CN105354829A - Self-adaptive point cloud data segmenting method - Google Patents

Abstract

The present invention discloses a self-adaptive point cloud data segmenting method. The method overcomes limitations of a regional growing method to select a seed point randomly. A local distance and a close distance of each data point are obtained by calculation. The data points are sorted to obtain a sorted set of the seed points. In a regional growing process, the seed point can be selected effectively to avoid too many parameter adjustment operations, thereby facilitating point cloud segmentation and improving a processing rate. According to a comparative experiment, it is found that the method provided by the present invention has a better effect on the point cloud segmentation.

Description

A kind of adaptive point cloud data segmentation method

Technical field

The invention belongs to computer graphics techniques field, relate to a kind of adaptive point cloud data segmentation method.

Background technology

Along with the development of computer graphical, image technique, virtual visualization technology has more wide application prospect at multiple fields such as ecology, agronomy and forestry.Along with the improvement and popularization of Laser Scanning Equipment and technology, allow to carry out reconstruction of three-dimensional model comparatively accurately based on the method for reconstructing of three-dimensional data model and structure measuring method, the model that reconstructs is simultaneously also realistic strong and precision is high and the advantage such as the characteristic information of good recovery crops, is subject to the increasing researcher's attention of industry.But the difference of same model structure can cause the curve of reconstruct, curved surface rough, so need to adopt different modeling methods for the different parts of model.

Traditional some cloud dividing method is for structurized point cloud model or 2.5D depth image mostly, and for inorganization at random some cloud, not there is adaptability, a lot of restriction can be run into time above-mentioned traditional some cloud dividing method promotes the use of in destructuring three-dimensional point cloud simultaneously, non-structured some cloud at random is difficult to reach good segmentation effect.From current both at home and abroad Study and appliance, although to a cloud segmentation carried out a large amount of, towards the research of different application problem, also do not have a kind of partitioning algorithm being applicable to all application, most algorithm all proposes for particular problem.

Summary of the invention

For above-mentioned problems of the prior art and defect, the object of the invention is to, a kind of adaptive point cloud data segmentation method is provided.

To achieve these goals, the present invention adopts following technical scheme:

A kind of adaptive point cloud data segmentation method, specifically comprises the following steps:

Step 1: obtain cloud data, carry out feature extraction to cloud data, obtains the local density and closely of each data point;

Step 2: the local density of each data point obtained according to step 1 and closely, sorts to each data point, obtains the Seed Points set after sorting;

Step 3: calculate normal vector corresponding to each data point and curvature;

Step 4: the normal vector that each data point obtained according to step 3 is corresponding and curvature, utilize region growing algorithm, data point in the Seed Points set obtained according to step 2 carries out automatically choosing of Seed Points, classifies to all data points, realizes the segmentation of cloud data.

Further, described step 1 the implementation procedure that cloud data carries out feature extraction is comprised:

Step 1.1: calculate data point x _i, the distance in 1≤i≤N between any two data points;

Step 1.2: the distance between any two data points step 1.1 obtained, generating new set d according to sorting from small to large, calculating and blocking distance d _c;

Step 1.3: what obtain according to step 1.2 blocks distance d _c, calculate each data point x _ilocal density ρ _i;

Step 1.4: the local density ρ that step 1.3 is obtained _icarry out descending sort, regenerate a subscript sequence;

Step 1.5: according to the distance between any two data points that step 1.1 obtains, calculate each data point closely q _irepresent the subscript numerical value of the subscript sequence described in step 1.4.

Further, the implementation procedure of described step 2 comprises:

Step 2.1: the local density ρ of each data point utilizing step 1 to obtain _iclosely obtain the judgment value PU of each data point _i;

Step 2.2: by data point x _iaccording to PU _ivalue carry out descending sort, generate Seed Points S set EEDS.

Further, the implementation procedure of described step 3 comprises:

Step 3.1: for each data point x _i, utilize the Neighborhood-region-search algorithm based on FDN to obtain each data point x _ineighbor point;

Step 3.2: calculate each data point x _ithe three-dimensional barycenter of all neighbor points;

Step 3.3: each data point x of the centroid calculation utilizing step 3.2 to obtain _icorresponding covariance matrix Cov _i;

Step 3.4: according to the covariance matrix Cov obtained of step 3.3 _icalculate each data point x _icharacteristic of correspondence value and proper vector, obtain data point x _icorresponding normal vector

Step 3.5: each the data point x obtained according to step 3.4 _icharacteristic of correspondence value, calculates each data point x _icurvature Cur _i.

Further, the implementation procedure of described step 4 comprises:

Step 4.1: note set Nbhd is data neighborhood of a point, and the element in set Nbhd is the neighbor point of data point, and set Rc is current region, and initial value is empty, S _cfor current seed region, initial value is empty; Select first data point seed in Seed Points S set EEDS ₁as initial current seed point, be inserted into current seed region S _cin;

Step 4.2: utilize FDN Neighborhood-region-search algorithm to obtain the neighborhood Nbhd of current seed point, to each neighbor point in neighborhood Nbhd l represents l neighbor point of current seed point, judges its normal vector with the normal vector of current seed point between angle whether meet the normal vector threshold condition of setting, if meet, then by this neighborhood point join in current region Rc;

Step 4.3: judge current neighborhood point curvature whether meet the curvature threshold condition of setting, if meet, then by neighborhood point join current seed region S _cin;

Step 4.4: delete current seed region S _cin current seed point, and by this data point from set U delete;

Step 4.5: choose current seed region S _cin next data point as current seed point, return step 4.2, until seed region S _cfor sky, perform step 4.6;

Step 4.6: if set U be empty, then the data point in current region Rc preserved hereof, current region Rc is set to sky again, chooses next data point in SEEDS as current seed point, returns step 4.1, until set U be empty, split end; Data point in then different current region Rc is stored in different files respectively, realizes the classification of data point, namely realizes the segmentation of cloud data.

Compared with prior art, the present invention has following technique effect:

1, the present invention is by calculating the local distance and closely of each data point, data point is sorted, obtain the Seed Points set after sorting, in the propagation process of region, according to the order of the data point in Seed Points set, automatically choosing of Seed Points can be carried out, avoid too much parameter adjustment operation, for a cloud segmentation provides convenient, improve processing speed; Found by contrast experiment, the method in the present invention has some cloud segmentation effect preferably.

2, the process of choosing of Seed Points of the present invention has adaptivity, and to insensitive for noise, can get rid of to a great extent and choose noise as this situation of Seed Points.

Accompanying drawing explanation

Fig. 1 is method flow diagram of the present invention;

Fig. 2 is the three-dimensional point cloud model segmentation effect figure of milpa; Milpa original point cloud model shown in Fig. 2 (a), Fig. 2 (b) is the some cloud design sketch after adopting the inventive method segmentation, Fig. 2 (c) is the some cloud design sketch after adopting region growing algorithm segmentation, and Fig. 2 (d) is the some cloud design sketch after adopting DBSCAN segmentation;

Fig. 3 is the three-dimensional point cloud model segmentation effect figure of tree; Fig. 3 (a) represents the original point cloud data model of tree, Fig. 3 (b) is the some cloud design sketch after adopting the inventive method segmentation, point cloud design sketch after Fig. 3 (c) adopts region growing algorithm to split, Fig. 3 (d) is the some cloud design sketch after adopting DBSCAN segmentation.

Below in conjunction with the drawings and specific embodiments the solution of the present invention done and explain in further detail and illustrate.

Embodiment

Defer to technique scheme, see Fig. 1, self-adaptation point cloud data segmentation method of the present invention, specifically comprises the following steps:

Step 1: utilize spatial digitizer to scan a cloud, obtains cloud data, cloud data is carried out to the extraction of cloud data feature, and obtain each data point local density and closely, implementation procedure comprises:

Step 1.1: the set of note cloud data is U={x _i| 1≤i≤N}, N is total number of data point in set U, by data point x _i, import with x, y, z be coordinate axis three-dimensional system of coordinate in, according to import data point x _i, calculate data point x _iin distance between any two data points, obtain N* (N-1)/2 distance value;

$dist(i,j)=\sqrt{{(x_i-x_j)}^2+{(y_i-y_j)}^2+{(z_i-z_j)}^2}---\left(1\right)$

Wherein, (x _i, y _i, z _i) and (x _j, y _j, z _j) represent data point x respectively _imid point p _iand p _jcoordinate, and j ≠ i, wherein 1≤i≤N, 1≤j≤N;

Step 1.2: the distance between any two data points step 1.1 obtained, generating new set d according to sorting from small to large, calculating and blocking distance d _c:

d _c＝d _{σ*(N*(N-1)/2)}(2)

Wherein, the value of σ is between 1%-2%, different for different model values.

Step 1.3: what obtain according to step 1.2 blocks distance d _c, calculate each data point x _ilocal density ρ _i:

${\mathrm{\ρ}}_i=\underset{j=1}{\overset{N}{\Σ}}X\left(dist\right(i,j)-d_c)---\left(3\right)$

Wherein, $X\left(x\right)=\left\{\begin{array}{c}1,x<0\\ 0,x\&GreaterEqual;0\end{array}\right..$

Step 1.4: the close together of two Seed Points chosen in region growing algorithm, causes the generation splitting excessive situation, by local density ρ _icarry out descending sort, regenerate a subscript sequence, described subscript sequence meets the condition described in formula (4):

${\mathrm{\&rho;}}_{Q_1}\&GreaterEqual;{\mathrm{\&rho;}}_{Q_2}\&GreaterEqual;{\mathrm{\&rho;}}_{Q_3}......\&GreaterEqual;{\mathrm{\&rho;}}_{Q_N}---\left(4\right)$

Wherein, Q _irepresent the subscript of the subscript sequence regenerated, and 1≤i≤N.

Step 1.5: according to the distance between any two data points that step 1.1 obtains, calculate each data point closely

${\mathrm{\&delta;}}_{Q_i}=\left\{\begin{array}{c}{\min }_{j<i}\left(dist\right(Q_i,Q_j\left)\right),i\&GreaterEqual;2\\ {\max }_{j\&GreaterEqual;2}\left(dist\right(Q_i,Q_j\left)\right),i=1\end{array}\right.---\left(5\right)$

Wherein, Q _i, Q _jbe the subscript numerical value of the subscript sequence in formula (4), the described distance closely referred between two Seed Points.

Step 2: according to the local density ρ of each data point that step 1 obtains _iclosely to data point x _isort, obtain the Seed Points S set EEDS after sorting, it specifically comprises the following steps:

Step 2.1: utilize local density ρ _iclosely obtain each data point x _ijudgment value PU _i:

PU _i＝ρ _i*δ _i(6)

Step 2.2: by data point x _iaccording to PU _ivalue carry out descending sort, generate Seed Points S set EEDS, the Seed Points after the data point in set is and sequences sequence;

SEEDS(s)＝{seed ₁，seed ₂，。。。。。。，seed _N}(7)

Step 3: calculate each data point x _inormal vector and curvature

Step 3.1: for each data point x _i, utilize the searching algorithm based on FDN (fixed range neighborhood search) ^[1]obtain each data point x _ineighbor point 1≤l≤k _iif, data point x _ithe number of neighbor point be k _i;

Step 3.2: calculate each data point x _ik _iindividual neighbor point three-dimensional barycenter

${\stackrel{\&RightArrow;}{X}}_{ic}=\frac{1}{k_i}\underset{l=1}{\overset{k_i}{\&Sigma;}}{\stackrel{\&RightArrow;}{x}}_l---\left(8\right)$

Step 3.3: utilize the barycenter obtained calculate data point x _icorresponding covariance matrix Cov _i:

$C=\frac{1}{k_i}\underset{l=1}{\overset{k_i}{\&Sigma;}}({\stackrel{\&RightArrow;}{x}}_l-{\stackrel{\&RightArrow;}{X}}_{ic}){({\stackrel{\&RightArrow;}{x}}_l-{\stackrel{\&RightArrow;}{X}}_{ic})}^T---\left(9\right)$

Step 3.4: according to the Cov obtained _icalculate data point x _icharacteristic of correspondence value and proper vector, obtain data point x _icorresponding normal vector:

${\mathrm{Cov}}_i*\stackrel{\&RightArrow;}{V_{im}}=\stackrel{\&RightArrow;}{{\mathrm{\&lambda;}}_{im}}*\stackrel{\&RightArrow;}{V_{im}}---\left(10\right)$

Wherein, represent data point x _icorresponding m (1≤m≤3) individual eigenwert, representation feature value characteristic of correspondence vector.

Theoretical according to principal element, data point x _icorresponding minimum eigenwert characteristic of correspondence vector be data point x _icorresponding normal vector.

Step 3.5: each the data point x obtained according to step 3.4 _icharacteristic of correspondence value, calculates each data point x _icurvature Cur _i:

${\mathrm{Cur}}_i=\frac{{\mathrm{\&lambda;}}_{i1}}{{\mathrm{\&lambda;}}_{i1}+{\mathrm{\&lambda;}}_{i2}+{\mathrm{\&lambda;}}_{i3}}---\left(11\right)$

Step 4: obtain each data point x according to step 3 _inormal vector with curvature Cur _i, utilize region growing algorithm, the data point in the Seed Points set obtained according to step 2 carries out automatically choosing of Seed Points, carries out cluster to all data points, realizes the segmentation of cloud data.Specific implementation step comprises:

Step 4.1: note set Nbhd is data neighborhood of a point, and the element in set Nbhd is the neighbor point of data point, and set Rc is current region, and initial value is empty, S _cfor current seed region, initial value is empty; Select first data point seed in Seed Points S set EEDS ₁as initial current seed point, be inserted into current seed region S _cin;

Step 4.2: utilize FDN (fixed range neighbour) searching algorithm to obtain the neighborhood Nbhd of current seed point, to each neighbor point in neighborhood Nbhd l represents l neighbor point of current seed point, utilizes formula (12) to judge the normal vector of each neighborhood point with the normal vector of current seed point between angle whether meet normal vector threshold condition:

$cos\left(\right|\stackrel{\&OverBar;}{v_{x_l}},\stackrel{\&OverBar;}{v_{seed}}\left|\right)<r_{th}---\left(12\right)$

Wherein, r _threpresentation vector threshold, this value of different model slightly difference, this value needs artificial appointment.If meet above-mentioned condition, then by this neighborhood point join in current region Rc.

Step 4.3: judge current neighborhood point curvature whether meet the curvature threshold condition shown in formula (13), if met; by neighborhood point join current seed region S _cin:

$Cur\left(\stackrel{\&RightArrow;}{x_l}\right)<c_{th}---\left(13\right)$

Wherein, c _threpresent curvature threshold, this value of different model slightly difference, this value needs artificial appointment;

Step 4.4: delete current seed region S _cin current seed point, and by this data point from set U delete;

Step 4.5: choose current seed region S _cin next data point as current seed point, return step 4.2, until seed region S _cfor sky, perform step 4.6;

Step 4.6: if set U be empty, then the data point in current region Rc preserved hereof, current region Rc is set to sky again, chooses next data point in SEEDS as current seed point, returns step 4.1, until set U be empty, split end; Data point in then different current region Rc is stored in different files respectively, realizes the classification of data point, namely realizes the segmentation of cloud data.

In above-mentioned zone propagation process, a data point successively in selection Seed Points S set EEDS is as current seed point, until after this Seed Points operated, a bit as current seed point after selecting again, can realize Seed Points by this algorithm automatically to select, thus the result avoiding region to increase depends on that Seed Points chooses this limitation condition to a great extent.

Experimental result

Experiment 1

Adopt method of the present invention, region growing algorithm respectively ^[2]and DBSCAN ^[3]the three-dimensional point cloud model data of algorithm to milpa are split.The different part that numeral three-dimensional point cloud model in Fig. 2 is divided into.Can be found out by Fig. 2 (b), for the multiple objects in model, the present invention is divided into 6 parts, from the position 5 figure, position 6 can find out that the present invention has carried out effective segmentation to 2 independent milpas, corn has been carried out effective extraction from complex environment, simultaneously also comparatively regular for position 1-4 segmentation effect of the present invention.Can find out that the position 6 Fig. 2 (b) is divided into 4 parts by the region growing algorithm do not improved from Fig. 2 (c), compared with the present invention, there is the problem of over-segmentation in this algorithm.Can find out that the position 6 Fig. 2 (b) has been divided into 5 parts by DBSCAN algorithm from Fig. 2 (d), not keep the integrality of milpa individuality, also there are some and split excessive problem in its segmentation result.Can be found by the contrast of above-mentioned experimental result, the present invention can distinguish object comparatively accurately, keeps the integrality of object, can reach a good segmentation effect.

Experiment 2

Method of the present invention, region growing algorithm and the DBSCAN algorithm three-dimensional point cloud model data to tree are adopted to split respectively.The different part that numeral three-dimensional point cloud model in Fig. 3 is divided into.Can be found out by Fig. 3 (b), model is divided into 7 parts by the present invention, and part 1-6 represents leaf, regular to the blade segmentation of each part, does not destroy the integrality of leaf, has also carried out effective extraction to branch is dry.The position 1 of comparison diagram 3 (c) position 1-4 and Fig. 3 (b), can find out that leaf is divided into 4 parts by region growing algorithm, destroy the integrality of leaf, the position 5 of contrast position 5-8 and Fig. 3 (b) can find out that limb and leaf are effectively distinguished, and also destroys the integrality of leaf simultaneously.The position 1 of 1-5 partly and in Fig. 3 (b) of comparison diagram 3 (d) can be found out, DBSCAN algorithm causes the over-segmentation of leaf, the 7-8 part of comparison diagram 3 (d) and the position 5 of Fig. 3 (b), the leaf in position 8 and limb segmentation dynamics deficiency can be found out, do not distinguish limb and leaf.Can find out that both region growing algorithm and DBSCAN algorithm exist the segmentation of some position excessively from Fig. 3 (c) and Fig. 3 (d), the problem that the segmentation of some position is not enough, relatively can find out by above-mentioned segmentation result, for discrete objects, the present invention can effectively distinguish limb, leaf, and segmentation effect is relatively better.

List of references:

[1]RabbaniT,vandenHeuvelF,VosselmannG.Segmentationofpointcloudsusingsmoothnessconstraint[J].InternationalArchivesofPhotogrammetry,RemoteSensingandSpatialInformationSciences,2006,36(5):248-253.

[2] Hu Huaiyu, etc. based on the dispersion point cloud partition method [J] of city, district growth method. computer utility, 2009,29 (10): 2716-2718.

[3]EsterM,KriegelHP,SanderJ,etal.Adensity-basedalgorithmfordiscoveringclustersinlargespatialdatabaseswithnoise[J].ProceedingsofInternationalConferenceonKnowledgeDiscovery&DataMining,1996:226--231.

Claims (5)

1. an adaptive point cloud data segmentation method, is characterized in that, specifically comprises the following steps:

Step 1: obtain cloud data, carry out feature extraction to cloud data, obtains the local density and closely of each data point;

Step 2: the local density of each data point obtained according to step 1 and closely, sorts to each data point, obtains the Seed Points set after sorting;

Step 3: calculate normal vector corresponding to each data point and curvature;

Step 4: the normal vector that each data point obtained according to step 3 is corresponding and curvature, utilize region growing algorithm, data point in the Seed Points set obtained according to step 2 carries out automatically choosing of Seed Points, classifies to all data points, realizes the segmentation of cloud data.

2. adaptive point cloud data segmentation method as claimed in claim 1, is characterized in that, comprising the implementation procedure that cloud data carries out feature extraction of described step 1:

Step 1.1: calculate data point x _i, the distance in 1≤i≤N between any two data points;

Step 1.2: the distance between any two data points step 1.1 obtained, generating new set d according to sorting from small to large, calculating and blocking distance d _c;

Step 1.3: what obtain according to step 1.2 blocks distance d _c, calculate each data point x _ilocal density ρ _i;

Step 1.4: the local density ρ that step 1.3 is obtained _icarry out descending sort, regenerate a subscript sequence;

Step 1.5: according to the distance between any two data points that step 1.1 obtains, calculate each data point closely q _irepresent the subscript numerical value of the subscript sequence described in step 1.4.

3. adaptive point cloud data segmentation method as claimed in claim 2, is characterized in that, the implementation procedure of described step 2 comprises:

Step 2.1: the local density ρ of each data point utilizing step 1 to obtain _iclosely obtain the judgment value PU of each data point _i;

Step 2.2: by data point x _iaccording to PU _ivalue carry out descending sort, generate Seed Points S set EEDS.

4. adaptive point cloud data segmentation method as claimed in claim 2, is characterized in that, the implementation procedure of described step 3 comprises:

Step 3.1: for each data point x _i, utilize the Neighborhood-region-search algorithm based on FDN to obtain each data point x _ineighbor point;

Step 3.2: calculate each data point x _ithe three-dimensional barycenter of all neighbor points;

Step 3.3: each data point x of the centroid calculation utilizing step 3.2 to obtain _icorresponding covariance matrix Cov _i;

Step 3.4: according to the covariance matrix Cov obtained of step 3.3 _icalculate each data point x _icharacteristic of correspondence value and proper vector, obtain data point x _icorresponding normal vector

Step 3.5: each the data point x obtained according to step 3.4 _icharacteristic of correspondence value, calculates each data point x _icurvature Cur _i.

5. adaptive point cloud data segmentation method as claimed in claim 4, is characterized in that, the implementation procedure of described step 4 comprises:

Step 4.1: note set Nbhd is data neighborhood of a point, and the element in set Nbhd is the neighbor point of data point, and set Rc is current region, and initial value is empty, S _cfor current seed region, initial value is empty; Select first data point seed in Seed Points S set EEDS ₁as initial current seed point, be inserted into current seed region S _cin;

Step 4.2: utilize FDN Neighborhood-region-search algorithm to obtain the neighborhood Nbhd of current seed point, to each neighbor point in neighborhood Nbhd l represents l neighbor point of current seed point, judges its normal vector with the normal vector of current seed point between angle whether meet the normal vector threshold condition of setting, if meet, then by this neighborhood point join in current region Rc;

Step 4.3: judge current neighborhood point curvature whether meet the curvature threshold condition of setting, if meet, then by neighborhood point join current seed region S _cin;

Step 4.4: delete current seed region S _cin current seed point, and by this data point from set U delete;

Step 4.5: choose current seed region S _cin next data point as current seed point, return step 4.2, until seed region S _cfor sky, perform step 4.6;

Step 4.6: if set U be empty, then the data point in current region Rc preserved hereof, current region Rc is set to sky again, chooses next data point in SEEDS as current seed point, returns step 4.1, until set U be empty, split end; Data point in then different current region Rc is stored in different files respectively, realizes the classification of data point, namely realizes the segmentation of cloud data.