CN110046661A - A kind of vehicle-mounted cloud clustering method cutting algorithm based on contextual feature and figure - Google Patents
A kind of vehicle-mounted cloud clustering method cutting algorithm based on contextual feature and figure Download PDFInfo
- Publication number
- CN110046661A CN110046661A CN201910285459.3A CN201910285459A CN110046661A CN 110046661 A CN110046661 A CN 110046661A CN 201910285459 A CN201910285459 A CN 201910285459A CN 110046661 A CN110046661 A CN 110046661A
- Authority
- CN
- China
- Prior art keywords
- super voxel
- super
- voxel
- algorithm
- cloud
- 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.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 36
- 238000002372 labelling Methods 0.000 claims abstract description 16
- 238000004364 calculation method Methods 0.000 claims abstract description 8
- 238000011156 evaluation Methods 0.000 claims abstract description 8
- 230000011218 segmentation Effects 0.000 claims description 13
- 238000004458 analytical method Methods 0.000 claims description 5
- 238000005259 measurement Methods 0.000 claims description 3
- 238000010586 diagram Methods 0.000 description 13
- VMXUWOKSQNHOCA-UKTHLTGXSA-N ranitidine Chemical compound [O-][N+](=O)\C=C(/NC)NCCSCC1=CC=C(CN(C)C)O1 VMXUWOKSQNHOCA-UKTHLTGXSA-N 0.000 description 8
- 238000005457 optimization Methods 0.000 description 5
- 230000000694 effects Effects 0.000 description 3
- 230000008569 process Effects 0.000 description 3
- NAWXUBYGYWOOIX-SFHVURJKSA-N (2s)-2-[[4-[2-(2,4-diaminoquinazolin-6-yl)ethyl]benzoyl]amino]-4-methylidenepentanedioic acid Chemical compound C1=CC2=NC(N)=NC(N)=C2C=C1CCC1=CC=C(C(=O)N[C@@H](CC(=C)C(O)=O)C(O)=O)C=C1 NAWXUBYGYWOOIX-SFHVURJKSA-N 0.000 description 1
- 238000007792 addition Methods 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 238000000605 extraction Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000008520 organization Effects 0.000 description 1
- 238000000513 principal component analysis Methods 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005070 sampling Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/20—Analysis of motion
- G06T7/246—Analysis of motion using feature-based methods, e.g. the tracking of corners or segments
Landscapes
- Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Image Analysis (AREA)
Abstract
The present invention provides a kind of vehicle-mounted cloud clustering method that algorithm is cut based on contextual feature and figure, including cut-point cloud forms super voxel, divides point cloud data according to being connected property of space density and forms different join domains, each join domain becomes super voxel；Context relation feature calculation between super voxel, the context relation feature between affiliated super voxel include space correlation feature and semantic association feature, and affiliated space correlation includes direction, distance and topology aspect, and the semantic association includes dimension and vpg connection；The figure of multiple labeling, which is cut, to cluster, including using super voxel as node, is constituted graph model as side using the line between super voxel, the connectivity of the super voxel of algorithm total evaluation is cut with figure, reconsolidates to form new point cloud and cluster.The present invention does not need priori knowledge, can not only be effectively improved over-segmentation phenomenon in cloud cluster, and greatly improves a cloud and cluster the precision of result, provides the basic data of high quality for subsequent target identification.
Description
Technical field
The present invention relates to a kind of vehicle-mounted cloud clustering methods, cut algorithm based on contextual feature and figure more particularly, to one kind
Vehicle-mounted cloud clustering method.
Background technique
Vehicle-carried mobile scanning system provides the acquisition mode of quick obtaining city streetscape point cloud data, from a large amount of, unordered
Point cloud data in accurately analyze, interpret the target that scene objects are cloud clusters and sort research.The precision of point cloud cluster is straight
Connect influence subsequent classification, information extraction and target Geometric model reconstruction etc. as a result, for clustering in existing cloud clustering method
Be generally characterized by calculating based on point or voxel, the feature based on point will receive point cloud data quality such as density unevenness, different
The influence of constant value and noise etc., voxel-based feature then will receive the influence of point cloud data voxel discretization resolution ratio.It clusters
Constraint gradually turns under the conditions of global optimization the connectivity between consideration data from subrange, as condition random field and figure are cut
Equal Global Optimization Models have been used to point cloud data and cluster, and wherein the weight on the side of correlation is determining very heavy between expression node
It wants, the distance restraint being mostly used between vertex neighborhood in existing algorithm defines the weight on side, and the relevance contained between data point considers to owe
It lacks.
Summary of the invention
Present invention mainly solves the technical problem present in the prior art, one kind is provided using super voxel as object, in conjunction with
Spatial context association and figure cut vehicle-mounted cloud clustering method of algorithm.
Above-mentioned technical problem of the invention is addressed by following technical proposals:
The present invention provides a kind of vehicle-mounted cloud clustering method that algorithm is cut based on contextual feature and figure, including following step
It is rapid:
Step 1, cut-point cloud forms super voxel, different including being formed according to being connected property of space density segmentation point cloud data
Join domain, each join domain become super voxel；
Step 2, context relation feature calculation between super voxel, the context relation feature between affiliated super voxel includes space
Linked character and semantic association feature, affiliated space correlation include direction, distance and topology aspect, and the semantic association includes dimension
Degree and vpg connection；
Step 3, the figure of multiple labeling, which is cut, clusters, including using super voxel as node, is constituted figure by side of the line between super voxel
Model cuts the connectivity of the super voxel of algorithm total evaluation with figure, reconsolidates to form new point cloud and cluster.
Moreover, the implementation that cut-point cloud forms super voxel is as follows in step 1,
Step 1.1, analysis site cloud density, set point neighborhood search range Eps；Cloud is segmented according to height value, setting
The corresponding smallest sample number Minpts of cloud is put within the scope of different elevations；
Step 1.2, according to vertex neighborhood search range Eps and smallest sample number Minpts, using DBSCAN algorithm to a cloud
It is split；
Step 1.3, the point cluster obtained to step 1.2 segmentation carries out points constraint, if points are too small, marks in the cluster
Point be noise spot, remaining point cluster then become super voxel.
Moreover, in step 2, context relation feature calculation between super voxel is realized using ambiguity function metric form.
Moreover, the directional correlation of super voxel is with two super voxel spatial position lines in horizontal plane and vertical plane projection and X
The angle theta of axishAnd θzIt indicates.
Moreover, two super voxel spatial position plan range dD of the distance relation of super voxelijWith height difference dHijIt indicates.
Moreover, the topological relation of super voxel with the corresponding cuboid bounding box of two super voxels space intersection with mutually fetch
Indicate
Moreover, ratio shared by one-dimensional, two-dimentional and three-dimensional point number is in the super voxel of the similitude of super voxel dimension come table
Show.
Moreover, the similitude of super voxel shape is indicated with cuboid bounding box side ratio.
Moreover, step 3 includes following sub-step,
Step 3.1, the graph model building of super voxel segmentation is carried out, the node of graph model includes super voxel and label endpoint,；
Step 3.2, graph model globally optimal solution is sought, including energy is minimized using approximate to multiple labeling energy function E (f)
Quantity algorithm obtains multiple labeling globally optimal solution, determines that optimum label configures,
The multiple labeling energy function E (f) is as follows,
E (f)=Edata(f)+Esmooth(f)
In formula, Edata(f),Esmooth(f) data item and smooth item are respectively represented；
Data item Edata(f) label f and the inconsistency for being observed data are described,
In formula, p is arbitrarily super voxel in data set P, fpFor the label of p, DpMeasurement labels fpWith observation data it is inconsistent
Property；In the present invention when super voxel and tag number are consistent or two super voxel semantic contexts be associated withs big, punishment Dp(fp) smaller,
Otherwise D is punishedp(fp) bigger；
Smooth item Esmooth(f) the unsmooth degree of piecemeal of label f is described,
Esmooth=∑p,q∈NVp,q(fp,fq)
Wherein N is super voxel Neighbourhood set, and p and q are the super voxel in N, fpAnd fqThe respectively corresponding label of p, q.Work as neighbour
In domain when the directional correlation of super voxel, distance association and very big topological correlation degree, if label is inconsistent, V is punishedp,qIt is bigger；It is no
Then, V is punishedp,qWith regard to smaller.
The present invention does not need priori knowledge, graph model is constituted as node using the super voxel that density is reachable, between super voxel
Space is associated with the weight for defining graph model side with property context, cuts optimization algorithm using figure and obtains most preferably clustering for super voxel.
This method can not only be effectively improved over-segmentation phenomenon in cloud cluster, and greatly improve a cloud and cluster the precision of result,
The basic data of high quality is provided for subsequent target identification.
Detailed description of the invention
Fig. 1 is context relation schematic diagram in space in the embodiment of the present invention, and wherein Fig. 1 (a) is directional correlation schematic diagram, figure
1 (b) is distance association schematic diagram, and Fig. 1 (c) is that topological correlation (connects and intersect) schematic diagram.
Fig. 2 is that the figure of multiple labeling in the embodiment of the present invention cuts algorithm schematic diagram, and wherein Fig. 2 (a) is graph model schematic diagram, figure
2 (b) be label schematic diagram.
Fig. 3 is flow diagram of the embodiment of the present invention.
Specific embodiment
Below in conjunction with drawings and examples, the technical solutions of the present invention will be further described.
For problem of the prior art, the present invention proposes a kind of new point cloud clustering method, carries out first to point cloud data
Over-segmentation, the unit that obtained super voxel is calculated as subsequent cluster feature, then introduces space and property context feature is come
The association between point cloud data is described, and further defines the weight on the graph model side of super voxel building, is finally based on multi-signature
It cuts optimization algorithm and realizes that optimal super voxel clusters.The voxel that traditional clustering method uses is specification cube, and the present invention mentions
Super voxel out is the point set of preliminary clusters.Relatively traditional clustering method, this method can effectively improve point cloud data and clustered
Divide phenomenon, the effect that over-segmentation improves in the case where dispersing particularly with atural object is more obvious.It is in super voxel due to clustering
On the basis of heavy label, the time for calculating cost greatly reduces compared with the clustering method based on point, greatly high clustering algorithm
Efficiency.In addition, the global optimization that the anti-noise and figure of super voxel cut algorithm ensure that the method for the present invention clusters result with high-precision
And high reliability.
Referring to Fig. 3, a kind of vehicle-mounted cloud clustering method that algorithm is cut based on contextual feature and figure that embodiment provides is real
It is now as follows:
Step 1, cut-point cloud forms super voxel.Divide point cloud data according to being connected property of space density and forms different connections
Region, each join domain become super voxel.Opposite individual data point, super voxel are the description closer data blocks of atural object, together
When also can more preferably embody the attribute and spatial correlation of data.
In embodiment, cut-point cloud forms super voxel and includes the following steps: in step 1
Step 1.1, analysis site cloud density, set point neighborhood search range Eps；Cloud is segmented according to height value, is given
The corresponding smallest sample number Minpts of cloud is put within the scope of different elevations；
Step 1.2, according to vertex neighborhood search range Eps and smallest sample number Minpts, DBSCAN (Density- is utilized
Based Spatial Clustering of Applications with Noise, has noisy density clustering
Method) algorithm is split a cloud.
For the sake of ease of implementation, embodiment is provided and realizes that process is as follows:
To be processed queue is established, takes untreated any point p in a cloud to be added in be processed queue, according to point p's
Height value obtains its corresponding smallest sample number Minpts.At neighborhood search range Eps, the neighborhood density value of point p is calculated
pValue.It Minpts and is not labeled if pValue is greater than, mark point p and its neighborhood point are new point cluster, and by neighborhood point
It is added in be processed queue；If pValue is greater than Minpts, and point p has been marked as some point cluster, then by the neighborhood of point p
Point is labeled as the cluster, and is added in be processed queue；Otherwise pValue is less than Minpts, and mark point p is noise spot.According to
The secondary point in queue carries out aforesaid operations, if queue is sky, reselects untreated any point in a cloud and is added wait locate
In reason point queue, until all the points are labeled, then complete to divide；
Step 1.3, points constraint is carried out to the obtained point cluster of segmentation, if points are too small, marks the point in the cluster to be
Noise spot, remaining point cluster then become super voxel.When it is implemented, respective threshold can be set, marked when points are less than threshold value
For noise spot.Threshold value setting is related with point cloud data density and scene atural object, such as the value takes 10 in the present embodiment.
Step 2, context relation feature calculation between super voxel.Context relation feature between super voxel includes that super voxel exists
Direction, distance and topology etc. space correlation, and dimension and in terms of semantic association.The suitably fuzzy letter of selection
Number measure, calculates the context relation characteristic value of super voxel, these values will divide energy letter in subsequent super voxel graph model
Assignment data item and smooth item in number.
Context relation feature calculation includes the following steps: between super voxel in embodiment
Step 2.1, spatial context linked character calculates, including direction, distance and topological correlation feature calculation.Two super
Voxel viAnd vjSpace correlation between (i, j=1 ..., n) can be described as directional correlation, distance association and topological correlation (phase
Connect and intersect), as shown in Figure 1, Fig. 1 (a) is directional correlation schematic diagram in wherein, Fig. 1 (b) is distance association schematic diagram, Fig. 1
It (c) is that topological correlation (connects and intersect) schematic diagram.
Fuzzy membership functions v (vi,vj) indicate two super voxel vsiAnd vjSpace correlation degree, as shown in formula (1).
Wherein, θijFor the directional correlation between two super voxels, dijDistance association between two super voxels, ρijIt is two
Topological correlation between super voxel.Work as θij、dijOr ρijWhen meeting corresponding threshold value constraint condition T, degree of association function f (θij,
dij,ρij) ∈ [0 1] expression.
(1) directional correlation
The spatial position v of super voxeli(Xi, Yi, Zi) by n point (x all in super voxeli,yi,zi) three-dimensional coordinate mean value
It indicates.
The directional correlation of super voxel is with two super voxel spatial position lines in horizontal plane (O-XY) and vertical plane (O-XZ)
The angle theta of projection and X-axishAnd θzIt indicates, the horizontal direction degree of association are as follows:
The vertical direction degree of association are as follows:
T in formulahθAnd TzθThe respectively threshold value at horizontal direction angle and vertical direction angle.
When it is implemented, some atural objects there are horizontal relevance, there is vertical association in some atural objects, the form connection that can be summed it up
It closes and uses.
(2) distance association
Two super voxel spatial position plan range dD of the distance relation of super voxelijWith height difference dHijIt indicates, plane
Apart from the degree of association are as follows:
The height difference degree of association are as follows:
T in formuladDAnd TdHRespectively horizontal distance and height difference threshold value.
When it is implemented, there are plan range associations for some atural objects, some atural objects are there are height difference association, the shape that can be summed it up
Formula is used in combination.
(3) topological correlation
In the corresponding cuboid bounding box in space, the length of bounding box and its mutual ratio are used to super voxel
The spatial form of super voxel is described.The topological relation of super voxel is with the corresponding cuboid bounding box of two super voxels in space intersection
Expression is mutually fetched, as shown in Fig. 1 (c).Topological correlation degree are as follows:
L, w, h are respectively the length and width and height of super voxel bounding box, i.e. l in formulai、wi、hiRespectively super voxel viIt is corresponding super
The length and width and height of voxel bounding box, lj、wj、hjRespectively super voxel vjThe length and width and height of corresponding super voxel bounding box；(Xi,Yi,
Zi) it is super voxel viPosition coordinates, (Xj,Yj,Zj) it is super voxel vjPosition coordinates.When two super voxels are at X, Y, Z tri-
When the coordinate range in direction has coincidence, super voxel viAnd vjIntersection.If two super voxels, 2 direction coordinate ranges in X, Y, Z
There is coincidence, third direction coordinate connects, then super voxel viAnd vjConnect.
Step 2.2, semantic context linked character value calculates, and the semantic association of super voxel can use the dimension of two super voxels
Counting with shape similarity indicates.
For space any point (abbreviation spatial point), PCA in the prior art (principal component is utilized
Analysis, principal component analysis) algorithm calculates its corresponding dimension, spatial point is described with one-dimensional, two-dimentional and three-dimensional point respectively
Linearly, flatness and dotted attribute.λD1、λD2And λD3Ratio shared by one-dimensional, two-dimentional and three-dimensional point number in respectively super voxel
Example, for indicating the dimension information of super voxel.To a point, λ can be calculated by PCA simultaneouslyD1、λD2And λD3.Belong to same
The super voxel of class atural object, dimension information have correlation, such as belong to the super voxel of building facade, dimension information should include
Two-dimentional planar point, the super voxel dimension for belonging to trees should be comprising one-dimensional linear point and through point etc..Therefore, the present invention selects
Ratio shared by one-dimensional, two-dimentional and three-dimensional point number indicates in the super voxel of the similitude of super voxel dimension.
(1) the dimension degree of association are as follows:
C () is vi, vjSuper voxel dimension information (λD1、λD2And λD3) similarity degree comparison function, when two super voxels 3
Dimension value otherness is in threshold value TλWithin, then two super voxel dimension associations.In embodiment, c () uses super voxel viAnd vjIt is right
The λ answeredD1、λD2And λD3Three absolute values for subtracting each other difference respectively and.
(2) similitude of super voxel shape indicates respectively S with cuboid bounding box side ratio1=h/l, S2=h/w and
S3=l/w, corresponding shape correlation degree are as follows:
D () is vi, vjSuper voxel shape (S1、S2And S3) degree of association comparison function, when two super voxel shape similarities are poor
It is different in threshold value TsWithin, then two super voxel shape correlations.In embodiment, d () uses super voxel viAnd vjCorresponding S1、S2With
S3Three absolute values for subtracting each other difference respectively and.
Figure is cut in algorithm, and semantic association is used to determine data item, and space correlation is used to determine smooth item.
Step 3, the figure of multiple labeling, which is cut, clusters.Using super voxel as node, the line between super voxel is that side constitutes graph model, is used
Figure cuts the connectivity of the super voxel of algorithm total evaluation, reconsolidates to form new point cloud and cluster.
In the step 3 of embodiment, the figure of multiple labeling, which cuts to cluster, to be included the following steps:
Step 3.1, the graph model building of super voxel segmentation is carried out.
If using super voxel as the node of graph model, side of their line as graph model then may be constructed super body
The markov random file (MRF) of element segmentation.In MRF model, super voxel segmentation can be regarded as a multiple labeling problem, such as
Shown in Fig. 2.Fig. 2 (a) is graph model schematic diagram, and Fig. 2 (b) is label schematic diagram, and the node of graph model includes super voxel and label
Endpoint, super voxel p are indicated with square, mark endpoint f1, f2... fnIt indicates, initial markers endpoint number is in the present invention
Super voxel number, the side of graph model include between line n-link and super voxel and label endpoint between super voxel and the super voxel of neighborhood
Line t-link.For arbitrarily super voxel p, finite aggregate (f is distributed1, f2... fnAs soon as) in label, the target of segmentation
It is to find optimal label configuration, it is smooth that it will not only meet piecemeal, to be also consistent with data are observed.
Step 3.2 seeks graph model globally optimal solution.
In computer vision field, optimal label configuration is found, energy minimization problem can be expressed as, according to
Formula (10) constructs a multiple labeling energy function.
E (f)=Edata(f)+Esmooth(f) (10)
E in formuladata(f),Esmooth(f) data item and smooth item are respectively represented.Data item Edata(f) describe label f with
It is observed the inconsistency of data.
In the embodiment of the present invention, data item are as follows:
In formula, p is arbitrarily super voxel in data set P, fpFor the label of p, DpMeasurement labels fpWith observation data it is inconsistent
Property.
In the present invention when super voxel and tag number are consistent or two super voxel semantic contexts be associated withs big, punishment Dp(fp)
It is smaller, otherwise punish Dp(fp) bigger.
Smooth item Esmooth(f) the unsmooth degree of piecemeal of label f is described.Smooth item indicates in the embodiment of the present invention are as follows:
Esmooth=Σp,q∈NVp,q(fp,fq) (12)
Wherein N is super voxel Neighbourhood set, and p, q are the super voxel in N, fpAnd fqThe respectively corresponding label of p, q.Work as neighbour
In domain when the directional correlation of super voxel, distance association (plan range is associated with height difference) and very big topological correlation degree, if label is not
Unanimously, then V is punishedp,qIt is bigger；Otherwise, V is punishedp,qWith regard to smaller.
In embodiment, if the size of relating value is 0-1,0 is minimum, and 1 is maximum.
ENERGY E (f) algorithm is minimized using approximation, multiple labeling globally optimal solution is obtained and is configured to get to optimum label.
Therefore, the present invention has the advantage that in vehicle-mounted cloud cluster, with super group of voxels organization data, tap point cloud number
According to space be associated with property context, under the conditions of figure cuts algorithm globally optimal solution obtain point cloud data most preferably cluster.
When it is implemented, computer software technology, which can be used, in the above process realizes automatic running.This method is realized in operation
Hardware should also be as within the scope of the present invention.
For ease of illustration technical effect of the invention is the specific example data using the present embodiment method below:
Experimental data is the city streetscape point cloud data that Reigle vux-1 terrestrial Laser scanner obtains, and point cloud point number is
2180726.The sampling interval of data is 0.01 °, mainly has building surface, power line, electric pole, trees and street lamp comprising object
Traffic sign bar etc..
Experimental data calculating process and result output
Step 1. pair point cloud is split, and generates super voxel.Super voxel segmentation is carried out using DBSCAN algorithm, searches for neighborhood
Radius Eps is 0.6m, and when the height value of point is located at 14.34m between 22.34m, MinPts 450, otherwise MinPts is 50,
212 super voxels are obtained after segmentation.
Step 2. surpasses voxel and constitutes graph model, by whether there is space correlation between the super voxel of context relation threshold decision
Related to semanteme, specific threshold parameter is shown in Table 1.
1 context relation threshold value of table
Step 3. figure cuts algorithm and realizes that super voxel multiple labeling best configuration, each super voxel initial markers state are respectively one
Other super voxel semantic relevancies are used to judge that super voxel to be processed becomes in data set in cluster, super voxel to be processed and data set
A possibility that other super voxel labels, so that it is determined that figure cuts the data item penalty value of algorithm.Super voxel to be processed and the super body of neighborhood
Plain spatial correlation is used to determine that figure cuts the smooth item penalty value of algorithm.Figure cuts the design parameter in algorithm and is shown in Table 2, data item by
Super voxel D to be markedpSurpass voxel D with otherpIt constitutes.Since super voxels all in data set are respectively initially cluster, so wait mark
Remember super voxel DpIt is 0, i.e., without punishment, and surpasses voxel D with otherpThen according to the condition assignment in table 2, for example, if to be marked
Super voxel and other super voxel λ1DRelated and TSCorrelation, DpIt is 2, i.e., punishes the inconsistent of label.Smooth item Vp,qValue root
It is obtained according to the relevance of two in table 2 super voxel p and q, for example, if distance, direction and space intersection connect, any one is associated with,
Vp,qIt is 0, i.e., does not punish.
2 figure of table cuts algorithm parameter
In order to have objective appraisal analysis to cluster result, the present invention uses common F_measure evaluation of classification method, should
Evaluation of classification method is combined to precision ratio in information retrieval (Precision) and recall ratio (Recall), to realize
The evaluation of cluster result.The Precision and Recall of certain cluster j and its relevant classification i is defined as:
P=precision (i, j)=Nij/Nj (13)
R=recall (i, j)=Nij/Ni
Wherein, NijIt is the number of the classification i in cluster j；NjIt is the number for clustering all objects in j；NiIt is institute in classification i
There is the number of object.The F_measure of classification i is defined as:
F (i)=2PR/ (P+R) (14)
Total F_measure can be weighted and averaged to obtain by the F_measure of each classification i.
The present invention is clustered with major surface features in experimental data (building, electric pole, trees and power line) statistic algorithm
Precision ratio and recall ratio, the results are shown in Table shown in 3.
Table 3 recall ratio and precision ratio of the invention
The results show that four kinds of major surface features types have higher precision ratio and recall ratio, wherein power line look into it is complete
Rate is opposite to want lower, and this is mainly due to electric wire and electric pole, power line and trees, in junction, there are a small amount of mistakes to gather
Cluster.If the practical total points of power line be that 28331, about 1553 points are gathered in electric pole cluster, about 1998 point mistakes gather to set by mistake
In the wooden cluster.For the effect of overall evaluation this paper algorithm, the average precision of major surface features and recall ratio in data and total are calculated
F_measure value, respectively 96.17%, 91.42% and 93.67%.
Specific embodiment described herein is only an example for the spirit of the invention.The neck of technology belonging to the present invention
The technical staff in domain can make various modifications or additions to the described embodiments or replace by a similar method
In generation, however, it does not deviate from the spirit of the invention or beyond the scope of the appended claims.
Claims (9)
1. a kind of vehicle-mounted cloud clustering method for cutting algorithm based on contextual feature and figure, comprising the following steps:
Step 1, cut-point cloud forms super voxel, forms different connections including dividing point cloud data according to being connected property of space density
Region, each join domain become super voxel；
Step 2, context relation feature calculation between super voxel, the context relation feature between affiliated super voxel includes space correlation
Feature and semantic association feature, affiliated space correlation include direction, distance and topology aspect, the semantic association include dimension and
Vpg connection；
Step 3, the figure of multiple labeling, which is cut, clusters, including using super voxel as node, is constituted graph model as side using the line between super voxel,
The connectivity that the super voxel of algorithm total evaluation is cut with figure reconsolidates to form new point cloud and cluster.
2. vehicle-mounted cloud clustering method of algorithm is cut based on contextual feature and figure according to claim 1, it is characterised in that:
In step 1, the implementation that cut-point cloud forms super voxel is as follows,
Step 1.1, analysis site cloud density, set point neighborhood search range Eps；Cloud is segmented according to height value, setting is different
The corresponding smallest sample number Minpts of cloud is put within the scope of elevation；
Step 1.2, according to vertex neighborhood search range Eps and smallest sample number Minpts, using DBSCAN algorithm to point Yun Jinhang
Segmentation；
Step 1.3, the point cluster obtained to step 1.2 segmentation carries out points constraint, if points are too small, marks the point in the cluster
For noise spot, remaining point cluster then becomes super voxel.
3. vehicle-mounted cloud clustering method of algorithm is cut based on contextual feature and figure according to claim 1, it is characterised in that:
In step 2, context relation feature calculation between super voxel is realized using ambiguity function metric form.
4. vehicle-mounted cloud clustering method of algorithm is cut based on contextual feature and figure according to claim 3, it is characterised in that:
The directional correlation of super voxel projects the angle theta with X-axis in horizontal plane and vertical plane with two super voxel spatial position lineshAnd θz
It indicates.
5. vehicle-mounted cloud clustering method of algorithm is cut based on contextual feature and figure according to claim 3, it is characterised in that:
Two super voxel spatial position plan range dD of the distance relation of super voxelijWith height difference dHijIt indicates.
6. vehicle-mounted cloud clustering method of algorithm is cut based on contextual feature and figure according to claim 3, it is characterised in that:
The topological relation of super voxel is indicated in space intersection with mutually fetching with the corresponding cuboid bounding box of two super voxels.
7. vehicle-mounted cloud clustering method of algorithm is cut based on contextual feature and figure according to claim 3, it is characterised in that:
Ratio shared by one-dimensional, two-dimentional and three-dimensional point number indicates in the super voxel of the similitude of super voxel dimension.
8. vehicle-mounted cloud clustering method of algorithm is cut based on contextual feature and figure according to claim 3, it is characterised in that:
The similitude of super voxel shape is indicated with cuboid bounding box side ratio.
9. the according to claim 1 or 2 or 3 or 4 or 5 or 6 or 7 or 8 vehicle-mounted points for cutting algorithm based on contextual feature and figure
Cloud clustering method, it is characterised in that: step 3 includes following sub-step,
Step 3.1, the graph model building of super voxel segmentation is carried out, the node of graph model includes super voxel and label endpoint,；
Step 3.2, graph model globally optimal solution is sought, including multiple labeling energy function E (f) is calculated using the approximate energy that minimizes
Method obtains multiple labeling globally optimal solution, determines that optimum label configures,
The multiple labeling energy function E (f) is as follows,
E (f)=Edata(f)+Esmooth(f)
In formula, Edata(f),Esmooth(f) data item and smooth item are respectively represented；
Data item Edata(f) label f and the inconsistency for being observed data are described,
In formula, p is arbitrarily super voxel in data set P, fpFor the label of p, DpMeasurement labels fpWith the inconsistency of observation data；
In the present invention when super voxel and tag number are consistent or two super voxel semantic contexts be associated withs big, punishment Dp(fp) smaller, otherwise
Punish Dp(fp) bigger；
Smooth item Esmooth(f) the unsmooth degree of piecemeal of label f is described,
Esmooth=Σp,q∈NVp,q(fp,fq)
Wherein N is super voxel Neighbourhood set, and p and q are the super voxel in N, fpAnd fqThe respectively corresponding label of p, q.When in neighborhood
When the directional correlation of super voxel, distance are associated with and topological correlation degree is very big, if label is inconsistent, V is punishedp,qIt is bigger；Otherwise,
Punish Vp,qWith regard to smaller.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910285459.3A CN110046661A (en) | 2019-04-10 | 2019-04-10 | A kind of vehicle-mounted cloud clustering method cutting algorithm based on contextual feature and figure |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910285459.3A CN110046661A (en) | 2019-04-10 | 2019-04-10 | A kind of vehicle-mounted cloud clustering method cutting algorithm based on contextual feature and figure |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN110046661A true CN110046661A (en) | 2019-07-23 |
Family
ID=67276651
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910285459.3A Pending CN110046661A (en) | 2019-04-10 | 2019-04-10 | A kind of vehicle-mounted cloud clustering method cutting algorithm based on contextual feature and figure |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110046661A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111582352A (en) * | 2020-04-30 | 2020-08-25 | 上海高仙自动化科技发展有限公司 | Object-based sensing method and device, robot and storage medium |
| CN111815776A (en) * | 2020-02-04 | 2020-10-23 | 山东水利技师学院 | Three-dimensional building fine geometric reconstruction method integrating airborne and vehicle-mounted three-dimensional laser point clouds and streetscape images |
| CN112070769A (en) * | 2020-09-18 | 2020-12-11 | 福州大学 | Layered point cloud segmentation method based on DBSCAN |
| CN112233237A (en) * | 2020-10-23 | 2021-01-15 | 广州建通测绘地理信息技术股份有限公司 | Water area leveling processing method and computer equipment for manufacturing digital elevation model |
| KR20210043518A (en) * | 2020-06-28 | 2021-04-21 | 베이징 바이두 넷컴 사이언스 테크놀로지 컴퍼니 리미티드 | High-precision mapping method and device |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106780524A (en) * | 2016-11-11 | 2017-05-31 | 厦门大学 | A kind of three-dimensional point cloud road boundary extraction method |
| CN107292276A (en) * | 2017-06-28 | 2017-10-24 | 武汉大学 | A kind of vehicle-mounted cloud clustering method and system |
-
2019
- 2019-04-10 CN CN201910285459.3A patent/CN110046661A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106780524A (en) * | 2016-11-11 | 2017-05-31 | 厦门大学 | A kind of three-dimensional point cloud road boundary extraction method |
| CN107292276A (en) * | 2017-06-28 | 2017-10-24 | 武汉大学 | A kind of vehicle-mounted cloud clustering method and system |
Non-Patent Citations (1)
| Title |
|---|
| 郑特 等: "基于图割算法的摄影测量点云面向对象分类方法", 《测绘工程》 * |
Cited By (9)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111815776A (en) * | 2020-02-04 | 2020-10-23 | 山东水利技师学院 | Three-dimensional building fine geometric reconstruction method integrating airborne and vehicle-mounted three-dimensional laser point clouds and streetscape images |
| CN111582352A (en) * | 2020-04-30 | 2020-08-25 | 上海高仙自动化科技发展有限公司 | Object-based sensing method and device, robot and storage medium |
| CN111582352B (en) * | 2020-04-30 | 2023-06-27 | 上海高仙自动化科技发展有限公司 | Object-based perception method, object-based perception device, robot and storage medium |
| KR20210043518A (en) * | 2020-06-28 | 2021-04-21 | 베이징 바이두 넷컴 사이언스 테크놀로지 컴퍼니 리미티드 | High-precision mapping method and device |
| US20210405200A1 (en) * | 2020-06-28 | 2021-12-30 | Beijing Baidu Netcome Science Technology Co. Ltd. | High-Precision Mapping Method And Device |
| US11668831B2 (en) * | 2020-06-28 | 2023-06-06 | Beijing Baidu Netcom Science Technology Co., Ltd. | High-precision mapping method and device |
| KR102548282B1 (en) * | 2020-06-28 | 2023-06-26 | 베이징 바이두 넷컴 사이언스 테크놀로지 컴퍼니 리미티드 | High-precision mapping method and device |
| CN112070769A (en) * | 2020-09-18 | 2020-12-11 | 福州大学 | Layered point cloud segmentation method based on DBSCAN |
| CN112233237A (en) * | 2020-10-23 | 2021-01-15 | 广州建通测绘地理信息技术股份有限公司 | Water area leveling processing method and computer equipment for manufacturing digital elevation model |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110046661A (en) | A kind of vehicle-mounted cloud clustering method cutting algorithm based on contextual feature and figure | |
| Liu et al. | Cylinder detection in large-scale point cloud of pipeline plant | |
| CN112070769B (en) | Layered point cloud segmentation method based on DBSCAN | |
| Ji et al. | A novel simplification method for 3D geometric point cloud based on the importance of point | |
| Zhao et al. | Scene understanding in a large dynamic environment through a laser-based sensing | |
| US20040249809A1 (en) | Methods, systems, and data structures for performing searches on three dimensional objects | |
| CN106599915B (en) | A kind of vehicle-mounted laser point cloud classifications method | |
| CN110390683A (en) | A kind of Old City Wall three-dimensional cracking detection method based on point off density cloud | |
| Poux et al. | Self-learning ontology for instance segmentation of 3d indoor point cloud | |
| Xiang et al. | Segmentation-based classification for 3D point clouds in the road environment | |
| Wei et al. | Automatic coarse registration of point clouds using plane contour shape descriptor and topological graph voting | |
| CN115661374B (en) | Rapid retrieval method based on space division and model voxelization | |
| Tian et al. | 3D object recognition method with multiple feature extraction from LiDAR point clouds | |
| CN110780276A (en) | Tray identification method and system based on laser radar and electronic equipment | |
| Kukolj et al. | Road edge detection based on combined deep learning and spatial statistics of LiDAR data | |
| CN114332172A (en) | Improved laser point cloud registration method based on covariance matrix | |
| Boerner et al. | Voxel based segmentation of large airborne topobathymetric lidar data | |
| Sun et al. | Automated segmentation of LiDAR point clouds for building rooftop extraction | |
| Liu et al. | M: N Object matching on multiscale datasets based on MBR combinatorial optimization algorithm and spatial district | |
| CN117253205A (en) | Road surface point cloud rapid extraction method based on mobile measurement system | |
| Lin et al. | Minimum spanning tree based spatial outlier mining and its applications | |
| Yu et al. | Bidirectionally greedy framework for unsupervised 3D building extraction from airborne-based 3D meshes | |
| CN114758323A (en) | Urban road sign extraction method based on vehicle-mounted laser point cloud | |
| Sellamani et al. | PCS: prominent cross-sections for mesh models | |
| Yuan et al. | Automatic cables segmentation from a substation device based on 3D point cloud |
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 | ||
| RJ01 | Rejection of invention patent application after publication |
Application publication date: 20190723 |
|
| RJ01 | Rejection of invention patent application after publication |