne LiDAR point clouds; 2) and obtaining aerial image dense matching point clouds through an aerial image and a dense matching algorithm.
Step 2, segmenting an original input point cloud into a multi-scale initial plane hyper-voxel and an independent point by using TBBP (heated between predicted hyper-voxel segmentation for 3D point clusters) [1] hyper-voxel segmentation method and hyper-voxel significance geometric characteristics with different resolution parameters;
[1]Lin,Y.,Wang,C.,Zhai,D.,Li W.,Li,J.,2018.Toward better boundary preserved supervoxel segmentation for 3D point clouds.ISPRS J.Photogramm.Rem.Sens.143,39-47.
and 2.1, carrying out hyper-voxel segmentation. And carrying out voxel segmentation on the input point cloud by adopting a TBBP algorithm to obtain a voxel set.
And 2.2, classifying the hyper-voxels. Method using Dong et al (2018) [2 ]]And classifying the generated hyper-voxel set into planar hyper-voxels and non-planar hyper-voxels according to the significance geometrical characteristics. For a hyper-voxel SV ═ p containing arbitrarily n points1，...，pnOf its covariance matrix M3×3The calculation formula is as follows:
Figure BDA0002625184870000061
according to M3×3Characteristic value λ of1，λ2，λ3；(λ1≥λ2≥λ3) And a feature vector e1，e2，e3The significance characteristics g of the SV can be further calculated1，g2，g3And curvature fsAs follows:
Figure BDA0002625184870000062
whether a hyper-voxel SV is classified as a planar hyper-voxel or a non-planar hyper-voxel is determined according to the following formula:
Figure BDA0002625184870000063
wherein T isThe curvature threshold value is generally 0.05.
[2]Dong,Z.,Yang,B.,Hu,P.,Scherer,S.,2018.An efficient global energy optimization approach for robust 3D plane segmentation of point clouds.ISPRS J.Photogramm.Remote Sens.137,112-133.
And 2.3, inputting the points classified into the non-planar hyper-voxels into the next layer of scale, and continuously repeating the step 2.1 and the step 2.2 until the scale is reduced to the minimum value or the number of the non-planar hyper-voxels is less than a certain specified number, wherein the minimum number of the points is in a direct proportion relation with the average point distance of the input point cloud.
Step 3, extracting the optimal plane subset of each multi-scale initial plane superpixel according to an opponent reasoning theory, thereby separating initial strict plane superpixel and independent points from each multi-scale initial plane superpixel;
and 3.1, calculating the geometric consistency measure. For a random three-dimensional point set S comprising n points, a virtual hypothesis H is first defined0It is used to indicate that there is no meaningful planar structure in S. The method comprises the following specific steps:
step 3.1.1, for a particular candidate plane PcCalculating an arbitrary point p in SiAt PcGeometric consistency measure of (2):
Figure BDA0002625184870000071
wherein d (p)i，Pc) To represent a point piTo PcThe distance tolerance value τ is used to indicate that there is a greater likelihood that the point P belongs to the planar model P when d (P, P) ≦ τ.
Step 3.1.2, denote S as belonging to an arbitrary subset of planes of S and satisfying
Figure BDA0002625184870000072
s is at PcThe geometric consistency measure of (a) is the maximum of the geometric consistency measures of the points within the s subset:
Figure BDA0002625184870000073
step 3.1.3, the global geometric consistency measure of S is defined as being at a certain PcObtained as above
Figure BDA0002625184870000074
Minimum value:
Figure BDA0002625184870000075
step 3.1.4, when there is a candidate plane model P derived from a subset of three pointscSuch that a subset s of k points satisfies:
Figure BDA0002625184870000076
then the geometric consistency of the subset s is said to reach the α level (or is said to satisfy the α -rigid condition), where α is a positive number. Under the null hypothesis H0The geometric consistency of the subset s satisfying the α -rigid condition is as follows:
Figure BDA0002625184870000077
wherein
Figure BDA0002625184870000078
Is shown in the imaginary hypothesis H0Lower part
Figure BDA0002625184870000079
A probability measure of (d).
And 3.2, calculating the disorder measure. The disorder metric value is obtained by calculating the NFA value of the subset of planes. Null hypothesis H as defined above0Next, the NFA value for the subset s that satisfies α -rigid is defined as:
Figure BDA00026251848700000710
wherein N issRepresenting the total number of subsets S that may exist in S. For any subset
Figure BDA00026251848700000711
The measure of disorder of s is obtained by computing (α, n, k):
Figure BDA00026251848700000712
n represents the number of points in the point set S; k represents the number of points in the subset s;
and 3.3, obtaining the optimal plane subset. Number of points contained in the optimal subset
Figure BDA00026251848700000713
And corresponding thereto
Figure BDA00026251848700000714
Is defined as:
Figure BDA0002625184870000081
optimal plane subset
Figure BDA0002625184870000082
Corresponding plane model
Figure BDA0002625184870000083
And the number of points contained
Figure BDA0002625184870000084
Is defined as:
Figure BDA0002625184870000085
and the initial optimal plane subset separated from the multi-scale plane superpixel is the initial strict plane superpixel.
Step 4, searching strict plane hyper-voxels with the smallest distance of the tangent planes in the neighborhood range of each independent point, judging whether to combine the strict plane hyper-voxels according to a distance threshold value, and further taking the remaining independent points as new input point clouds to repeat the steps 1 to 3 to generate more strict plane hyper-voxels;
step 4.1, independent points are classified again: for all points in the hyper-voxel of an initial strict plane, all neighborhood independent points of the points are found in an independent point set through a k-d tree and a neighborhood search radius gamma, repeated neighborhood points are deleted, and the rest points are used as a candidate point set omega. Let an arbitrary point p in ΩΩPlanar model to initial strict planar hyper-voxels
Figure BDA0002625184870000086
Is expressed as
Figure BDA0002625184870000087
When the condition is satisfied
Figure BDA0002625184870000088
Then point p will be pointed outΩReclassification into corresponding strict plane hyper-voxels, TdTypically set to tau/2.
Step 4.2, the remaining independent point set is taken as a new input point set and steps 1 to 4.1 are performed until no new strict plane hyper-voxels are generated.
And 5, taking strict plane hyper-voxels as basic units, and performing region growing according to tangent plane distances and normal vector angles among the hyper-voxels to generate a final complete plane set.
The above description of the embodiments is merely illustrative of the basic technical solutions of the present invention and is not limited to the above embodiments. It should be noted that: any simple modification, addition, equivalent change or modification of the described embodiments may be made by a person or team in the field to which the invention pertains without departing from the essential spirit of the invention or exceeding the scope defined by the claims.

Claims (5)

1. The aviation three-dimensional point cloud plane segmentation method based on the opponent inference theory is characterized by comprising the following steps of:
step 1, point cloud data preparation;
step 2, segmenting the original input point cloud into multi-scale initial plane hyper-voxels and independent points by using TBBP hyper-voxel segmentation methods with different resolution parameters and hyper-voxel significance geometric features;
step 3, extracting the optimal plane subset of each multi-scale initial plane superpixel according to an opponent reasoning theory, thereby separating initial strict plane superpixel and independent points from each multi-scale initial plane superpixel;
step 4, searching strict plane hyper-voxels with the smallest distance between the tangent planes in the neighborhood range of each independent point, judging whether to combine the tangent planes according to a distance threshold value, and further taking the remaining independent points as new input point clouds to repeat the steps 1 to 3 to generate more strict plane hyper-voxels;
and 5, taking strict plane hyper-voxels as basic units, and performing region growing according to tangent plane distances and normal vector angles among the hyper-voxels to generate a final complete plane set.
2. The three-dimensional point cloud plane segmentation method based on the opponent inference theory as claimed in claim 1, wherein: in step 2, the specific method of multi-scale hyper-voxel segmentation is as follows:
step 2.1, voxel segmentation is carried out; performing voxel segmentation on the input point cloud by adopting a TBBP algorithm to obtain a voxel set;
step 2.2, classifying the hyper-voxels; classifying the generated hyper-voxel set into planar hyper-voxels and non-planar hyper-voxels according to the significance geometrical characteristics, and for any hyper-voxel SV containing n points, the method is { p }1，...，PnOf its covariance matrix M3×3The calculation formula is as follows:
Figure FDA0002625184860000011
according to M3×3Characteristic value λ of1，λ2，λ3And a feature vector e1，e2，e3And further calculating to obtain the significance characteristics g of the SV1，g2，g3And curvature fsAs follows:
Figure FDA0002625184860000012
whether a hyper-voxel SV is classified as a planar hyper-voxel or a non-planar hyper-voxel is determined according to the following formula:
Figure FDA0002625184860000013
wherein T isIs a curvature threshold;
and 2.3, inputting the points classified into the non-planar hyper-voxels into the next layer of scale, and continuously repeating the step 2.1 and the step 2.2 until the scale is reduced to the minimum value or the number of the non-planar hyper-voxels is less than a certain specified number.
3. The three-dimensional point cloud plane segmentation method based on the opponent inference theory as claimed in claim 1, wherein: in step 3, the optimal three-dimensional plane subset extraction method is as follows:
step 3.1, calculating geometric consistency measure; for a three-dimensional point set S randomly comprising n points, a virtual hypothesis H is first defined0The method is used for indicating that no meaningful planar structure exists in S, and comprises the following specific steps:
step 3.1.1, for a particular candidate plane PcCalculating an arbitrary point p in SiAt PcGeometric consistency measure of (2):
Figure FDA0002625184860000021
wherein d (p)i，Pc) To represent a point piTo PcThe distance tolerance value tau is used to indicate that the point P has a higher probability of belonging to the planar model P when d (P, P) is less than tau;
step 3.1.2, denote S as belonging to an arbitrary subset of planes of S and satisfying
Figure FDA0002625184860000022
s is at PcThe geometric consistency measure of (a) is the maximum of the geometric consistency measures of the points within the s subset:
Figure FDA0002625184860000023
step 3.1.3, the global geometric consistency measure of S is defined as being at a certain PcObtained as above
Figure FDA0002625184860000024
Minimum value:
Figure FDA0002625184860000025
step 3.1.4, when there is a candidate plane model P derived from a subset of three pointscSuch that a subset s of k points satisfies:
Figure FDA0002625184860000026
then the geometric consistency of the subset s is said to reach the α level, or is said to satisfy the α -rigid condition, where α is a positive number; under the null hypothesis H0The geometric consistency of the subset s satisfying the α -rigid condition is as follows:
Figure FDA0002625184860000027
wherein
Figure FDA0002625184860000028
Is shown in the imaginary hypothesis H0Lower part
Figure FDA0002625184860000029
A probability measure of (d);
step 3.2, calculating the disorder measure; the measure of the disorder is obtained by computing the NFA values of a subset of planes, the imaginary hypothesis H defined above0Next, the NFA value for the subset s that satisfies α -rigid is defined as:
Figure FDA00026251848600000210
wherein N issRepresenting the total number of subsets S that may be present in S, the measure of disorder of S is obtained by calculating (α, n, k) for any subset S:
Figure FDA0002625184860000031
n represents the number of points in the point set S; k represents the number of points in the subset s;
step 3.3, obtaining an optimal plane subset; number of points contained in the optimal subset and corresponding thereto
Figure FDA0002625184860000032
Is defined as:
Figure FDA0002625184860000033
optimal plane subset
Figure FDA00026251848600000310
Corresponding plane model
Figure FDA0002625184860000034
And the number of points contained
Figure FDA0002625184860000035
Is defined as:
Figure FDA0002625184860000036
the optimal plane subset separated from the multi-scale plane superpixel is the initial strict plane superpixel.
4. The three-dimensional point cloud plane segmentation method based on the opponent inference theory as claimed in claim 1, wherein: in step 4, a strict plane hyper-voxel with the smallest distance between planes in the neighborhood range of each independent point is searched, and whether to combine the planes is determined according to a distance threshold in the following specific implementation manner:
step 4.1, for all points in a hyper-voxel of an initial strict plane, finding all neighborhood independent points in an independent point set through a k-d tree and a neighborhood search radius gamma, deleting repeated neighborhood points, and taking the rest points as a candidate point set omega; let an arbitrary point p in ΩΩPlanar model to initial strict planar hyper-voxels
Figure FDA0002625184860000037
Is expressed as
Figure FDA0002625184860000038
When the condition is satisfied
Figure FDA0002625184860000039
Then point p will be pointed outΩReclassification into corresponding strict plane hyper-voxels, TdIs a set distance threshold;
step 4.2, the remaining independent point set is taken as a new input point set and steps 1 to 4.1 are performed until no new strict plane hyper-voxels are generated.
5. The three-dimensional point cloud plane segmentation method based on the opponent inference theory as claimed in claim 1, wherein: the point cloud data in step 1 includes: 1) airborne LiDAR point clouds; 2) and obtaining aerial image dense matching point clouds through an aerial image and a dense matching algorithm.
CN202010794927.2A 2020-08-10 2020-08-10 Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory Active CN112070787B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010794927.2A CN112070787B (en) 2020-08-10 2020-08-10 Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010794927.2A CN112070787B (en) 2020-08-10 2020-08-10 Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory

Publications (2)

Publication Number Publication Date
CN112070787A true CN112070787A (en) 2020-12-11
CN112070787B CN112070787B (en) 2022-06-07

Family

ID=73660953

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010794927.2A Active CN112070787B (en) 2020-08-10 2020-08-10 Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory

Country Status (1)

Country Link
CN (1) CN112070787B (en)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112712596A (en) * 2021-03-29 2021-04-27 深圳大学 Dense matching point cloud building structured model fine reconstruction method
CN114463396A (en) * 2022-01-07 2022-05-10 武汉大学 Point cloud registration method using plane shape and topological graph voting

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1741071A (en) * 2005-09-19 2006-03-01 威盛电子股份有限公司 Three-dimensional space drawing processing method and apparatus
US20150178573A1 (en) * 2013-12-19 2015-06-25 Texas Instruments Incorporated Ground plane detection
CN106600622A (en) * 2016-12-06 2017-04-26 西安电子科技大学 Point cloud data partitioning method based on hyper voxels
CN107862738A (en) * 2017-11-28 2018-03-30 武汉大学 One kind carries out doors structure three-dimensional rebuilding method based on mobile laser measurement point cloud
CN109063663A (en) * 2018-08-10 2018-12-21 武汉大学 A kind of spissatus detection of timing remote sensing image and minimizing technology by slightly to essence
CN110110802A (en) * 2019-05-14 2019-08-09 南京林业大学 Airborne laser point cloud classification method based on high-order condition random field
CN110222642A (en) * 2019-06-06 2019-09-10 上海黑塞智能科技有限公司 A kind of planar architectural component point cloud contour extraction method based on global figure cluster
US20190310378A1 (en) * 2018-04-05 2019-10-10 Apex.AI, Inc. Efficient and scalable three-dimensional point cloud segmentation for navigation in autonomous vehicles
CN111275724A (en) * 2020-02-26 2020-06-12 武汉大学 Airborne point cloud roof plane segmentation method based on octree and boundary optimization

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN1741071A (en) * 2005-09-19 2006-03-01 威盛电子股份有限公司 Three-dimensional space drawing processing method and apparatus
US20150178573A1 (en) * 2013-12-19 2015-06-25 Texas Instruments Incorporated Ground plane detection
CN106600622A (en) * 2016-12-06 2017-04-26 西安电子科技大学 Point cloud data partitioning method based on hyper voxels
CN107862738A (en) * 2017-11-28 2018-03-30 武汉大学 One kind carries out doors structure three-dimensional rebuilding method based on mobile laser measurement point cloud
US20190310378A1 (en) * 2018-04-05 2019-10-10 Apex.AI, Inc. Efficient and scalable three-dimensional point cloud segmentation for navigation in autonomous vehicles
CN109063663A (en) * 2018-08-10 2018-12-21 武汉大学 A kind of spissatus detection of timing remote sensing image and minimizing technology by slightly to essence
CN110110802A (en) * 2019-05-14 2019-08-09 南京林业大学 Airborne laser point cloud classification method based on high-order condition random field
CN110222642A (en) * 2019-06-06 2019-09-10 上海黑塞智能科技有限公司 A kind of planar architectural component point cloud contour extraction method based on global figure cluster
CN111275724A (en) * 2020-02-26 2020-06-12 武汉大学 Airborne point cloud roof plane segmentation method based on octree and boundary optimization

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
ZHEN DONG ETAL.: "An efficient global energy optimization approach for robust 3D plane segmentation of point clouds", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》 *
ZHEN DONG ETAL.: "An efficient global energy optimization approach for robust 3D plane segmentation of point clouds", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》, vol. 2018, no. 137, 31 December 2018 (2018-12-31) *
姜媛媛: "基于超体素区域增长的点云分割算法研究", 《中国优秀博硕士学位论文全文数据库（硕士）》 *
姜媛媛: "基于超体素区域增长的点云分割算法研究", 《中国优秀博硕士学位论文全文数据库（硕士）》, vol. 2018, no. 4, 15 April 2018 (2018-04-15) *

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112712596A (en) * 2021-03-29 2021-04-27 深圳大学 Dense matching point cloud building structured model fine reconstruction method
CN114463396A (en) * 2022-01-07 2022-05-10 武汉大学 Point cloud registration method using plane shape and topological graph voting
CN114463396B (en) * 2022-01-07 2024-02-06 武汉大学 Point cloud registration method utilizing plane shape and topological graph voting

Also Published As

Publication number Publication date
CN112070787B (en) 2022-06-07

Similar Documents

Publication Publication Date Title
CN106780524B (en) Automatic extraction method for three-dimensional point cloud road boundary
CN103839261A (en) SAR image segmentation method based on decomposition evolution multi-objective optimization and FCM
Galvanin et al. Extraction of building roof contours from LiDAR data using a Markov-random-field-based approach
CN112070787B (en) Aviation three-dimensional point cloud plane segmentation method based on opponent reasoning theory
CN113628263A (en) Point cloud registration method based on local curvature and neighbor characteristics thereof
CN113484875B (en) Laser radar point cloud target hierarchical identification method based on mixed Gaussian ordering
CN111783722B (en) Lane line extraction method of laser point cloud and electronic equipment
Sun et al. Roads and intersections extraction from high-resolution remote sensing imagery based on tensor voting under big data environment
CN111462017A (en) Denoising method for tunnel laser point cloud data
CN115049925A (en) Method for extracting field ridge, electronic device and storage medium
CN110348478B (en) Method for extracting trees in outdoor point cloud scene based on shape classification and combination
Ruan et al. Review of 3d point cloud data segmentation methods
Wang et al. Roof plane segmentation from lidar point cloud data using region expansion based l 0 gradient minimization and graph cut
CN114463396A (en) Point cloud registration method using plane shape and topological graph voting
CN114332172A (en) Improved laser point cloud registration method based on covariance matrix
He et al. Progressive filtering of airborne LiDAR point clouds using graph cuts
CN109241628B (en) Three-dimensional CAD model segmentation method based on graph theory and clustering
Sun et al. Automated segmentation of LiDAR point clouds for building rooftop extraction
Omidalizarandi et al. Segmentation and classification of point clouds from dense aerial image matching
CN117253205A (en) Road surface point cloud rapid extraction method based on mobile measurement system
CN111986223B (en) Method for extracting trees in outdoor point cloud scene based on energy function
CN113077473A (en) Three-dimensional laser point cloud pavement segmentation method, system, computer equipment and medium
Santos et al. Classification of LiDAR data over building roofs using k-means and principal component analysis
CN115619977A (en) High-order dangerous rock monitoring method based on airborne laser radar
Huang et al. Semantic labeling and refinement of LiDAR point clouds using deep neural network in urban areas

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