CN108665491A - A kind of quick point cloud registration method based on local reference - Google Patents
A kind of quick point cloud registration method based on local reference Download PDFInfo
- Publication number
- CN108665491A CN108665491A CN201810241936.1A CN201810241936A CN108665491A CN 108665491 A CN108665491 A CN 108665491A CN 201810241936 A CN201810241936 A CN 201810241936A CN 108665491 A CN108665491 A CN 108665491A
- Authority
- CN
- China
- Prior art keywords
- point
- cloud
- angle
- local
- neighborhood
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/30—Determination of transform parameters for the alignment of images, i.e. image registration
- G06T7/33—Determination of transform parameters for the alignment of images, i.e. image registration using feature-based methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T2207/00—Indexing scheme for image analysis or image enhancement
- G06T2207/10—Image acquisition modality
- G06T2207/10028—Range image; Depth image; 3D point clouds
Abstract
The invention belongs to three-dimensional reconstruction fields, disclose a kind of quick point cloud registration method based on local reference, it is down-sampled to the original point cloud progress of input to obtain its corresponding sparse cloud, local angle feature as a reference point of getting is put with each in sparse cloud, kd tree are established with the local angle feature at a visual angle, search for the arest neighbors feature of each local angle feature of another visual angle, the k1 of Euclidean range error minimum is to as seed matching pair in arest neighbors feature, from its neighborhood search this refer under other matchings pair, then adjacent view transformation matrix is solved to complete the rough registration of a cloud using rigid body translation, rough registration result is optimized using partly overlapping ICP registration Algorithms.The present invention establishes initial matching using quick angle feature, is then propagated using matched neighborhood, improves point cloud registering speed while also assuring the precision of registration.
Description
Technical field
The invention belongs to three-dimensional reconstruction field more particularly to a kind of quick point cloud registering sides based on local reference
Method.
Background technology
Currently, the prior art commonly used in the trade is such：Vision is the most important mode in the human perception world, and video
Just play critical role in the various applications with visual correlation.In three-dimensional reconstruction algorithm based on RGB-D images still
Be so according to the movable information of depth camera restore target object structure, i.e., restore structure from movement, the technology be based on
The three-dimensional reconstruction algorithm of image sequence is identical, but the mode of the movable information of two kinds of algorithm for reconstructing estimation cameras has very big difference
Different, the target at each visual angle during camera motion is calculated according to the texture information of acquisition and depth information first in the former
Point cloud information, then according to the transformation matrix of the point cloud estimation camera at obtained each visual angle, which is also referred to as point cloud registering,
To restore the complete point cloud of target.And the latter is then first to analyze characteristics of image and extract characteristic point present in image, then
The characteristic point between adjacent two frame is matched, the transformation matrix of camera is then estimated according to matching double points, after obtaining camera pose
Will matching in information MAP to three dimensions, to obtain the point cloud information of target.Most for both the above method for reconstructing
Latter step is all to carry out resurfacing to the point cloud of acquisition to obtain reconstruction target to the end.Three-dimensional reconstruction based on RGB-D images
Point cloud registration algorithm in algorithm can be divided into following three kinds according to the condition difference in registration process, i.e., pure coordinate matching joins
Close the matching of coordinate and texture information, the matching of analysis site cloud structure feature.In pure coordinate matching, each visual angle point is used only
The coordinate information of cloud does not need the registration that a cloud can be completed in other auxiliary informations.The most classical algorithm of such registration Algorithm is
It is iteration closest approach algorithm.The algorithm assumes the movement very little between the corresponding depth camera of two clouds to be registered, i.e., two
There is very big Duplication between a cloud, Kd-tree then is established according to its coordinate to one of point cloud, then at this
The closest approach each put in another point cloud is searched in Kd-tree, is constantly iterated and is approached very according to obtained closest approach relationship
Real camera transformation matrix, to obtain the registration result of two clouds.Such algorithm requires the corresponding phase in two neighboring visual angle
Machine movement is minimum, and the Duplication between two clouds to be registered wants high in addition, and otherwise the algorithm, which is iterated, is unable to get one
A convergence is as a result, cause final point cloud registering to fail.The matching algorithm of co-ordinates and texture information is only calculated in pure coordinate matching
The texture information that each pair of point is answered in a cloud is added in method, i.e., when building Kd-tree by pervious three-dimensional extended to the four-dimension,
The point cloud registering condition at two visual angles is further reinforced.But it remains that two clouds Duplication to be registered is high
The corresponding camera motion in two visual angles cannot be excessive simultaneously, and otherwise last registration can also fail.Of analysis site cloud structure feature
It is to establish relevant description information according to existing certain denominators between two clouds with algorithm, is then believed according to description
The matching pair between two clouds is calculated in breath, final to obtain then according to matching to the transformation matrix of estimation camera
Registration result.The algorithm is established description to each of cloud point and is taken considerable time.
In conclusion problem of the existing technology is：Existing three kinds of point cloud registration algorithms movement is small, Duplication is high；
It needs to establish description to each putting in cloud, take larger.
Solve the difficulty and meaning of above-mentioned technical problem：The time for reducing registration while ensureing point cloud registering precision is multiple
Miscellaneous degree is the difficult point of current research.Reduce point cloud registering time complexity can make three-dimensional reconstruction system be calculated as can in real time
Energy.
Invention content
In view of the problems of the existing technology, the present invention provides a kind of quick point cloud registering side based on local reference
Method.
The invention is realized in this way a kind of quick point cloud registration method based on local reference, described based on part
The quick point cloud registration method of reference point is down-sampled to the original point cloud progress of input to obtain its corresponding sparse cloud；With sparse
Point cloud in each point it is as a reference point establish local angle feature and multiple dimensioned local approach vector description respectively, obtain local folder
After corner characteristics, kd-tree is established with the local angle feature at a visual angle；Each office of another visual angle is searched in kd-tree
The arest neighbors feature of portion's angle feature, the k1 of Euclidean range error minimum is to as seed matching pair in arest neighbors feature, then
Respectively using the seed match to as with reference to match, from its neighborhood search this refer under other matchings pair, it is to be matched to number
It stops search when reaching specified threshold, solves adjacent view transformation matrix using rigid body translation and match to complete the thick of a cloud
Standard optimizes rough registration result using partly overlapping ICP registration Algorithms.
Further, the quick point cloud registration method based on local reference includes the following steps：
Step 1 carries out down-sampled obtaining corresponding sparse cloud Pd and Qd to being originally inputted cloud P, a Q；
Step 2, to as a reference point each to be put in sparse cloud, in the r1 neighborhoods in its corresponding original point cloud
Its k nearest and farthest point is found, then respectively using the angle that closest approach, reference point, farthest point are constituted as feature；
Step 3, it is as a reference point each to be put in sparse cloud, in L neighborhood of its corresponding original point cloud, root
Approximate local approach vector description is obtained according to the data dependence put in reference point and neighborhood；
Step 4 establishes kd-tree with the angle feature each put in Pd, is then searched in the kd-tree every in Qd
The arest neighbors of a angle feature finally only chooses the k1 of Euclidean error minimum to as seed matching pair；
Step 5 obtains the angle between other points and the normal vector of seed point in Pd, then in seed neighborhood of a point,
Other match points are found with angle constraint, and rough registration is completed according to match point；
Step 6 optimizes the rough registration result use that step 5 the obtains ICP algorithm that partly overlaps.
Further, in the step 1 down-sampled specifically include is carried out to being originally inputted a cloud：It is empty that cloud will be put using grid
Between be divided into several small cubes, the point put in cloud will be respectively fallen in the form of cluster in corresponding cube grid,
If entirely point cloud is divided into and does gathering, which is indicated with the coordinate mean value of all the points in each gathering, by with
The upper obtained point cloud that handles is sparse cloud after original point cloud is down-sampled.
Further, the computational methods of local angle feature include in the step 2：Made with each of sparse cloud point
K nearest and farthest point is found in r1 neighborhoods in its corresponding original point cloud for reference point, and this is a with farthest k recently
What point was also ranked up according to the distance away from reference point, it is arranged from small to large according to distance；Current reference point is x, and its
Corresponding k closest approach and farthest point are expressed as X={ x1..., x2k, then its vector constituted between reference point indicates
For：
Wherein x1For the closest approach in nearest k point, and x2kFor the farthest point in farthest k point；
So angle feature calculation expression formula is：
Wherein calculate every subscript in angle formulae indicate its in vector in which.
Further, multiple dimensioned local approach vector calculation includes in the step 3：It calculates separately for each neighbour
Difference in point and sparse cloud in the radius of domain between current reference point is expressed as：
Wherein l=1 ..., L, while Sl={ xi|||xi-x||≤rl, for the correlation matrix C under each scalei, make
With singular value decomposition, three eigenvalue λs are obtainedl1≥λl2≥λl3And corresponding feature vector nl1, nl2, nl3, selection minimum
Eigenvalue λl3Corresponding nl3The layout description vectors small as each scale；Combine under multiple scales and is expressed as rectangular
Formula：
N=(n1..., nL)；
Wherein n1And nLThe partial descriptions vector of each scale is corresponded to respectively.
Further, the choosing method of initial seed matching pair specifically includes in the step 4：To the folder each put in Pd
Corner characteristics establish kd-tree；The corresponding arest neighbors feature of each angle feature in Qd, the measurement side of arest neighbors are searched in the tree
Formula is the Euclidean distance error between two angle features；The k1 for choosing Euclidean distance error minimum is used as initially matching double points
Seed matches, and is denoted as S.
Further, the matched neighborhood transmission method of initial seed includes in the step 5：(xi, yj) it is seed set of matches S
In a matching, corresponding set of matches is initialized as M={ (xi, yj), it is found away from x in PdiK2 nearest point,
The point set that k2 point is constituted is expressed as P 'd, and calculate this k2 middle-range xiFarthest distance d1；With yjIt is as a reference point in Qd
In find point in its d1 neighborhood, the point set that the point in this neighborhood is constituted is expressed as Q 'd.For arbitrary x ∈ Pd′\xiAnd d=| |
x-xi| |, define xiAngle between the vector description N of the point x in its d neighborhood is θ, is expressed as θ=(θ1..., θL)T,
Middle θlIt is in rlX under neighborhoodiAngle between the vector description of x indicates such as following formula：
In Q 'dMiddle yjD ' neighborhoods in search for xiD neighborhoods in all the points corresponding points, be expressed as：
y∈Qd' ∩ y | | and d '-d | ＜ ∈1}；
Wherein d '=| | y-yj| | and θlIt is similar, y and yjBetween angle between vector be expressed as θ '=(θ '1..., θ
′L)T, similarly θ 'lIt is in rlY under neighborhoodjAngle between the vector description of y, is expressed as following formula：
Then by xiD neighborhoods in put vector between angle theta and yjD ' neighborhoods in put vector between angle theta '
Between singularity be defined as follows：
When carrying out seed matching neighborhood propagation, to possessing the point of minimum angle singularity in the matched neighborhood of seed to making
It is concentrated for new matching to being added to initial matching, initial matching collection M=M ∪ (x, y), the angle singularity of all the points in neighborhood
It is all Inf to be all, then M is remained unchanged；As reference point xiD neighborhoods in all the points all traversed, a number of pairs reaches pre-
If value Nlim, then search stops, the matching otherwise added using last time is searched for next time as new reference point；Just
There is n matching pair in the matching of beginning seed, it is M to obtain set of matches1..., Mn, can be converted in the hope of one according to each set of matches
Matrix can complete point cloud registering according to transformation matrix, choose a minimum transformation of registration error to be originally inputted a cloud into
Row transformation.
Further, the ICP algorithm that partly overlaps in the step 6 includes to rough registration result optimizing method：Use arest neighbors
It searches for and is finding the closest approach that each pair of point answers in the point cloud by rough registration, k3 point pair of arest neighbors error minimum
Mean square error is as iteration initial error and by this k3 point to the initial matching pair as iteration；According to matching to solving two
Transformation matrix between point cloud, and one of visual angle is converted, reuse two of nearest neighbor search after the conversion
Arest neighbors matching is found in point cloud, while calculating the mean square error of k3 point pair of arest neighbors error minimum, if adjacent change twice
The error in generation is less than threshold value, and current square mean error amount is less than error threshold or reaches maximum iteration, iteration stopping；It is on the contrary
Iteration process is until iteration stopping.
Another object of the present invention is to provide the quick point cloud registration methods based on local reference described in a kind of application.
In conclusion advantages of the present invention and good effect are：
Registration of the present invention suitable for point cloud Duplication when low, has improvement to the registration effect of multiple point cloud models.Cause
This present invention is the registration Algorithm that a kind of computation complexity is low and applicability is high.
Description of the drawings
Fig. 1 is the quick point cloud registration method flow chart provided in an embodiment of the present invention based on local reference.
Fig. 2 is provided in an embodiment of the present invention to bun model measurement figures.
Fig. 3 is provided in an embodiment of the present invention to Arm model measurement figures.
Fig. 4 is provided in an embodiment of the present invention to dragon model measurement figures.
Fig. 5 is provided in an embodiment of the present invention to happy model measurement figures.
Specific implementation mode
In order to make the purpose , technical scheme and advantage of the present invention be clearer, with reference to embodiments, to the present invention
It is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to
Limit the present invention.
The present invention provides a kind of quick point cloud registration method based on local reference, it is intended to solve existing algorithm satisfaction and match
Under quasi- precise manner, the high problem of time complexity.
As shown in Figure 1, the quick point cloud registration method provided in an embodiment of the present invention based on local reference includes following
Step：
S101：It carries out down-sampled obtaining corresponding sparse cloud Pd and Qd to being originally inputted cloud P, a Q；
S102：It is as a reference point each to be put in sparse cloud, it is found in the r1 neighborhoods in its corresponding original point cloud
Its k nearest and farthest point, then respectively using the angle that closest approach, reference point, farthest point are constituted as feature；
S103：It is as a reference point each to be put in sparse cloud, according to ginseng in L neighborhood of its corresponding original point cloud
Examination point obtains approximate local approach vector description with the data dependence put in neighborhood；
S104：Kd-tree is established with the angle feature each put in Pd, is then searched in the kd-tree each in Qd
The arest neighbors of angle feature finally only chooses the k1 of Euclidean error minimum to as seed matching pair；
S105：The angle between other points and the normal vector of seed point in Pd is obtained first, then in seed neighborhood of a point
In, other match points are found with angle constraint, and rough registration is completed according to match point；
S106：The rough registration result use that S105 the is obtained ICP algorithm that partly overlaps is optimized.
In embodiments of the present invention, down-sampled specially to being originally inputted a cloud progress in step S101：It will using grid
Point cloud space is divided into several small cubes, thus the point in entire point cloud will be respectively fallen in the form of cluster it is corresponding
It is then equal with the coordinate of all the points in each gathering if final entire point cloud, which is divided into, is done gathering in cube grid
Value indicates the gathering, by the point cloud handled above be sparse cloud after original point cloud is down-sampled.Use the party
Formula greatly maintains the structure of origin cloud.
In the embodiment of the present invention, the calculating process of local angle feature is：First using each of sparse cloud point as
Reference point finds k nearest and farthest point, and this k nearest and farthest point in its corresponding original point cloud in r1 neighborhoods
Also it is ranked up according to the distance away from reference point, i.e., is arranged from small to large according to distance.If current reference point is
X, and its corresponding k closest approach and farthest point can be expressed as X={ x1..., x2k, then it is constituted between reference point
Vector can be expressed as：
Wherein x1For the closest approach in nearest k point, and x2kFor the farthest point in farthest k point；
So angle feature calculation expression formula is：
Wherein calculate every subscript in angle formulae indicate its in vector in which.Such as Pl 1It indicates to be located at PlIn
First item, and Pl 2kIt indicates to be located at PlIn 2k, that is, last.
In the embodiment of the present invention, multiple dimensioned part normal vector calculating process is in step S103：Calculate separately first for
Difference in point and sparse cloud in each radius of neighbourhood between current reference point, physical relationship can be expressed as：
Wherein l=1 ..., L, while Sl={ xi|||xi-x||≤rl), for the correlation matrix C under each scalei, make
With singular value decomposition, three eigenvalue λs can be obtainedl1≥λl2≥λl3And corresponding feature vector nl1, nl2, nl3, only select
Select minimal eigenvalue λl3Corresponding nl3The layout description vectors small as each scale.Therefore combine under multiple scales be expressed as
Lower matrix form：
N=(n1..., nL) (4)
Wherein n1And nLThe partial descriptions vector of each scale is corresponded to respectively.
In the embodiment of the present invention, the selection process of initial seed matching pair is as follows in step 104：To the folder each put in Pd
Corner characteristics establish kd-tree, and the corresponding arest neighbors feature of each angle feature in Qd is then searched in the tree, the arest neighbors
Metric form is the Euclidean distance error between two angle features.Then the k1 of Euclidean distance error minimum is only chosen to matching
Point is denoted as S to being matched as initial seed.Specific k1 parameter selections can be configured according to actual needs.
In the embodiment of the present invention, the matched neighborhood communication process of initial seed is as follows in S105：Assuming that (xi, yj) it is seed
A matching in set of matches S, corresponding set of matches are initialized as M={ (xi, yj), it is found first away from x in PdiMost
K2 close point, the point set that this k2 point is constituted can be expressed as P 'd, and calculate this k2 middle-range xiFarthest distance d1, connects
It with yjThe point as a reference point found in Qd in its d1 neighborhood, the point set that the point in this neighborhood is constituted can be expressed as Q 'd。
For arbitrary x ∈ Pd′\xiAnd d=| | x-xi| |, define xiAngle between the vector description N of the point x in its d neighborhood is θ,
It can be expressed as θ=(θ1..., θL)T, wherein θlIt is in rlX under neighborhoodiAngle between the vector description of x indicates such as
Following formula：
Then in Q 'dMiddle yjD ' neighborhoods in search for xiD neighborhoods in all the points corresponding points, these point can be expressed as：
y∈Qd' ∩ y | | and d '-d | ＜ ∈1} (6)
Wherein d '=| | y-yj| | and θlIt is similar, y and yjBetween angle between vector can be expressed as θ '=
(θ′1..., θ 'L)T, similarly θ 'lIt is in rlY under neighborhoodjAngle between the vector description of y can be expressed as following formula：
Then by xiD neighborhoods in put vector between angle theta and yjD ' neighborhoods in put vector between angle theta '
Between singularity be defined as follows：
When carrying out seed matching neighborhood propagation, to possessing the point of minimum angle singularity in the matched neighborhood of seed to making
Be that new matching is concentrated to being added to initial matching, initial matching collection M=M ∪ (x, y) at this time, if in neighborhood all the points angle
Singularity is all to be all Inf, then M is remained unchanged.As reference point xiD neighborhoods in all the points all traversed, if at this time match
Preset value N is reached to numberlim, then search stops, the matching otherwise added using last time carries out down as new reference point
Primary search.There is n matching pair in being matched due to initial seed, it is M that may finally obtain set of matches1..., Mn, according to each
Then a set of matches can complete point cloud registering in the hope of a transformation matrix according to transformation matrix, choose registration error most
A small transformation is converted to being originally inputted a cloud.
During present example is implemented, the ICP algorithm that partly overlaps is as follows to the rough registration result optimizing process in step 5：It is first
The closest approach that each pair of point is answered first is found in the point cloud in process rough registration using nearest neighbor search, then arest neighbors error
The mean square error of k3 minimum point pair is as iteration initial error and by this k3 point to the initial matching pair as iteration.So
Afterwards according to the matching to the transformation matrix between solving two clouds, and one of visual angle is converted, is reused most
Arest neighbors matching is found in two clouds of neighbor search after the conversion, while calculating k3 point pair of arest neighbors error minimum
Mean square error, if the error of adjacent iteration twice is less than threshold value, current square mean error amount is less than error threshold or reaches maximum
Iterations, iteration stopping.Otherwise iteration process is until iteration stopping.Iteration error threshold value therein and iterations threshold
Value can require to be configured according to specific precision and time complexity.
The application effect of the present invention is explained in detail with reference to experiment.
Point cloud registering performance can be promoted in order to illustrate the present invention while being applicable in the registration of various point cloud models.To a variety of moulds
Type is tested, and experimental data provides to each model under the cloud miss rate of difference point cloud registering performance test such as by 1~table of table 4
Fig. 2~Fig. 5,
1 bun model test results of table
2 Arm model test results of table
3 dragon model test results of table
4 happy model test results of table
Time, Rot and Tran in aforementioned four table characterize the performance of point cloud registering, three smaller present invention of parameter value
Performance it is better, from above-mentioned four tables result it is found that the time complexity of the present invention is low with respect to the other three algorithm, at the same with
The decline of quasi- precision can be ignored for entirely converting.Fig. 2~Fig. 5 tests each model under different miss rates, test
The result shows that algorithm is registrated better performances.
The foregoing is merely illustrative of the preferred embodiments of the present invention, is not intended to limit the invention, all essences in the present invention
All any modification, equivalent and improvement etc., should all be included in the protection scope of the present invention made by within refreshing and principle.
Claims (9)
1. a kind of quick point cloud registration method based on local reference, which is characterized in that described based on the fast of local reference
Fast point cloud registration method is down-sampled to the original point cloud progress of input to obtain its corresponding sparse cloud；With each in sparse cloud
A point is as a reference point to establish local angle feature and multiple dimensioned local approach vector description respectively, after obtaining local angle feature,
Kd-tree is established with the local angle feature at a visual angle；Each local angle feature of another visual angle is searched in kd-tree
Arest neighbors feature, the k1 of Euclidean range error minimum is to as seed matching pair in arest neighbors feature, then respectively with this kind
Son matching to as with reference to match, from its neighborhood search this refer under other matchings pair, it is to be matched that specified threshold is reached to number
It stops search when value, solves adjacent view transformation matrix to complete the rough registration of a cloud using rigid body translation, use part
The ICP registration Algorithms of overlapping optimize rough registration result.
2. the quick point cloud registration method based on local reference as described in claim 1, which is characterized in that described to be based on office
The quick point cloud registration method of portion's reference point includes the following steps：
Step 1 carries out down-sampled obtaining corresponding sparse cloud Pd and Qd to being originally inputted cloud P, a Q；
Step 2, it is as a reference point each to be put in sparse cloud, find it in the r1 neighborhoods in its corresponding original point cloud
K nearest and farthest point, then respectively using the angle that closest approach, reference point, farthest point are constituted as feature；
Step 3, it is as a reference point each to be put in sparse cloud, in L neighborhood of its corresponding original point cloud, according to ginseng
Examination point obtains approximate local approach vector description with the data dependence put in neighborhood；
Step 4 establishes kd-tree with the angle feature each put in Pd, then searches in Qd in the kd-tree and each presss from both sides
The arest neighbors of corner characteristics finally only chooses the k1 of Euclidean error minimum to as seed matching pair；
Step 5 obtains the angle between other points and the normal vector of seed point in Pd, then in seed neighborhood of a point, with folder
Other match points are found in angle constraint, and complete rough registration according to match point；
Step 6 optimizes the rough registration result use that step 5 the obtains ICP algorithm that partly overlaps.
3. the quick point cloud registration method based on local reference as claimed in claim 2, which is characterized in that the step 1
In carry out down-sampled specifically include to being originally inputted cloud：A cloud space is divided into several small cubes using grid, puts cloud
In point will be respectively fallen in the form of cluster in corresponding cube grid, if entirely point cloud is divided into and does cluster
Collection, the gathering is indicated with the coordinate mean value of all the points in each gathering, is original point by the point cloud handled above
Sparse cloud after cloud is down-sampled.
4. the quick point cloud registration method based on local reference as claimed in claim 2, which is characterized in that the step 2
The computational methods of middle local angle feature include：It is as a reference point in its corresponding original point cloud with each point in sparse cloud
K nearest and farthest point is found in middle r1 neighborhoods, and this is also carried out according to the distance away from reference point with k farthest point recently
Sequence, it is arranged from small to large according to distance；Current reference point is x, and its corresponding k closest approach and farthest point indicate
For X={ x1,...,x2k, then its vector constituted between reference point is expressed as：
Wherein x1For the closest approach in nearest k point, and x2kFor the farthest point in farthest k point；
So angle feature calculation expression formula is：
Wherein calculate every subscript in angle formulae indicate its in vector in which.
5. the quick point cloud registration method based on local reference as claimed in claim 2, which is characterized in that the step 3
In multiple dimensioned local approach vector calculation include：Calculate separately in each radius of neighbourhood point in sparse cloud when
Difference between preceding reference point is expressed as：
Wherein l=1 ..., L, while Sl={ xi|||xi-x||≤rl, for the correlation matrix C under each scalei, using strange
Different value is decomposed, and three eigenvalue λs are obtainedl1≥λl2≥λl3And corresponding feature vector nl1,nl2,nl3, select minimal characteristic
Value λl3Corresponding nl3The layout description vectors small as each scale；Combine under multiple scales and is expressed as matrix form：
N=(n1,...,nL)；
Wherein n1And nLThe partial descriptions vector of each scale is corresponded to respectively.
6. the quick point cloud registration method based on local reference as claimed in claim 2, which is characterized in that the step 4
The choosing method of middle initial seed matching pair specifically includes：Kd-tree is established to the angle feature each put in Pd；In the tree
Each the corresponding arest neighbors feature of angle feature, the metric form of arest neighbors are the Euclidean between two angle features in search Qd
Range error；The k1 for choosing Euclidean distance error minimum matches matching double points as initial seed, is denoted as S.
7. the quick point cloud registration method based on local reference as claimed in claim 2, which is characterized in that the step 5
The middle matched neighborhood transmission method of initial seed includes：(xi,yj) matched for one in seed set of matches S, corresponding
It is initialized as M={ (x with collectioni,yj), it is found away from x in PdiK2 nearest point, the point set that k2 point is constituted are expressed as P 'd,
And calculate this k2 middle-range xiFarthest distance d1；With yjThe point as a reference point found in Qd in its d1 neighborhood, this neighborhood
The point set that interior point is constituted is expressed as Q 'd；For arbitrary x ∈ Pd′\xiAnd d=| | x-xi| |, define xiWith the point x in its d neighborhood
Vector description N between angle be θ, be expressed as θ=(θ1,...,θL)T, wherein θlIt is in rlX under neighborhoodiIt is retouched with the vector of x
Angle between stating indicates such as following formula：
In Q 'dMiddle yjD ' neighborhoods in search for xiD neighborhoods in all the points corresponding points, be expressed as：
y∈Qd' ∩ y | | and d '-d | ＜ ∈1}；
Wherein d '=| | y-yj| | and θlIt is similar, y and yjBetween angle between vector be expressed as θ '=(θ '1,...,θ'L)T,
Similarly θ 'lIt is in rlY under neighborhoodjAngle between the vector description of y, is expressed as following formula：
Then by xiD neighborhoods in put vector between angle theta and yjD ' neighborhoods in put vector between angle theta ' between
Singularity be defined as follows：
When carrying out seed matching neighborhood propagation, to possessing the point of minimum angle singularity in the matched neighborhood of seed to as new
Matching concentrated to being added to initial matching, initial matching collection M=M ∪ (x, y), the angle singularity of all the points is all in neighborhood
All it is Inf, then M is remained unchanged；As reference point xiD neighborhoods in all the points all traversed, a number of pairs reaches preset value
Nlim, then search stops, the matching otherwise added using last time is searched for next time as new reference point；Initial kind
There is n matching pair in son matching, it is M to obtain set of matches1,...,Mn, can be in the hope of a transformation square according to each set of matches
Battle array can complete point cloud registering according to transformation matrix, choose a minimum transformation of registration error and carried out to being originally inputted a cloud
Transformation.
8. the quick point cloud registration method based on local reference as claimed in claim 2, which is characterized in that the step 6
In the ICP algorithm that partly overlaps include to rough registration result optimizing method：Using nearest neighbor search in the point cloud by rough registration
In find the closest approach that each pair of point is answered, using the mean square error of k3 point pair of arest neighbors error minimum as iteration initial error
And by this k3 point to the initial matching pair as iteration；According to matching to the transformation matrix between solving two clouds, and will
One of visual angle is converted, and is reused and is found arest neighbors matching in two clouds of nearest neighbor search after the conversion, together
When calculate arest neighbors error minimum k3 point pair mean square error, it is currently equal if the error of adjacent iteration twice is less than threshold value
Square error amount is less than error threshold or reaches maximum iteration, iteration stopping；Otherwise iteration process is until iteration is stopped
Only.
9. a kind of quick point cloud registration method using based on local reference described in claim 1~8 any one.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810241936.1A CN108665491B (en) | 2018-03-22 | 2018-03-22 | Rapid point cloud registration method based on local reference points |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810241936.1A CN108665491B (en) | 2018-03-22 | 2018-03-22 | Rapid point cloud registration method based on local reference points |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN108665491A true CN108665491A (en) | 2018-10-16 |
| CN108665491B CN108665491B (en) | 2022-04-12 |
Family
ID=63782241
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810241936.1A Active CN108665491B (en) | 2018-03-22 | 2018-03-22 | Rapid point cloud registration method based on local reference points |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN108665491B (en) |
Cited By (12)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109859256A (en) * | 2019-03-13 | 2019-06-07 | 大连理工大学 | A kind of three-dimensional point cloud method for registering based on automatic corresponding point matching |
| CN109887009A (en) * | 2019-01-24 | 2019-06-14 | 西北大学 | A kind of point cloud local matching process |
| CN109887012A (en) * | 2019-01-09 | 2019-06-14 | 天津大学 | A kind of point cloud registration method of combining adaptive search point set |
| CN109919984A (en) * | 2019-04-15 | 2019-06-21 | 武汉惟景三维科技有限公司 | A kind of point cloud autoegistration method based on local feature description's |
| CN111192300A (en) * | 2019-12-30 | 2020-05-22 | 南京航空航天大学 | Sparse mark point data registration method and device for local deformation measurement |
| CN111223132A (en) * | 2019-12-25 | 2020-06-02 | 华东师范大学 | Object registration method and system |
| CN112102376A (en) * | 2020-08-04 | 2020-12-18 | 广东工业大学 | Multi-view point cloud registration method and device for mixed sparse ICP (inductively coupled plasma) and storage medium |
| CN112381862A (en) * | 2020-10-27 | 2021-02-19 | 新拓三维技术(深圳)有限公司 | Full-automatic registration method and device for CAD (computer-aided design) model and triangular mesh |
| CN112861674A (en) * | 2021-01-28 | 2021-05-28 | 中振同辂(江苏)机器人有限公司 | Point cloud optimization method based on ground features and computer readable storage medium |
| CN113223062A (en) * | 2021-06-04 | 2021-08-06 | 武汉工控仪器仪表有限公司 | Point cloud registration method based on angular point feature point selection and quick descriptor |
| CN113239500A (en) * | 2021-07-12 | 2021-08-10 | 四川大学 | Reference point neighborhood feature matching method based on covariance matrix |
| CN113706587A (en) * | 2021-07-14 | 2021-11-26 | 西安交通大学 | Rapid point cloud registration method, device and equipment based on space grid division |
Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006065399A (en) * | 2004-08-24 | 2006-03-09 | Sony Corp | Image processing device and method, storage medium and program |
| CN102799763A (en) * | 2012-06-20 | 2012-11-28 | 北京航空航天大学 | Point cloud posture standardization-based method for extracting linear characteristic of point cloud |
| CN103236064A (en) * | 2013-05-06 | 2013-08-07 | 东南大学 | Point cloud automatic registration method based on normal vector |
| CN104112289A (en) * | 2014-01-29 | 2014-10-22 | 辽宁师范大学 | Three-dimensional object point cloud registration method based on parallel cascaded EM-ICP |
| CN104134216A (en) * | 2014-07-29 | 2014-11-05 | 武汉大学 | Laser point cloud auto-registration method and system based on 16-dimension feature description |
| CN104299260A (en) * | 2014-09-10 | 2015-01-21 | 西南交通大学 | Contact network three-dimensional reconstruction method based on SIFT and LBP point cloud registration |
| CN104463826A (en) * | 2013-09-18 | 2015-03-25 | 镇江福人网络科技有限公司 | Novel point cloud parallel Softassign registering algorithm |
| CN104794678A (en) * | 2015-05-04 | 2015-07-22 | 福建师范大学 | Automatic registration method for high-spatial-resolution remote-sensing images based on SIFI feature points |
| CN104952107A (en) * | 2015-05-18 | 2015-09-30 | 湖南桥康智能科技有限公司 | Three-dimensional bridge reconstruction method based on vehicle-mounted LiDAR point cloud data |
| CN105118059A (en) * | 2015-08-19 | 2015-12-02 | 哈尔滨工程大学 | Multi-scale coordinate axis angle feature point cloud fast registration method |
| US20180003825A1 (en) * | 2016-06-30 | 2018-01-04 | Topcon Corporation | Laser Scanner System And Registration Method Of Point Cloud Data |
-
2018
- 2018-03-22 CN CN201810241936.1A patent/CN108665491B/en active Active
Patent Citations (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2006065399A (en) * | 2004-08-24 | 2006-03-09 | Sony Corp | Image processing device and method, storage medium and program |
| CN102799763A (en) * | 2012-06-20 | 2012-11-28 | 北京航空航天大学 | Point cloud posture standardization-based method for extracting linear characteristic of point cloud |
| CN103236064A (en) * | 2013-05-06 | 2013-08-07 | 东南大学 | Point cloud automatic registration method based on normal vector |
| CN104463826A (en) * | 2013-09-18 | 2015-03-25 | 镇江福人网络科技有限公司 | Novel point cloud parallel Softassign registering algorithm |
| CN104112289A (en) * | 2014-01-29 | 2014-10-22 | 辽宁师范大学 | Three-dimensional object point cloud registration method based on parallel cascaded EM-ICP |
| CN104134216A (en) * | 2014-07-29 | 2014-11-05 | 武汉大学 | Laser point cloud auto-registration method and system based on 16-dimension feature description |
| CN104299260A (en) * | 2014-09-10 | 2015-01-21 | 西南交通大学 | Contact network three-dimensional reconstruction method based on SIFT and LBP point cloud registration |
| CN104794678A (en) * | 2015-05-04 | 2015-07-22 | 福建师范大学 | Automatic registration method for high-spatial-resolution remote-sensing images based on SIFI feature points |
| CN104952107A (en) * | 2015-05-18 | 2015-09-30 | 湖南桥康智能科技有限公司 | Three-dimensional bridge reconstruction method based on vehicle-mounted LiDAR point cloud data |
| CN105118059A (en) * | 2015-08-19 | 2015-12-02 | 哈尔滨工程大学 | Multi-scale coordinate axis angle feature point cloud fast registration method |
| US20180003825A1 (en) * | 2016-06-30 | 2018-01-04 | Topcon Corporation | Laser Scanner System And Registration Method Of Point Cloud Data |
Non-Patent Citations (6)
| Title |
|---|
| DMITRY CHETVERIKOV 等: "Robust Euclidean alignment of 3D point sets", 《IMAGE AND VISION COMPUTING》 * |
| HUAN LEI 等: "Fast Descriptors and Correspondence Propagation for Robust Global Point Cloud Registration", 《IEEE TRANSACTIONS ON IMAGE PROCESSING》 * |
| YING HE 等: "An Iterative Closest Points Algorithm for Registration of 3D Laser Scanner Point Clouds with Geometric Features", 《SENSORS 2017,》 * |
| 吕所军 等: "基于双向随机 KD 树的点云配准", 《工业控制计算机》 * |
| 孙军华 等: "基于分层块状全局搜索的三维点云自动配准", 《光学精密工程》 * |
| 陈学伟 等: "基于FPFH特征的点云配准技术", 《电脑知识与技术》 * |
Cited By (19)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109887012A (en) * | 2019-01-09 | 2019-06-14 | 天津大学 | A kind of point cloud registration method of combining adaptive search point set |
| CN109887012B (en) * | 2019-01-09 | 2023-03-31 | 天津大学 | Point cloud registration method combined with self-adaptive search point set |
| CN109887009A (en) * | 2019-01-24 | 2019-06-14 | 西北大学 | A kind of point cloud local matching process |
| CN109887009B (en) * | 2019-01-24 | 2022-12-09 | 西北大学 | Point cloud local matching method |
| CN109859256A (en) * | 2019-03-13 | 2019-06-07 | 大连理工大学 | A kind of three-dimensional point cloud method for registering based on automatic corresponding point matching |
| CN109859256B (en) * | 2019-03-13 | 2023-03-31 | 大连理工大学 | Three-dimensional point cloud registration method based on automatic corresponding point matching |
| CN109919984A (en) * | 2019-04-15 | 2019-06-21 | 武汉惟景三维科技有限公司 | A kind of point cloud autoegistration method based on local feature description's |
| CN111223132A (en) * | 2019-12-25 | 2020-06-02 | 华东师范大学 | Object registration method and system |
| CN111192300B (en) * | 2019-12-30 | 2023-11-07 | 南京航空航天大学 | Sparse marker point data registration method and device for local deformation measurement |
| CN111192300A (en) * | 2019-12-30 | 2020-05-22 | 南京航空航天大学 | Sparse mark point data registration method and device for local deformation measurement |
| CN112102376B (en) * | 2020-08-04 | 2023-06-06 | 广东工业大学 | Multi-view cloud registration method, device and storage medium of hybrid sparse ICP |
| CN112102376A (en) * | 2020-08-04 | 2020-12-18 | 广东工业大学 | Multi-view point cloud registration method and device for mixed sparse ICP (inductively coupled plasma) and storage medium |
| CN112381862A (en) * | 2020-10-27 | 2021-02-19 | 新拓三维技术(深圳)有限公司 | Full-automatic registration method and device for CAD (computer-aided design) model and triangular mesh |
| CN112861674A (en) * | 2021-01-28 | 2021-05-28 | 中振同辂(江苏)机器人有限公司 | Point cloud optimization method based on ground features and computer readable storage medium |
| CN113223062A (en) * | 2021-06-04 | 2021-08-06 | 武汉工控仪器仪表有限公司 | Point cloud registration method based on angular point feature point selection and quick descriptor |
| CN113239500B (en) * | 2021-07-12 | 2021-09-21 | 四川大学 | Reference point neighborhood feature matching method based on covariance matrix |
| CN113239500A (en) * | 2021-07-12 | 2021-08-10 | 四川大学 | Reference point neighborhood feature matching method based on covariance matrix |
| CN113706587B (en) * | 2021-07-14 | 2022-12-09 | 西安交通大学 | Rapid point cloud registration method, device and equipment based on space grid division |
| CN113706587A (en) * | 2021-07-14 | 2021-11-26 | 西安交通大学 | Rapid point cloud registration method, device and equipment based on space grid division |
Also Published As
| Publication number | Publication date |
|---|---|
| CN108665491B (en) | 2022-04-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108665491A (en) | A kind of quick point cloud registration method based on local reference | |
| CN108090960B (en) | A kind of Object reconstruction method based on geometrical constraint | |
| CN107886529B (en) | Point cloud registration method for three-dimensional reconstruction | |
| CN111968217B (en) | SMPL parameter prediction and human body model generation method based on picture | |
| CN106023298B (en) | Point cloud Rigid Registration method based on local Poisson curve reestablishing | |
| CN106780751A (en) | Three-dimensional point cloud method for reconstructing based on improved shielding Poisson algorithm | |
| CN105654483B (en) | The full-automatic method for registering of three-dimensional point cloud | |
| CN109325993B (en) | Saliency feature enhanced sampling method based on class octree index | |
| WO2015188445A1 (en) | Point cloud three-dimensional model reconstruction method and system | |
| CN111524170B (en) | Pulmonary CT image registration method based on unsupervised deep learning | |
| CN111027140B (en) | Airplane standard part model rapid reconstruction method based on multi-view point cloud data | |
| CN112381862B (en) | Full-automatic registration method and device for CAD (computer-aided design) model and triangular mesh | |
| CN110232698A (en) | One kind is based on model segmentation and L1The successive frame 3 D human body curve skeleton extracting method of intermediate value skeleton | |
| CN109242769B (en) | A kind of image processing method and device | |
| CN109754459A (en) | It is a kind of for constructing the method and system of human 3d model | |
| CN106980812A (en) | Three-dimensional face features' independent positioning method based on concatenated convolutional neutral net | |
| CN114972366A (en) | Full-automatic segmentation method and system for cerebral cortex surface based on graph network | |
| CN109584355A (en) | Threedimensional model fast reconstructing method based on mobile phone GPU | |
| CN112862949A (en) | Object 3D shape reconstruction method based on multiple views | |
| CN113538689A (en) | Three-dimensional model mesh simplification method based on feature fusion of neural network | |
| CN113327276A (en) | Method for registering point cloud data of large volume facing mobile measurement | |
| CN109920050A (en) | A kind of single-view three-dimensional flame method for reconstructing based on deep learning and thin plate spline | |
| CN113593037A (en) | Building method and application of Delaunay triangulated surface reconstruction model | |
| CN110956601B (en) | Infrared image fusion method and device based on multi-sensor mode coefficients and computer readable storage medium | |
| CN117274339A (en) | Point cloud registration method based on improved ISS-3DSC characteristics combined with ICP |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |