CN104700398A - Point cloud scene object extracting method - Google Patents

Abstract

The invention discloses a point cloud scene object extracting method. The point cloud scene object extracting method specifically comprises the steps of 1 utilizing the distribution situation of point cloud data on a Gaussian sphere and utilizing a mean shift algorithm to conduct coarse segmentation on the point cloud data; 2 adopting a distance based point clustering method to conduct fine segmentation on the point cloud data subjected to coarse segmentation; 3 utilizing the nature and curvature information of the Gaussian sphere to conduct shape recognition on the point cloud data subjected to fine segmentation; 4 correcting the point cloud data subjected to shape recognition. The point cloud scene object extracting method can be applied to extraction of point cloud scenes including complex objects not just extraction of specific objects in the scenes or object extraction conducted on aviation LiDAR data including few objects.

Description

A kind of some cloud object scene extracting method

Technical field

The invention belongs to the cross discipline technical field that computer graphics and pattern-recognition combine, relate to a kind of some cloud object scene extracting method.

Background technology

The important research field that object always is computer graphics and pattern-recognition is extracted from a cloud scene.Mostly traditional extracting method is to extract for single object, is therefore difficult to the some cloud scene being applicable to comprise complex object.

Due to cloud data on a large scale, large scale and magnanimity, abundant object is included in scene, and characteristics of objects is different, exist and block between object when adding scanning scene, cause each object in scene can not obtain multi-faceted scanning, the cloud data causing object corresponding is imperfect.This makes from the extraction object of a cloud scene more difficult.

In in recent years, a lot of scholar concentrated on the research of research object extraction algorithm.According to the difference of input data type, existing method is divided into two classes: the object extraction based on image and the object extraction based on cloud data.

1. based on the object extraction of image

Image is the data type that a kind of ratio is easier to obtain, and has had a lot of method to take the image zooming-out object of image or Aerial photography from ground.

Yang determines the weights of laser scanning point by the spatial distribution characteristic (plan range, elevation difference and some dense degree etc.) of analysis site cloud, adopts distance weighted IDW reciprocal (Inverse DistanceWeighted) interpolating method to generate the characteristic image of Vehicle-borne Laser Scanning point cloud.Then, utilize the image processing means such as Iamge Segmentation, contours extract, the method combined with three dimensions geometric analysis, complete the extraction to buildings, trees in scene.Xiao proposes a kind of method extracting buildings from Aerial Images, and the method extracts buildings by the border and elevation information detecting Aerial Images, and wherein elevation information extracts from the matching technique of dense graph picture.King is different from the feature of other spatial object texture according to the buildings texture of remote sensing image, for improving image resolution, the remote sensing image object model approach proposing Gabor texture block is applied to the extraction of remote sensing image town buildings, with whole cities and towns for object, with different downtown areas such as buildings, road, greenery patchess for forming the texture block of object, set up the object model based on texture block, the texture utilizing model to carry out remote sensing image object is demarcated, and finally extracts town buildings.In proposing a kind of Color Remote Sensing Image buildings extracting method, this algorithm combines neutral collection and mean shift algorithm (Mean shift), mean shift algorithm segmentation is carried out to the image being transformed into neutral collection space, generate the spectrum class image being core with major surface features type in image, extract buildings.Lee, on the basis considering buildings spectral signature, establish parallel with the anisotropy neighborhood perpendicular to target structures owner direction, and the buildings adopting the sub-pixed mapping location model based on anisotropy Markov random field to carry out sub-pixed mapping yardstick extracts.

2. based on the object extraction of cloud data

Yao utilizes adaptive M ean shift algorithm to split airborne lidar cloud data, then extract the area of each segmentation bunch, planarity, upright position and vertical drop five features, utilize support vector machine to complete extraction to sports car in scene and stationary automobile.Boyko proposes a kind of method extracting road from the scape of scale non-structured three-dimensional point cloud city, first two-dimensional map is mapped to given three dimensional point cloud by the method, complete route map registration, then the segmentation completing a cloud about the same is produced according to the topological sum geometry of input map, then construct collection of illustrative plates, completed the extraction of road by the frontier point extracting road.Side, by analyzing space distribution and the statistical nature of laser point cloud on sweep trace, proposes a kind of road waypoint cloud extraction method being applicable to structured road environment.Xu Shouxian sets up gibbs free energy change model according to the geometric properties of target, the data item of this model is set up by compatibility of goals, the priori item of this model is set up by spatial characters such as the topological properties of target, then reversible jump markov Monte carlo algorithm is utilized to sample, and adopt simulated annealing to be optimized to solve, realize the automatic extraction of building target and tree crown target geometric object.Yang is according to after the globally optimal solution of acquisition Gibbs energy model, preliminarily can extract building target from airborne laser scanning data, then the geometric attribute integrate features algorithm of region growing of building target is utilized to reject ground point, tree crown point, the noise spot of error extraction, process of refinement is carried out to the building target tentatively extracted, finally from cloud data, has extracted each building target accurately.

Mostly these relevant methods are for object (such as buildings, road etc.) specific in scene at present, or for comprising the less aviation LiDAR data of object.In real City scenarios, often comprise the object of Various Complex, existing method is not also suitable for ground point cloud data.

Summary of the invention

The object of this invention is to provide a kind of some cloud object scene extracting method, solve prior art and only can carry out extracting for special object or from the problem comprising the less aviation LiDAR point cloud data of object and carry out object extraction.

The technical solution adopted in the present invention is, a kind of some cloud object scene extracting method, and concrete implementation step is:

Step 1, utilize the distribution situation of cloud data in Gaussian sphere, utilize mean shift algorithm to utilize mean shift algorithm to carry out coarse segmentation to cloud data;

Step 2, to carrying out the cloud data after coarse segmentation in step 1, adopting the some clustering method based on distance to carry out segmentation and cutting;

Step 3, the character of Gaussian sphere and curvature information is utilized to carry out shape recognition to segmenting the cloud data after cutting in step 2;

Step 4, the cloud data after carrying out shape recognition in step 3 to be revised.

Feature of the present invention is also,

In step 1, the concrete implementation step of coarse segmentation is:

Step 1.1, utilize the normal vector of principle component analysis calculation level cloud

Any point p in some cloud, finds k the neighbor point of a p the three rank covariance matrix M of some p are:

$M=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\Σ}}}(p_i-\stackrel{\‾}{p}){(p_i-\stackrel{\‾}{p})}^T---\left(1\right)$

Wherein, for the mean place of k the neighbor point of a p,

$\stackrel{\‾}{p}=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{p}_i---\left(2\right)$

Carry out Eigenvalues Decomposition by the three rank covariance matrix M of svd to a p, obtain the eigenvalue λ of covariance matrix M ₃> λ ₂> λ ₁> 0, the normal vector of some p is minimal eigenvalue λ ₁corresponding proper vector ${\stackrel{\→}{n}}_p=(n_x,n_y,n_z);$

Step 1.2, Gaussian mapping

According to the normal vector of step 1.1 mid point p the mapping position of some p in Gaussian sphere is obtained by formula (3):

$\left\{\begin{array}{c}\mathrm{ud}=\arcsin \left(n_y\right)\\ \mathrm{ld}=\arctan (n_x,n_z)\end{array}\right.---\left(3\right)$

To set up with the centre of sphere of the Gaussian sphere spherical coordinate system conversion formula that is coordinate origin as formula (4):

$\left\{\begin{array}{c}{x}_{\mathrm{norm}}=\cos \left(\mathrm{ud}\right)\·\cos \left(\mathrm{ld}\right)\\ y_{\mathrm{norm}}=\cos \left(\mathrm{ud}\right)\·\sin \left(\mathrm{ld}\right)\\ z_{\mathrm{norm}}=\sin \left(\mathrm{ud}\right)\end{array}\right.---\left(4\right)$

Owing to comprising multiple object in a cloud scene, after cloud data carries out Gaussian mapping, the point with identical normal vector can be mapped to same position in Gaussian sphere, splits the point in Gaussian sphere according to Mean-Shift method.

In step 1.1, any point p in cloud data, utilizes k-d to set and finds out the some q nearest with a p, if n _pn _q≈-1, reverses the normal vector direction of a p.

In step 1.2, the concrete implementation step of Mean-Shift method is:

1. select any point x in cloud data to be the center of circle, take h as radius, make a higher-dimension ball, and the institute that record drops in ball there is an x _i;

2. Meanshift vector m is calculated _h,Gx (), if m _h,G(x) < ε, now algorithm convergence is to the maximum point of data space Midst density, quits a program, otherwise, repeat step 1..

The scope of ε is 0-0.01.

Segmenting the concrete implementation step of cutting in step 2 is:

Step 2.1, for every bit p _i, utilize k-d to set and find out a p _ik neighbor point, screening with some p _idistance is less than the point set NN of threshold value r;

Cloud data after step 2.2, traversal coarse segmentation, as fruit dot p _ifirst point be traversed, some label=1 all in mark point set NN; As fruit dot p _inot first point be traversed, then whether the every bit traveled through in point set NN is labeled, if points all in point set NN is not all labeled, then label++, points all in point set NN is labeled as label, if there is the point be labeled in point set NN, find out mark value mLabel minimum in the point be labeled, points all in point set NN is labeled as mLabel;

Step 2.3, repetition step 2.1, step 2.2, until be a little all labeled in cloud data.

The concrete steps of the shape recognition of step 3 are:

Cloud data set after step 2 segmentation is cut is respectively P ₁, P ₂... P _m, the data set after segmenting the cloud data cut and carrying out Gaussian mapping is respectively G (P ₁), G (P ₂) ... G (P _m);

1. plane identification

First, formula (5) is utilized to calculate G (P _i) center

${\stackrel{\&OverBar;}{c}}_i=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}G\left(p_i\right)---\left(5\right)$

Wherein, p _ifor cloud data set P _iin any point, N is this cloud data set P _ithe number of point;

Secondly, formula (6) is utilized to calculate the variance var (G (P of the normal vector of every bit cloud data acquisition _i)):

$\mathrm{var}\left(G\left(P_i\right)\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(G\left(p_i\right)-{\stackrel{\&OverBar;}{c}}_i)}^2---\left(6\right)$

If variance var is (G (P _i)) be less than threshold epsilon, judging point cloud data acquisition P _ifor plane, if variance var is (G (P _i)) be greater than threshold epsilon, cloud data set P _iit is not plane;

2. cylinder identification

First, face of cylinder parametric equation is (ρ cosu, ρ sinu, v), and two curvature calculating cylinder are respectively:

|k ₁|＝1/ρ，

k ₂＝0，

Wherein, ρ is cylindrical radius,

Secondly, formula (7) is utilized to calculate every bit cloud data acquisition P _icurvature assemble center:

${\stackrel{\&OverBar;}{k}}_1=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i1}\right|---\left(7\right)$

Wherein, N is cloud data set P _iin | k _i2| the number that < is 1,

With for the center of circle, statistics and distance of center circle from the set of point being less than 1, with distance of center circle from being less than 1 count and account for cloud data set P _icount more than 70%, namely predicate cylinder;

3. spheroid identification

The parametric equation of spheroid is (ρ cosusinv, ρ sinucosv, ρ sinv), and two curvature calculating spheroid are:

|k ₁|＝|k ₂|＝1/ρ，

Wherein, ρ is radius of sphericity,

Formula (8), (9) are utilized to calculate the mean value of two curvature respectively:

${\stackrel{\&OverBar;}{k}}_1=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i1}\right|---\left(8\right)$

${\stackrel{\&OverBar;}{k}}_2=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i2}\right|---\left(9\right)$

Wherein, N is cloud data set P _ithe number of mid point,

With for the center of circle, statistics and distance of center circle from the set of point being less than 1, with distance of center circle from being less than 1 count and account for cloud data set P _icount more than 70%, namely predicate spheroid.

Cloud data carried out to segmentation and after shape recognition through step 1-3, a basic configuration can be divided into multiple, for avoiding dough sheet in small, broken bits and over-segmentation, the cloud data after segmentation also shape recognition being revised, in step 4 to the concrete steps that cloud data is revised is:

1. to plane correction

If two cluster data S _iand S _jtype is all plane, s respectively _i, S _jthe method of average vector, meet two cluster data normal vector directions identical simultaneously for a bundle of planes S _i, bundle of planes S _iinterior every bit looks for its k neighbor point, if wherein include bundle of planes S _iand S _jpoint, by two bundle of planes S _iand S _jmerge;

2. to cylinder correction

If two cluster data types are all cylinder, utilize k-d to set k the neighbor point finding out every bit in a cylinder bunch, judge whether to include the point of other cylinder bunch, if had, two bunches are merged;

3. to spheroid correction

If two cluster data types are all spheroid, find out k neighbor point of every bit in a cylinder bunch with k-d tree, judge whether to include the point of other spheroid bunch, if had, two bunches are merged.

The invention has the beneficial effects as follows: some cloud object scene extracting method of the present invention goes for extracting the some cloud scene comprising complex object, and not only for object specific in scene or carry out object extraction for the aviation LiDAR data comprising object less.

Accompanying drawing explanation

Fig. 1 is some cloud object scene extracting method point cloud contextual data figure of the present invention;

Fig. 2 is some cloud object scene extracting method point cloud contextual data coarse segmentation result figure of the present invention;

Fig. 3 is the schematic diagram based on the some clustering method of distance in some cloud object scene extracting method of the present invention;

Fig. 4 is cylindrical curvature mapping graph in some cloud object scene extracting method of the present invention;

Fig. 5 is that in some cloud object scene extracting method of the present invention, basic configuration extracts schematic diagram;

Fig. 6 is not substantially shape extracting schematic diagram in some cloud object scene extracting method of the present invention.

Embodiment

Below in conjunction with the drawings and specific embodiments, the present invention is described in detail.

As shown in Figure 1, the present invention's one puts cloud object scene extracting method to some cloud contextual data of the present invention, and concrete implementation step is:

Step 1, utilize the distribution situation of cloud data in Gaussian sphere, utilize mean shift algorithm to carry out coarse segmentation to cloud data, concrete implementation step is:

Step 1.1, utilize the normal vector of principle component analysis calculation level cloud

Any point p in some cloud, finds k the neighbor point of a p the three rank covariance matrix M of some p are:

$M=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}(p_i-\stackrel{\&OverBar;}{p}){(p_i-\stackrel{\&OverBar;}{p})}^T---\left(1\right)$

Wherein, for the mean place of k the neighbor point of a p,

$\stackrel{\&OverBar;}{p}=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{p}_i---\left(2\right)$

Carry out Eigenvalues Decomposition by the three rank covariance matrix M of svd to a p, obtain the eigenvalue λ of covariance matrix M ₃> λ ₂> λ ₁> 0, the normal vector of some p is minimal eigenvalue λ ₁corresponding proper vector because there is ambiguity in the direction of normal vector, so for any point p in cloud data, utilize k-d to set and find out the some q nearest with a p, if n _pn _q≈-1, then reverse the normal vector direction of a p.

Step 1.2, Gaussian mapping

According to the normal vector of step 1.1 mid point p the mapping position of some p in Gaussian sphere is obtained by formula (3):

$\left\{\begin{array}{c}\mathrm{ud}=\arcsin \left(n_y\right)\\ \mathrm{ld}=\arctan (n_x,n_z)\end{array}\right.---\left(3\right)$

To set up with the centre of sphere of the Gaussian sphere spherical coordinate system conversion formula that is coordinate origin as formula (4):

$\left\{\begin{array}{c}{x}_{\mathrm{norm}}=\cos \left(\mathrm{ud}\right)\&CenterDot;\cos \left(\mathrm{ld}\right)\\ y_{\mathrm{norm}}=\cos \left(\mathrm{ud}\right)\&CenterDot;\sin \left(\mathrm{ld}\right)\\ z_{\mathrm{norm}}=\sin \left(\mathrm{ud}\right)\end{array}\right.---\left(4\right)$

Owing to comprising multiple object in a cloud scene, after cloud data carries out Gaussian mapping, the point with identical normal vector can be mapped to same position in Gaussian sphere, splits the point in Gaussian sphere according to Mean-Shift method.

Mean-Shift method essence is a gradient ascent algorithm, and it is variable step, also can be called self-adaption gradient ascent algorithm, and its concrete implementation step is:

1. select any point x in cloud data to be the center of circle, take h as radius, make a higher-dimension ball, and the institute that record drops in ball there is an x _i;

2. Meanshift vector m is calculated _h,Gx (), if m _h,G(x) < ε (scope of ε is 0-0.01), now algorithm convergence is to the maximum point of data space Midst density, quits a program, otherwise, repeat step 1..

According to the character of Gaussian mapping, for the point that normal vector direction is identical, Gaussian sphere all maps and becomes same point.So the curved surface of identical type can there will be identical position after Gaussian mapping in Gaussian sphere, namely overlapping.This can cause the curved surface of the identical type in normal vector direction, utilizes Mean-shift clustering algorithm can be divided into identical bunch.

Result figure after coarse segmentation as shown in Figure 2.

Step 2, adopt the some clustering method based on distance to carry out segmentation to cut

Can find, between the curved surface of multiple overlap, all there is certain distance according to the observation, so the present invention adopts the some cloud clustering method (distance-based clustering) based on distance, by the curved surface of these identical types separately.Based on the main thought of the some cloud clustering method of distance, be that some distance being less than r is classified as a class.For a p _i, p _i∈ T _i, first, utilize k-d to set and find out a p _ik (k generalized case gets 30) individually closely to face a little, with p _ithe point that distance is less than r is classified as a class, i.e. P _r-distace={ { p _i, p _j, d _ij| d _ij≤ r, i ≠ j}, wherein, d _ijfor a p _iand p _jdistance as shown in Figure 3.

Segmenting the concrete implementation step of cutting is:

Step 2.1, for every bit p _i, utilize k-d to set and find out a p _ik neighbor point, screening with some p _idistance is less than the point set NN of threshold value r;

Cloud data after step 2.2, traversal coarse segmentation, as fruit dot p _ifirst point be traversed, some label=1 all in mark point set NN; As fruit dot p _inot first point be traversed, then whether the every bit traveled through in point set NN is labeled, if points all in point set NN is not all labeled, then label++, points all in point set NN is labeled as label, if there is the point be labeled in point set NN, find out mark value mLabel minimum in the point be labeled, points all in point set NN is labeled as mLabel;

Step 2.3, repetition step 2.1, step 2.2, until be a little all labeled in cloud data.

Step 3, the character utilizing Gaussian sphere and curvature information carry out shape recognition

Cloud data set after step 2 segmentation is cut is respectively P ₁, P ₂... P _m, the data set after segmenting the cloud data cut and carrying out Gaussian mapping is respectively G (P ₁), G (P ₂) ... G (P _m).Every cluster cloud data P will be judged respectively _itype, to realize the extraction to single body.

1. plane identification

Because the point of plane cloud data after Gaussian mapping is not that strict mapping becomes a point, they can be gathered in a little region.So cluster cloud data P will be judged _iwhether be plane, first judge the data set G (P after their Gaussian mapping _i) whether concentrate in a little region.

First, formula (5) is utilized to calculate G (P _i) center

${\stackrel{\&OverBar;}{c}}_i=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}G\left(p_i\right)---\left(5\right)$

Wherein, p _ifor cloud data set P _iin any point, N is this cloud data set P _ithe number of point;

Secondly, formula (6) is utilized to calculate the variance var (G (P of the normal vector of every bit cloud data acquisition _i)):

$\mathrm{var}\left(G\left(P_i\right)\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(G\left(p_i\right)-{\stackrel{\&OverBar;}{c}}_i)}^2---\left(6\right)$

If variance var is (G (P _i)) be less than threshold epsilon (ε scope is 0-0.15), then illustrate that the normal vector of this bunch is distributed in a less region, can judging point cloud data acquisition P _ifor plane.On the contrary, if variance is greater than threshold epsilon, illustrates that dispersion is compared in the normal vector distribution of this bunch, do not meet the requirement of plane.

2. cylinder identification

Because cloud data amount is more, comprise many noises, so in actual computation process, curvature mapping point can coincide with a region in coordinate axis simultaneously.So cluster cloud data P will be judged _iwhether be cylinder, first judge the data set C (P after this bunch of curvature mapping _i) whether concentrate in the little region of coordinate axis one.

First, face of cylinder parametric equation is (ρ cosu, ρ sinu, v), and two curvature calculating cylinder are respectively:

|k ₁|＝1/ρ，

k ₂＝0，

Wherein, ρ is cylindrical radius, which dictates that its curvature mapping point coincide with in coordinate axis positive dirction a bit, as shown in Figure 4,

Secondly, formula (7) is utilized to calculate every bit cloud data acquisition P _icurvature assemble center:

${\stackrel{\&OverBar;}{k}}_1=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i1}\right|---\left(7\right)$

Wherein, N is cloud data set P _iin | k _i2| the number that < is 1,

Due to noise or the reason such as curvature estimation is inaccurate, curvature maps and does not all concentrate on a little region, has a lot of scattered points to be distributed in the positive dirction of x-axis.So utilize formula (7) to add up | k ₂| the average coordinates of the point of ≈ 0, so that the center that correct determination curvature is assembled.A region in coordinate axis is coincided with due to curvature mapping point, so with for the center of circle, statistics and distance of center circle from the set of point being less than 1, with distance of center circle from being less than 1 count and account for cloud data set P _icount more than 70%, namely predicate cylinder;

3. spheroid identification

The parametric equation of spheroid is (ρ cosusinv, ρ sinucosv, ρ sinv), and two curvature calculating spheroid are:

|k ₁|＝|k ₂|＝1/ρ，

Wherein, ρ is radius of sphericity,

Formula (8), (9) are utilized to calculate the mean value of two curvature respectively:

${\stackrel{\&OverBar;}{k}}_1=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i1}\right|---\left(8\right)$

${\stackrel{\&OverBar;}{k}}_2=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i2}\right|---\left(9\right)$

Wherein, N is cloud data set P _ithe number of mid point,

With for the center of circle, statistics and distance of center circle from the set of point being less than 1, with distance of center circle from being less than 1 count and account for cloud data set P _icount more than 70%, namely predicate spheroid;

Be illustrated in figure 5 basic configuration and extract schematic diagram.

4. not substantially shape point cloud data identification

The differential character such as the character of Gaussian sphere, curvature are utilized to identify the basic configuration in a cloud scene.In addition, also comprise many objects that can not be made up of basic configuration, utilize exclusive method in scene, remove basic configuration, remaining cloud data is not substantially shape cloud data.

The cloud data after basic configuration is removed, comprising objects such as noise, trees, short shrubs in some cloud scene.For the cloud data of not substantially shape, utilize the some clustering method based on distance to split, trees etc. can not be become single body by the Object Segmentation that basic configuration represents.

Be illustrated in figure 6 not substantially shape extracting schematic diagram.

Step 4, correction

Cloud data carried out to segmentation and after shape recognition through step 1-3, a basic configuration can be divided into multiple, for avoiding dough sheet in small, broken bits and over-segmentation, the cloud data after segmentation also shape recognition being revised, in step 4 to the concrete steps that cloud data is revised is:

1. to plane correction

If two cluster data S _iand S _jtype is all plane, s respectively _i, S _jthe method of average vector, meet two cluster data normal vector directions identical simultaneously for a bundle of planes S _i, bundle of planes S _iinterior every bit looks for its k neighbor point, if wherein include bundle of planes S _iand S _jpoint, by two bundle of planes S _iand S _jmerge;

2. to cylinder correction

If two cluster data types are all cylinder, utilize k-d to set k the neighbor point finding out every bit in a cylinder bunch, judge whether to include the point of other cylinder bunch, if had, two bunches are merged;

3. to spheroid correction

If two cluster data types are all spheroid, find out k neighbor point of every bit in a cylinder bunch with k-d tree, judge whether to include the point of other spheroid bunch, if had, two bunches are merged.

Point cloud object scene extracting method of the present invention goes for extracting the some cloud scene comprising complex object, and not only for object specific in scene or carry out object extraction for the aviation LiDAR data comprising object less.

Claims (8)

1. a some cloud object scene extracting method, it is characterized in that, concrete implementation step is:

Step 1, utilize the distribution situation of cloud data in Gaussian sphere, utilize mean shift algorithm to carry out coarse segmentation to cloud data;

Step 2, to carrying out the cloud data after coarse segmentation in step 1, adopting the some clustering method based on distance to carry out segmentation and cutting;

Step 3, the character of Gaussian sphere and curvature information is utilized to carry out shape recognition to segmenting the cloud data after cutting in step 2;

Step 4, the cloud data after carrying out shape recognition in step 3 to be revised.

2. one point cloud object scene extracting method according to claim 1, it is characterized in that, in described step 1, the concrete implementation step of coarse segmentation is:

Step 1.1, utilize the normal vector of principle component analysis calculation level cloud

Any point p in some cloud, finds k the neighbor point of a p the three rank covariance matrix M of some p are:

$M=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}(p_i-\stackrel{\&OverBar;}{p}){(p_i-\stackrel{\&OverBar;}{p})}^T,$

Wherein, for the mean place of k the neighbor point of a p,

$\stackrel{\&OverBar;}{p}=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{p}_i,$

Carry out Eigenvalues Decomposition by the three rank covariance matrix M of svd to a p, obtain the eigenvalue λ of covariance matrix M ₃> λ ₂> λ ₁> 0, the normal vector of some p is minimal eigenvalue λ ₁corresponding proper vector ${\stackrel{\&RightArrow;}{n}}_p=(n_x,n_y,n_z);$

Step 1.2, Gaussian mapping

According to the normal vector of step 1.1 mid point p the mapping position of some p in Gaussian sphere is obtained by following formula:

$\left\{\begin{array}{c}\mathrm{ud}=\arcsin \left(n_y\right)\\ \mathrm{ld}=\arctan (n_x,n_z)\end{array}\right.,$

The spherical coordinate system conversion formula that foundation is coordinate origin with the centre of sphere of Gaussian sphere is:

$\left\{\begin{array}{c}{x}_{\mathrm{norm}}=\cos \left(\mathrm{ud}\right)\&CenterDot;\cos \left(\mathrm{ld}\right)\\ y_{\mathrm{norm}}=\cos \left(\mathrm{ud}\right)\&CenterDot;\sin \left(\mathrm{ld}\right)\\ z_{\mathrm{norm}}=\sin \left(\mathrm{ud}\right)\end{array}\right.,$

Owing to comprising multiple object in a cloud scene, after cloud data carries out Gaussian mapping, the point with identical normal vector can be mapped to same position in Gaussian sphere, splits the point in Gaussian sphere according to Mean-Shift method.

3. one point cloud object scene extracting method according to claim 2, it is characterized in that, in described step 1.1, any point p in cloud data, utilizes k-d to set and finds out the some q nearest with a p, if:

n _p·n _q≈-1，

Is reversed in the normal vector direction of a p.

4. one point cloud object scene extracting method according to claim 2, it is characterized in that, in described step 1.2, the concrete implementation step of Mean-Shift method is:

1. select any point x in cloud data to be the center of circle, take h as radius, make a higher-dimension ball, and the institute that record drops in ball there is an xi;

2. Meanshift vector m is calculated _h,Gx (), if m _h,G(x) < ε, now algorithm convergence is to the maximum point of data space Midst density, quits a program, otherwise, repeat step 1..

5. one point cloud object scene extracting method according to claim 4, it is characterized in that, the scope of described ε is 0-0.01.

6. one point cloud object scene extracting method according to claim 1, it is characterized in that, segmenting the concrete implementation step of cutting in described step 2 is:

Step 2.1, for every bit p _i, utilize k-d to set and find out a p _ik neighbor point, screening with some p _idistance is less than the point set NN of threshold value r;

Cloud data after step 2.2, traversal coarse segmentation, as fruit dot p _ifirst point be traversed, some label=1 all in mark point set NN; As fruit dot p _inot first point be traversed, then whether the every bit traveled through in point set NN is labeled, if points all in point set NN is not all labeled, then label++, points all in point set NN is labeled as label, if there is the point be labeled in point set NN, find out mark value mLabel minimum in the point be labeled, points all in point set NN is labeled as mLabel;

Step 2.3, repetition step 2.1, step 2.2, until be a little all labeled in cloud data.

7. one point cloud object scene extracting method according to claim 1, it is characterized in that, the concrete steps of the shape recognition of described step 3 are:

Cloud data set after step 2 segmentation is cut is respectively P ₁, P ₂... P _m, the data set after segmenting the cloud data cut and carrying out Gaussian mapping is respectively G (P ₁), G (P ₂) ... G (P _m);

1. plane identification

First, following formula is utilized to calculate G (P _i) center

${\stackrel{\&OverBar;}{c}}_i=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}G\left(p_i\right),$

Wherein, p _ifor cloud data set P _iin any point, N is this cloud data set P _ithe number of point;

Secondly, following formula is utilized to calculate the variance var (G (P of the normal vector of every bit cloud data acquisition _i)):

$\mathrm{var}\left(G\left(P_i\right)\right)=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}{(G\left(p_i\right)-{\stackrel{\&OverBar;}{c}}_i)}^2,$

If variance var is (G (P _i)) be less than threshold epsilon, judging point cloud data acquisition P _ifor plane, if variance var is (G (P _i)) be greater than threshold epsilon, cloud data set P _iit is not plane;

2. cylinder identification

First, face of cylinder parametric equation is (ρ cosu, ρ sinu, v), and two curvature calculating cylinder are respectively:

|k ₁|＝1/ρ，

k ₂＝0，

Wherein, ρ is cylindrical radius,

Secondly, following formula is utilized to calculate every bit cloud data acquisition P _icurvature assemble center:

${\stackrel{\&OverBar;}{k}}_1=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i1}\right|,$

Wherein, N is cloud data set P _iin | k _i2| the number that < is 1,

With for the center of circle, statistics and distance of center circle from the set of point being less than 1, with distance of center circle from being less than 1 count and account for cloud data set P _icount more than 70%, namely predicate cylinder;

3. spheroid identification

The parametric equation of spheroid is (ρ cosusinv, ρ sinucosv, ρ sinv), and two curvature calculating spheroid are:

|k ₁|＝|k ₂|＝1/ρ，

Wherein, ρ is radius of sphericity,

Following two formula are utilized to calculate the mean value of two curvature respectively:

${\stackrel{\&OverBar;}{k}}_1=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i1}\right|,$

${\stackrel{\&OverBar;}{k}}_2=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\&Sigma;}}}\left|k_{i2}\right|,$

Wherein, N is cloud data set P _ithe number of mid point,

With for the center of circle, statistics and distance of center circle from the set of point being less than 1, with distance of center circle from being less than 1 count and account for cloud data set P _icount more than 70%, namely predicate spheroid.

8. one point cloud object scene extracting method according to claim 1, it is characterized in that, cloud data carried out to segmentation and after shape recognition through described step 1-3, a basic configuration can be divided into multiple, for avoiding dough sheet in small, broken bits and over-segmentation, cloud data after segmentation also shape recognition is revised, in described step 4 to the concrete steps that cloud data is revised is:

1. to plane correction

If two cluster data S _iand S _jtype is all plane, s respectively _i, S _jthe method of average vector, meet two cluster data normal vector directions identical simultaneously for a bundle of planes S _i, bundle of planes S _iinterior every bit looks for its k neighbor point, if wherein include bundle of planes S _iand S _jpoint, by two bundle of planes S _iand S _jmerge;

2. to cylinder correction

If two cluster data types are all cylinder, utilize k-d to set k the neighbor point finding out every bit in a cylinder bunch, judge whether to include the point of other cylinder bunch, if had, two bunches are merged;

3. to spheroid correction

If two cluster data types are all spheroid, find out k neighbor point of every bit in a cylinder bunch with k-d tree, judge whether to include the point of other spheroid bunch, if had, two bunches are merged.