CN113838109A - Low-coincidence point cloud registration method - Google Patents

Low-coincidence point cloud registration method Download PDF

Info

Publication number
CN113838109A
CN113838109A CN202111437345.XA CN202111437345A CN113838109A CN 113838109 A CN113838109 A CN 113838109A CN 202111437345 A CN202111437345 A CN 202111437345A CN 113838109 A CN113838109 A CN 113838109A
Authority
CN
China
Prior art keywords
point
point cloud
information
convolution
feature
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Granted
Application number
CN202111437345.XA
Other languages
Chinese (zh)
Other versions
CN113838109B (en
Inventor
高庆
关海宁
吕金虎
张鹏
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Beihang University
Original Assignee
Beihang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Beihang University filed Critical Beihang University
Priority to CN202111437345.XA priority Critical patent/CN113838109B/en
Publication of CN113838109A publication Critical patent/CN113838109A/en
Application granted granted Critical
Publication of CN113838109B publication Critical patent/CN113838109B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/30—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
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T17/00—Three dimensional [3D] modelling, e.g. data description of 3D objects
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00—Image analysis
    • G06T7/10—Segmentation; Edge detection
    • G06T7/11—Region-based segmentation
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00—Indexing scheme for image analysis or image enhancement
    • G06T2207/10—Image acquisition modality
    • G06T2207/10028—Range image; Depth image; 3D point clouds

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Software Systems (AREA)
  • General Health & Medical Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Biophysics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Mathematical Physics (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computer Graphics (AREA)
  • Geometry (AREA)
  • Health & Medical Sciences (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses a low-coincidence point cloud registration method which can complete a point cloud registration task in a low-coincidence scene. Aiming at the problem that homonymous point pairs are difficult to search in a low-coincidence degree scene, an automatic attention mechanism is adopted to enable the integral point cloud of the aggregation point pair to achieve integral perception, meanwhile, an intersection attention mechanism is utilized to explicitly mine information of an overlapping region, confidence degrees of all points in the point cloud in the overlapping region are predicted, the point pairs in a matching stage are sampled in the overlapping region through probability selection, and the recall rate of registration is improved. Meanwhile, the perception field of the convolution kernel is dynamically limited in the overlapping region, so that the extraction of invalid geometric neighborhood information is avoided, and the precision and the accuracy of point-by-point characteristics are improved.

Description

Low-coincidence point cloud registration method
Technical Field
The invention belongs to the technical field of artificial intelligence, control science and engineering, and particularly relates to a low-coincidence point cloud registration method.
Background
The single photon imaging is a novel technology which adopts a pulse laser light source and utilizes nonlinear optical technologies such as wavelength division multiplexing and the like to realize ultrahigh precision and strong anti-noise imaging. Compared with the traditional laser radar system, the single photon radar system has remarkable advantages under the influence of long-distance and severe environment noise, and has huge application potential in the fields of weak signal detection, long-distance imaging, precise measurement and the like.
In three-dimensional imaging, in order to obtain a complete three-dimensional model of a real-world object or scene, a single photon radar system is generally required to perform multiple point cloud data acquisition on a target object point cloud at different spatial angles to obtain multiple point cloud segments partially overlapped in different spatial coordinate systems, and then complete three-dimensional reconstruction is realized through a point cloud registration technology.
The point cloud registration method can be divided into a method based on a feature matching formula and a method based on an optimization iterative formula from different optimization angles, and the effect of the method based on the feature matching formula is better than that of the iterative formula method from the viewpoint of generalization performance aiming at the point cloud registration problem with different characteristics. The way of extracting features is to use spatial statistical information such as normal vector information to construct manual features based on statistical histograms at first, and in recent years, with the development of deep learning, the way of generating deep features with rotation invariance based on deep learning gradually evolves. After dense feature extraction, the point cloud selects accurate corresponding relation of homonymous point pairs in a high-dimensional feature space through a robust error point pair removing method RANSAC, and then restores rotation and translation pose information among point cloud segments through methods such as Singular Value Decomposition (SVD).
Currently, a commonly used point cloud registration algorithm generally requires a high coincidence degree between point cloud segments shot from multiple angles to provide sufficient point pair matching information. At present, the registration rate of a registration algorithm data set based on deep learning is generally higher than 30%, and under the high-overlapping-degree scene, the registration recall rate of more than 90% is achieved by a feature matching algorithm and a global optimization algorithm.
However, due to the limited depth imaging range and the large point cloud density of the single photon radar, in many unconventional imaging tasks, multi-point cloud segments with multiple angles and high coincidence degree are difficult to obtain. The point cloud registration problem under the low-coincidence condition (the coincidence ratio is less than 30%) is still not well solved because the point cloud total number of the homonymous point pairs is low, the homonymous point pairs with enough number cannot be obtained by conventional random sampling or Furthest Point Sampling (FPS), and a large error is generated in the matching stage.
Disclosure of Invention
In order to solve the defects of the prior art, the invention provides a low-coincidence point cloud registration method, which is oriented to a single-photon radar detection system, solves the problems of small quantity of point cloud homonymous point pairs and poor registration precision in a low-coincidence scene, explicitly excavates overlapping region information, and improves the feature matching precision and the registration recall rate of point cloud registration in the low-coincidence scene. The specific technical scheme of the invention is as follows:
a low-coincidence point cloud registration method comprises the following steps:
s1: data preprocessing:
source point cloud acquisition by using single photon radar detection systemPWith a target point cloudQThe high-density array point cloud data is characterized in that an Octree Octree is established based on the existing point cloud data, and an index mechanism of disordered point cloud is established to realize quick neighbor searching;
s2: extracting full convolution characteristics:
aiming at the point cloud with space sparse characteristic, aiming at the source point cloudPWith a target point cloudQCarrying out full convolution aggregation point by point to obtain feature aggregation point coordinates representing each neighborhood information and feature aggregation point high-dimensional feature vectors;
s3: constructing an attention mechanism module and endowing the sensing capability of the full convolution characteristic overlapping area;
s4: and reversely solving a rigid body transformation matrix T through a rigid body transformation estimation module to realize the three-dimensional reconstruction of the complete scene.
Further, the specific process of step S2 is as follows:
s2-1: modeling a point cloud registration problem asGiven a cloud of source points under different coordinate systems
Figure 313332DEST_PATH_IMAGE001
With a target point cloud
Figure 262834DEST_PATH_IMAGE002
Solving the rotation matrix
Figure 820854DEST_PATH_IMAGE003
And translation vector
Figure 271427DEST_PATH_IMAGE004
To minimize the point pair error:
Figure 168976DEST_PATH_IMAGE005
wherein the content of the first and second substances,
Figure 414013DEST_PATH_IMAGE006
representing a cloud of sourcesPWith a target point cloudQThe same-name point pair in (2) exists only inPAndQin the region of the overlap of (a) and (b),
Figure 724908DEST_PATH_IMAGE007
is a set of all rotation matrices around the origin in the euclidean space,
Figure 323380DEST_PATH_IMAGE008
represents a real number;
the degree of overlap between two relative point clouds is defined as:
Figure 465648DEST_PATH_IMAGE009
wherein the content of the first and second substances,
Figure 819269DEST_PATH_IMAGE010
as a cloud of sourcesPTo the target point cloudQThe true value of (a) is rotationally transformed,
Figure 555144DEST_PATH_IMAGE011
as a cloud of sourcesPTo the target point cloudQThe true-value translation transformation of (1),
Figure 583405DEST_PATH_IMAGE012
in order to be the distance threshold value,
Figure 252284DEST_PATH_IMAGE013
representing a nearest neighbor operation;
s2-2: extracting sparse full-volume aggregation joint characteristics;
replacing the traditional voxel grid convolution with a sparse full convolution, the sparse convolution is defined as:
Figure 980068DEST_PATH_IMAGE014
wherein the content of the first and second substances,urepresenting the location of the input,
Figure 62294DEST_PATH_IMAGE015
the representation of the complex field is represented by a complex field,
Figure 268147DEST_PATH_IMAGE016
for the input dimension of a full convolutional neural network,
Figure 57112DEST_PATH_IMAGE017
for the output dimension of a full convolutional neural network,
Figure 80431DEST_PATH_IMAGE018
is defined as relative to the inputuIs selected from the group consisting of the bias of (c),
Figure 259740DEST_PATH_IMAGE019
in order to output the sparse convolution,
Figure 66022DEST_PATH_IMAGE020
to correspond to
Figure 302968DEST_PATH_IMAGE021
Relative offset of the convolution outputiThe input of the position(s) to be determined,
Figure 841397DEST_PATH_IMAGE022
representing the parameters of the convolution kernel, input
Figure 665259DEST_PATH_IMAGE020
If and only if
Figure 9652DEST_PATH_IMAGE023
Temporal presence definition, Source Point cloudPWith a target point cloudQGenerating characteristic aggregation point coordinates representing each neighborhood information through hierarchical convolution and downsampling
Figure 976471DEST_PATH_IMAGE024
And feature aggregation point high-dimensional vector
Figure 810435DEST_PATH_IMAGE025
。
Further, the specific process of step S3 is as follows:
s3-1: based on the characteristic aggregation point high-dimensional vector, the DGCNN is adopted to enhance local information so as to enable the local information to be enhanced
Figure 495494DEST_PATH_IMAGE025
Energy source point cloudPWith a target point cloudQThe integral perception is achieved;
s3-2: feature aggregation point high-dimensional vector passing through self-attention module
Figure 643579DEST_PATH_IMAGE025
The method comprises the steps of dynamically detecting whether corresponding aggregation points with similar feature spaces exist in a target point cloud by adopting a cross attention module transducer, and explicitly mining the confidence degree of whether the aggregation points and the fields thereof exist in an overlapping region based on semantic information of sequence level
Figure 323959DEST_PATH_IMAGE026
；
S3-3: the overlapping confidence degree of the dynamic attention convolution network GAC and a single aggregation point and the neighborhood thereof is utilized, the convolution kernel is dynamically limited to the overlapping area obtained by prediction according to the confidence degree, and the occupation ratio of invalid features in a matching stage is reduced;
s3-4: feature aggregation point high-dimensional vector based on output
Figure 735349DEST_PATH_IMAGE025
High-dimensional information decoding is carried out by adopting hole sparse convolution, and reception field dynamic selection is carried out by utilizing a GAC network, so that effective information extraction is realized, and point-by-point characteristics are obtained
Figure 438863DEST_PATH_IMAGE027
Confidence with overlapping region
Figure 452955DEST_PATH_IMAGE028
。
Further, the step S3-1 includes:
convolving the DGCNN with a dynamic graph to aggregate point coordinates with features
Figure 128787DEST_PATH_IMAGE029
And its high-dimensional vector
Figure 242236DEST_PATH_IMAGE025
For inputting and enhancing the characteristic information of the aggregation point, the DGCNN core step edge convolution EdgeConv is defined as:
Figure 262407DEST_PATH_IMAGE030
wherein for the input
Figure 689978DEST_PATH_IMAGE031
Constructing a neighborhood set
Figure 282633DEST_PATH_IMAGE032
，
Figure 160459DEST_PATH_IMAGE033
，hBy feeling of multiple layersThe composition of the base is known, and the base is,
Figure 776248DEST_PATH_IMAGE031
which represents the global information, is,
Figure 132143DEST_PATH_IMAGE034
representing local neighborhood information, the significance of the dynamic map update is
Figure 579305DEST_PATH_IMAGE031
No longer limited to the Euclidean space, but extends to the feature space, so in the hierarchical edge convolution, each time the edge convolution is carried out,
Figure 237820DEST_PATH_IMAGE031
the space definition of the system is dynamically changed, and the feature neighborhood is reconstructed, so that the receptive field covers the whole point cloud, and meanwhile, the sparse characteristic is kept;
Figure 731118DEST_PATH_IMAGE035
representing the mutual information of each edge in the feature space structure,
Figure 562808DEST_PATH_IMAGE036
represents a multilayer sensing base structure;
Figure 802159DEST_PATH_IMAGE037
is composed of
Figure 517393DEST_PATH_IMAGE031
At the final output convolution result, RELU represents the ramp activation function.
Further, the specific process of step S3-2 is:
s3-2-1: processing a kernel model Transformer in NLP by adopting a natural language, modeling the point cloud into an information sequence with continuous characteristics, and aggregating the characteristic points into high-dimensional vectors
Figure 170091DEST_PATH_IMAGE025
Performing information interaction to mine spatial overlapped information, and a Transformer module coreThe cardiac operation is defined as:
Figure 743155DEST_PATH_IMAGE038
wherein the indexQueryKey valueKeyInformation, informationValueThe equivalence is regarded as a database searching operation, is embodied into a point cloud registration problem,Query，Key，Valuehigh-dimensional vector of feature aggregation point
Figure 227226DEST_PATH_IMAGE025
，
Figure 289860DEST_PATH_IMAGE039
Is the vector dimension of Query;
s3-2-2: target point cloudQTo source point cloudPThe information flow of (1), then Query is
Figure 367538DEST_PATH_IMAGE040
Key is
Figure 868926DEST_PATH_IMAGE041
The Transformer operation is regarded as firstly calculating cosine similarity in a feature space, taking an index vector Query and a Key value vector Key as dot products, and taking vector dimensions
Figure 145187DEST_PATH_IMAGE042
Making a scale parameter;
s3-2-3: normalizing the obtained similarity degree Value into probability Value distribution which belongs to 0-1 and is added to 1 through a softmax function, multiplying the probability Value distribution by information Value corresponding to the Key Value Key respectively to obtain cross Attention information Attention (Query, Key, Value) of the index Query and all the Key Value keys, namely the target point cloudQCloud to sourcePThe cross information stream of (2); and isPToQThe reverse information flow calculation process follows the same calculation rule;
s3-2-4: the overlapping area information is implicit in the cross attention information, so that the source point clouds are respectively alignedPWith a target point cloudQRespectively carries out a cross injectionCalculating the intention information flow:
Figure 581984DEST_PATH_IMAGE043
aggregating the obtained attention information and the original feature into a high-dimensional vector
Figure 6012DEST_PATH_IMAGE044
Splicing, cat [ | ] non-calculation in the above formula]Representing the vector splicing operation on the feature dimension, and outputting the obtained mixed features as the confidence coefficient of a point-by-point overlapping region of aggregation points through a multi-layer perceptron MLP
Figure 920879DEST_PATH_IMAGE045
。
Further, the specific process of step S3-3 is:
s3-3-1: the high-dimensional vectors of the feature aggregation points are obtained by network forward propagation
Figure 317225DEST_PATH_IMAGE046
And overlap region confidence
Figure 551022DEST_PATH_IMAGE045
The method adopts a dynamic attention network mechanism GAC to dynamically limit the receptive field of a convolution kernel in a predicted overlapping area, and reduces invalid geometric information of a non-overlapping area to extract pollution point-by-point characteristics;
s3-3-2: the polymerization point at this time
Figure 337713DEST_PATH_IMAGE047
Using undirected graphs
Figure 853008DEST_PATH_IMAGE048
It is shown that,
Figure 166177DEST_PATH_IMAGE049
respectively representing the vertices and non-directional edges of the graph structure,
Figure 944778DEST_PATH_IMAGE050
the number of undirected graph vertices; order to
Figure 608977DEST_PATH_IMAGE051
To represent
Figure 927963DEST_PATH_IMAGE052
Neighborhood vertex of
Figure 971005DEST_PATH_IMAGE053
The set of components is composed of a plurality of groups,
Figure 45141DEST_PATH_IMAGE052
representation versus undirected graphGThe vertex of (1);
s3-3-3: reduced symbols, use
Figure 603161DEST_PATH_IMAGE054
Representing a point cloudPOrQThe feature of (2) is aggregated into a high-dimensional vector,brepresenting a characteristic dimension of a current incoming GAC network, the GAC network and
Figure 929100DEST_PATH_IMAGE055
selecting the probability for the prior information, selecting the aggregation point in the overlapping area, and aggregating the coordinates and high-dimensional characteristics of the overlapping area again by using the graph convolution operation defined in the GAC, namely performing the graph convolution operation only in the overlapping area:
Figure 718327DEST_PATH_IMAGE056
wherein the weight is
Figure 838730DEST_PATH_IMAGE057
The calculation is as follows:
Figure 211942DEST_PATH_IMAGE058
wherein the content of the first and second substances,
Figure 75993DEST_PATH_IMAGE059
in order to be a multi-layer perceptron MLP,
Figure 218261DEST_PATH_IMAGE060
，
Figure 775144DEST_PATH_IMAGE061
on behalf of the operation of the network,
Figure 635653DEST_PATH_IMAGE062
is a central point
Figure 834553DEST_PATH_IMAGE063
Is determined by the point of the neighborhood of the point,
Figure 972274DEST_PATH_IMAGE064
and
Figure 60578DEST_PATH_IMAGE065
are respectively a central point
Figure 18169DEST_PATH_IMAGE066
And adjacent neighbor point
Figure 614236DEST_PATH_IMAGE067
The characteristics of the respective intermediate layer(s),
Figure 403200DEST_PATH_IMAGE068
is an offset; therefore, the graph convolution operation is performed by simultaneously utilizing Euclidean space coordinate information and feature space distance information in the neighborhood for aggregation, and the definition of the neighborhood of the feature space is different in each forward propagation, so that the dynamic limitation of the receptive field is realized.
Further, the specific process of step S3-4 is:
s3-4-1: will be provided with
Figure 36307DEST_PATH_IMAGE069
And
Figure 871408DEST_PATH_IMAGE070
performing dimensionality splicing operation to obtain aggregation point characteristicsA feature map with an overlap region confidence;
s3-4-2: utilizing sparse hole convolution to carry out up-sampling operation on the characteristic graph, wherein parameters, namely convolution kernel size and sliding step length, are selected and down-sampled consistently to recover and input point cloudP、QThe confidence of point-by-point overlapping region with the same output point cloud size and output simultaneously
Figure 615373DEST_PATH_IMAGE071
。
Further, the specific process of step S4 is as follows:
s4-1: selecting a certain number of candidate matching points by adopting a probability sorting method based on point-by-point overlapping confidence coefficients, and obtaining an accurate corresponding relation of the homonymous point pairs in a high-dimensional feature space by using a robust outlier rejection method RANSAC;
s4-2: based on the corresponding relation of the homonymous point pairs, Singular Value Decomposition (SVD) is utilized to reversely solve pose transformation information of the rotation matrix and the translation matrix to form a rigid transformation matrix T;
s4-3: and (3) utilizing the relative pose transformation information to act on the target point cloud, transforming the target point cloud to a source point cloud coordinate system, and then carrying out point cloud density equalization operation to obtain a complete three-dimensional model.
Further, in the step S3-4-2, the output is up-sampledP、QPoint by point characteristics
Figure 586740DEST_PATH_IMAGE027
Dimension is 32-dimensional compact features, and simultaneously, the confidence coefficient of the point-by-point overlapping region is output
Figure 656327DEST_PATH_IMAGE072
。
The invention has the beneficial effects that:
1. compared with the traditional local geometric feature descriptor extraction mode, the method adopts the full-convolution neural network architecture to extract the point cloud features, and realizes the real-time and rapid generation of the point cloud features.
2. According to the method, the overlapped region point cloud region is predicted by the aid of the Transformer module in an explicit mode, points in a subsequent matching stage are sampled in the overlapped region, and accuracy of point cloud registration in a low-overlap degree scene is greatly improved.
3. The method adopts the GAC module to limit the sparse convolution kernel receptive field in the overlapping region, and reduces the influence of the geometric information of the non-overlapping region on the feature descriptor.
Drawings
In order to illustrate embodiments of the present invention or technical solutions in the prior art more clearly, the drawings which are needed in the embodiments will be briefly described below, so that the features and advantages of the present invention can be understood more clearly by referring to the drawings, which are schematic and should not be construed as limiting the present invention in any way, and for a person skilled in the art, other drawings can be obtained on the basis of these drawings without any inventive effort. Wherein:
FIG. 1 is a process diagram of the self-attention mechanism of the present invention;
FIG. 2 is a cross-attention mechanism process diagram of the present invention;
FIG. 3 is a schematic diagram of the GAC network mechanism of the present invention;
FIG. 4 is a flow chart of a method of the present invention;
fig. 5 is a diagram illustrating feature extraction robustness.
Detailed Description
In order that the above objects, features and advantages of the present invention can be more clearly understood, a more particular description of the invention will be rendered by reference to the appended drawings. It should be noted that the embodiments of the present invention and features of the embodiments may be combined with each other without conflict.
In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present invention, however, the present invention may be practiced in other ways than those specifically described herein, and therefore the scope of the present invention is not limited by the specific embodiments disclosed below.
The invention particularly relates to a feature matching type low-overlap-ratio point cloud registration method based on a sparse full convolution neural network, which is used for solving the problem of low-overlap-ratio point cloud registration of an object level generated by scanning of a single photon radar system and can be widely applied to other low-overlap-ratio scenes such as indoor scene point clouds and outdoor scanning point clouds.
The method can complete the point cloud registration task in a low-coincidence degree scene. Aiming at the problem that homonymous point pairs are difficult to search in a low-coincidence degree scene, an automatic attention mechanism is adopted to enable the integral point cloud of the aggregation point pair to achieve integral perception, meanwhile, an intersection attention mechanism is utilized to explicitly mine information of an overlapping region, confidence degrees of all points in the point cloud in the overlapping region are predicted, the point pairs in a matching stage are sampled in the overlapping region through probability selection, and the recall rate of registration is improved. Meanwhile, the perception field of the convolution kernel is dynamically limited in the overlapping region, so that the extraction of invalid geometric neighborhood information is avoided, and the precision and the accuracy of point-by-point characteristics are improved.
Specifically, as shown in fig. 4, a low-coincidence point cloud registration method includes the following steps:
s1: data preprocessing:
source point cloud acquisition by using single photon radar detection systemPWith a target point cloudQThe high-density array point cloud data is characterized in that an Octree Octree is established based on the existing point cloud data, and an index mechanism of disordered point cloud is established to realize quick neighbor searching;
s2: extracting full convolution characteristics:
aiming at the point cloud with space sparse characteristic, aiming at the source point cloudPWith a target point cloudQCarrying out full convolution aggregation point by point to obtain feature aggregation point coordinates representing each neighborhood information and feature aggregation point high-dimensional feature vectors; the specific process is as follows:
s2-1: modeling point cloud registration problem as a source point cloud given under different coordinate systems
Figure 208750DEST_PATH_IMAGE001
With a target point cloud
Figure 287565DEST_PATH_IMAGE002
Solving the rotation matrix
Figure 113438DEST_PATH_IMAGE003
And translation vector
Figure 88347DEST_PATH_IMAGE004
To minimize the point pair error:
Figure 570144DEST_PATH_IMAGE005
wherein the content of the first and second substances,
Figure 46125DEST_PATH_IMAGE006
representing a cloud of sourcesPWith a target point cloudQThe same-name point pair in (2) exists only inPAndQin the region of the overlap of (a) and (b),
Figure 195347DEST_PATH_IMAGE007
is a set of all rotation matrices around the origin in the euclidean space,
Figure 341157DEST_PATH_IMAGE008
represents a real number;
the degree of overlap between two relative point clouds is defined as:
Figure 874032DEST_PATH_IMAGE009
wherein the content of the first and second substances,
Figure 29070DEST_PATH_IMAGE010
as a cloud of sourcesPTo the target point cloudQThe true value of (a) is rotationally transformed,
Figure 95115DEST_PATH_IMAGE011
as a cloud of sourcesPTo the target point cloudQThe true-value translation transformation of (1),
Figure 677406DEST_PATH_IMAGE012
in order to be the distance threshold value,
Figure 399374DEST_PATH_IMAGE013
representing a nearest neighbor operation;
s2-2: extracting sparse full-volume aggregation joint characteristics;
replacing the traditional voxel grid convolution with a sparse full convolution, the sparse convolution is defined as:
Figure 217157DEST_PATH_IMAGE014
wherein the content of the first and second substances,urepresenting the location of the input,
Figure 13075DEST_PATH_IMAGE015
the representation of the complex field is represented by a complex field,
Figure 563005DEST_PATH_IMAGE016
for the input dimension of a full convolutional neural network,
Figure 804893DEST_PATH_IMAGE017
for the output dimension of a full convolutional neural network,
Figure 36154DEST_PATH_IMAGE018
is defined as relative to the inputuIs selected from the group consisting of the bias of (c),
Figure 483316DEST_PATH_IMAGE019
in order to output the sparse convolution,
Figure 532043DEST_PATH_IMAGE020
to correspond to
Figure 635129DEST_PATH_IMAGE021
Relative offset of the convolution outputiThe input of the position(s) to be determined,
Figure 794714DEST_PATH_IMAGE022
representing the parameters of the convolution kernel, input
Figure 96383DEST_PATH_IMAGE020
If and only if
Figure 925799DEST_PATH_IMAGE023
Temporal presence definition, Source Point cloudPWith a target point cloudQDownsampling via hierarchical convolutionGenerating feature aggregation point coordinates representing each neighborhood information
Figure 640814DEST_PATH_IMAGE024
And feature aggregation point high-dimensional vector
Figure 276194DEST_PATH_IMAGE025
。
S3: constructing an attention mechanism module and endowing the sensing capability of the full convolution characteristic overlapping area; as shown in fig. 1-3, the specific process is as follows:
s3-1: based on the characteristic aggregation point high-dimensional vector, the DGCNN is adopted to enhance local information so as to enable the local information to be enhanced
Figure 901211DEST_PATH_IMAGE025
Energy source point cloudPWith a target point cloudQThe integral perception is achieved;
in some embodiments, the step S3-1 is specifically performed as follows:
convolving the DGCNN with a dynamic graph to aggregate point coordinates with features
Figure 527626DEST_PATH_IMAGE029
And its high-dimensional vector
Figure 402041DEST_PATH_IMAGE025
For inputting and enhancing the characteristic information of the aggregation point, the DGCNN core step edge convolution EdgeConv is defined as:
Figure 778796DEST_PATH_IMAGE030
wherein for the input
Figure 117374DEST_PATH_IMAGE031
Constructing a neighborhood set
Figure 554171DEST_PATH_IMAGE032
，
Figure 915882DEST_PATH_IMAGE033
，hIs composed of a plurality of layers of sensing bases,
Figure 955383DEST_PATH_IMAGE031
which represents the global information, is,
Figure 554991DEST_PATH_IMAGE034
representing local neighborhood information, the significance of the dynamic map update is
Figure 959428DEST_PATH_IMAGE031
No longer limited to the Euclidean space, but extends to the feature space, so in the hierarchical edge convolution, each time the edge convolution is carried out,
Figure 605173DEST_PATH_IMAGE031
the space definition of the system is dynamically changed, and the feature neighborhood is reconstructed, so that the receptive field covers the whole point cloud, and meanwhile, the sparse characteristic is kept;
Figure 386047DEST_PATH_IMAGE035
representing the mutual information of each edge in the feature space structure,
Figure 840162DEST_PATH_IMAGE036
represents a multilayer sensing base structure;
Figure 74939DEST_PATH_IMAGE037
is composed of
Figure 83346DEST_PATH_IMAGE031
At the final output convolution result, RELU represents the ramp activation function.
S3-2: feature aggregation point high-dimensional vector passing through self-attention module
Figure 402332DEST_PATH_IMAGE025
The method comprises the steps of dynamically detecting whether corresponding aggregation points with similar feature spaces exist in a target point cloud by adopting a cross attention module Transformer, and explicitly mining whether the aggregation points and the field thereof exist in the target point cloud or not based on semantic information of sequence levelDegree of confidence that there is an overlapping region
Figure 835587DEST_PATH_IMAGE026
；
In some embodiments, the specific process of step S3-2 is:
s3-2-1: processing a kernel model Transformer in NLP by adopting a natural language, modeling the point cloud into an information sequence with continuous characteristics, and aggregating the characteristic points into high-dimensional vectors
Figure 519510DEST_PATH_IMAGE025
And carrying out information interaction so as to mine spatial overlapping information, wherein the kernel operation of the Transformer module is defined as:
Figure 343109DEST_PATH_IMAGE038
the index Query, the Key Value Key and the information Value are equivalently regarded as database searching operation, and are embodied in the point cloud registration problem, and the Query, the Key Value Key and the Value are regarded as high-dimensional vectors of feature aggregation points
Figure 528103DEST_PATH_IMAGE025
，
Figure 691231DEST_PATH_IMAGE039
Is the vector dimension of Query;
s3-2-2: target point cloudQTo source point cloudPThe information flow of (1), then Query is
Figure 936267DEST_PATH_IMAGE040
Key is
Figure 981584DEST_PATH_IMAGE041
The Transformer operation is regarded as firstly calculating cosine similarity in a feature space, taking an index vector Query and a Key value vector Key as dot products, and taking vector dimensions
Figure 580055DEST_PATH_IMAGE042
Making a scale parameter;
s3-2-3: normalizing the obtained similarity degree Value into probability Value distribution which belongs to 0-1 and is added to 1 through a softmax function, multiplying the probability Value distribution by information Value corresponding to the Key Value Key respectively to obtain cross Attention information Attention (Query, Key, Value) of the index Query and all the Key Value keys, namely the target point cloudQCloud to sourcePThe cross information stream of (2); and isPToQThe reverse information flow calculation process follows the same calculation rule;
s3-2-4: the overlapping area information is implicit in the cross attention information, so that the source point clouds are respectively alignedPWith a target point cloudQThe aggregation points of (a) perform cross attention information flow calculation for one time respectively:
Figure 489368DEST_PATH_IMAGE043
aggregating the obtained attention information and the original feature into a high-dimensional vector
Figure 577410DEST_PATH_IMAGE044
Splicing, cat [ | ] non-calculation in the above formula]Representing the vector splicing operation on the feature dimension, and outputting the obtained mixed features as the confidence coefficient of a point-by-point overlapping region of aggregation points through a multi-layer perceptron MLP
Figure 47705DEST_PATH_IMAGE045
。
S3-3: the overlapping confidence degree of the dynamic attention convolution network GAC and a single aggregation point and the neighborhood thereof is utilized, the convolution kernel is dynamically limited to the overlapping area obtained by prediction according to the confidence degree, and the occupation ratio of invalid features in a matching stage is reduced;
the specific process of step S3-3 is:
s3-3-1: the high-dimensional vectors of the feature aggregation points are obtained by network forward propagation
Figure 840081DEST_PATH_IMAGE046
And overlap region confidence
Figure 774539DEST_PATH_IMAGE045
The method adopts a dynamic attention network mechanism GAC to dynamically limit the receptive field of a convolution kernel in a predicted overlapping area, and reduces invalid geometric information of a non-overlapping area to extract pollution point-by-point characteristics;
s3-3-2: the polymerization point at this time
Figure 236744DEST_PATH_IMAGE047
Using undirected graphs
Figure 318970DEST_PATH_IMAGE048
It is shown that,
Figure 790402DEST_PATH_IMAGE049
respectively representing the vertices and non-directional edges of the graph structure,
Figure 579367DEST_PATH_IMAGE050
the number of undirected graph vertices; order to
Figure 71528DEST_PATH_IMAGE051
To represent
Figure 781995DEST_PATH_IMAGE052
Neighborhood vertex of
Figure 588277DEST_PATH_IMAGE053
The set of components is composed of a plurality of groups,
Figure 61109DEST_PATH_IMAGE052
representation versus undirected graphGThe vertex of (1);
s3-3-3: reduced symbols, use
Figure 865117DEST_PATH_IMAGE054
Representing a point cloudPOrQThe feature of (2) is aggregated into a high-dimensional vector,brepresenting a characteristic dimension of a current incoming GAC network, the GAC network andbselecting probability for prior information, selecting aggregation point in the overlapping area, and aggregating the coordinates and high-dimensional features of the overlapping area again by using graph convolution operation defined in GAC (generalized open area code) to obtain the aggregation point only in the overlapping areaPerforming graph convolution operation on the domain:
Figure 921934DEST_PATH_IMAGE056
wherein the weight is
Figure 531907DEST_PATH_IMAGE057
The calculation is as follows:
Figure 967568DEST_PATH_IMAGE058
wherein the content of the first and second substances,
Figure 332690DEST_PATH_IMAGE059
in order to be a multi-layer perceptron MLP,
Figure 548908DEST_PATH_IMAGE060
，
Figure 900255DEST_PATH_IMAGE061
on behalf of the operation of the network,
Figure 580635DEST_PATH_IMAGE062
is a central point
Figure 992025DEST_PATH_IMAGE063
Is determined by the point of the neighborhood of the point,
Figure 695538DEST_PATH_IMAGE064
and
Figure 476675DEST_PATH_IMAGE065
are respectively a central point
Figure 886928DEST_PATH_IMAGE066
And adjacent neighbor point
Figure 265956DEST_PATH_IMAGE067
The characteristics of the respective intermediate layer(s),
Figure 784662DEST_PATH_IMAGE068
is an offset; therefore, the graph convolution operation is performed by simultaneously utilizing Euclidean space coordinate information and feature space distance information in the neighborhood for aggregation, and the definition of the neighborhood of the feature space is different in each forward propagation, so that the dynamic limitation of the receptive field is realized.
S3-4: feature aggregation point high-dimensional vector based on output
Figure 212233DEST_PATH_IMAGE069
High-dimensional information decoding is carried out by adopting hole sparse convolution, and reception field dynamic selection is carried out by utilizing a GAC network, so that effective information extraction is realized, and point-by-point characteristics are obtained
Figure 132784DEST_PATH_IMAGE073
Confidence with overlapping region
Figure 417135DEST_PATH_IMAGE074
。
In some embodiments, the specific process of step S3-4 is:
s3-4-1: will be provided with
Figure 298503DEST_PATH_IMAGE069
And
Figure 654398DEST_PATH_IMAGE070
performing dimension splicing operation to obtain a feature map of the polymerization point features and the confidence coefficient of the overlapping region;
s3-4-2: utilizing sparse hole convolution to carry out up-sampling operation on the characteristic graph, wherein parameters, namely convolution kernel size and sliding step length, are selected and down-sampled consistently to recover and input point cloudP、QThe confidence of point-by-point overlapping region with the same output point cloud size and output simultaneously
Figure 101560DEST_PATH_IMAGE071
。
In some embodiments, in step S3-4-2, the output is upsampledP、QPoint by point characteristics
Figure 25654DEST_PATH_IMAGE073
Dimension is 32-dimensional compact features, and simultaneously, the confidence coefficient of the point-by-point overlapping region is output
Figure 483399DEST_PATH_IMAGE071
。
S4: the method comprises the following steps of reversely solving a rigid body transformation matrix T through a rigid body transformation estimation module to realize the three-dimensional reconstruction of a complete scene, and specifically comprising the following steps:
s4-1: selecting a certain number of candidate matching points by adopting a probability sorting method based on point-by-point overlapping confidence coefficients, and obtaining an accurate corresponding relation of the homonymous point pairs in a high-dimensional feature space by using a robust outlier rejection method RANSAC;
s4-2: based on the corresponding relation of the homonymous point pairs, Singular Value Decomposition (SVD) is utilized to reversely solve pose transformation information of the rotation matrix and the translation matrix to form a rigid transformation matrix T;
s4-3: and (3) utilizing the relative pose transformation information to act on the target point cloud, transforming the target point cloud to a source point cloud coordinate system, and then carrying out point cloud density equalization operation to obtain a complete three-dimensional model.
The point cloud registration method with low overlap ratio is researched, and the point cloud registration method has a key effect on improving the integral three-dimensional imaging capability of the single photon radar and improving the application value of the single photon radar.
For the convenience of understanding the above technical aspects of the present invention, the following detailed description will be given of the above technical aspects of the present invention by way of specific examples.
Example 1
In order to verify the effectiveness of the invention, the invention is compared with the latest algorithm in the aspect of point cloud registration on a public data set through experimental tests, and the superiority of the invention in the actual application scene compared with the contemporary algorithm is verified.
Preparation of data set: the invention is also applicable to general point cloud registration under general contact ratio by verifying on an indoor data set 3DMatch commonly used by the current point cloud registration algorithm, wherein the 3DMatch contains point cloud data of 62 different indoor scenes in total, wherein 54 scenes are used as training sets, and 8 verification sets. The invention is verified on a low-coincidence scene data set 3DLoMatch to have superior performance in a low-coincidence scene.
Evaluation indexes are as follows: the invention belongs to a point cloud Registration algorithm based on a Feature Matching formula, and therefore, the point cloud Registration algorithm is mainly used for evaluating a Feature Matching Recall ratio (Feature Matching Recall) and a Registration success ratio (Registration success ratio), wherein the Feature Matching Recall ratio (Feature Matching Recall) is used for measuring the description capacity of a Feature extraction module for extracting features, and the Registration success ratio (Registration success ratio) is used for representing the proportion of logarithms successfully matched in Registration to the total number of point clouds. The two indexes have a positive correlation with each other, and the higher the recall rate of feature matching is, the higher the success rate of registration is.
The specific experimental process comprises the following steps:
(1) model training: the method of the invention is used for training on a training set of a public data set 3DMatch, an SGD optimizer is selected, the learning rate is set to be 5e-3, a TITAN X is selected as a GPU, 1.5h is needed for training one period, and 40 periods of integral model training can be converged. The back-end RANSAC is implemented using open3D (version 0.9.0).
(2) And (3) testing a model: and testing the model on a 3D match test set and a low-coincidence scene data set 3D LoMatch, respectively verifying the effectiveness of the model on the two data sets, and simultaneously performing an ablation experiment to verify the effectiveness of the method.
(3) The experimental results are as follows: the performance on the two evaluation indices is shown in the following table:
TABLE 13 model representation on DMatch
Figure 518351DEST_PATH_IMAGE075
TABLE 23 DLoMatch modeling Performance
Figure 413495DEST_PATH_IMAGE076
From the above experimental results, it can be seen that, no matter in the standard coincidence scene 3d match or in the low coincidence scene 3d lomatch data set, the method of the present invention obtains an improvement in the index compared with some mainstream feature matching point cloud registration algorithms at present, and especially in the low coincidence scene, obtains a large improvement under the condition of 1000 sampling points.
As shown in fig. 5, the left graph illustrates that the feature matching recall rate of the method of the present invention is most pronounced when the inner point distance threshold is increased to the same value as compared to other methods; in the right side diagram, three-dimensional feature matching is 3DMatch, a fast point feature histogram is FPFH, a rotation diagram feature is SpinImage, a direction histogram feature is SHOT, a core point convolution feature in an overlapping area is PREDATOR, a sparse full convolution feature FCGF, a core point convolution feature D3Feat, a compact geometric feature CGF, a point pair feature PPFNet and a three-dimensional smooth feature 3DSmoothNet, and it can be known from the diagram that when an internal point proportion threshold value is increased, the speed of curve reduction corresponding to the method is slowest, and further, the extracted point pair feature in the model has obvious robustness improvement compared with other methods;
the ablation experiment is performed on an overlap area prediction module (overlap assessment module) and a receptive field restriction module (receptive field restriction module) in the method of the present invention, and the results on the registration success rate are as follows:
TABLE 3 ablation experiment
Figure 305227DEST_PATH_IMAGE077
Ablation experiments show that the method has remarkable advantages particularly in registration in a low-coincidence-degree scene, and registration indexes in a common coincidence-degree scene are improved.
(4) And (4) experimental conclusion: table 1 shows that the present invention exhibits significant advantages in a low coincidence scene registration scenario, table 2 shows that the present invention can be improved in a standard coincidence scene, and is not limited to a low coincidence scene registration scenario, and table 3 shows that the present invention is effective for a module specifically designed for a low coincidence scene. The above conclusions illustrate the effectiveness of the present invention.
The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention, and various modifications and changes may be made by those skilled in the art. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims (9)

1. A low-coincidence point cloud registration method is characterized by comprising the following steps:
s1: data preprocessing:
source point cloud acquisition by using single photon radar detection systemPWith a target point cloudQThe high-density array point cloud data is characterized in that an Octree Octree is established based on the existing point cloud data, and an index mechanism of disordered point cloud is established to realize quick neighbor searching;
s2: extracting full convolution characteristics:
aiming at the point cloud with space sparse characteristic, aiming at the source point cloudPWith a target point cloudQCarrying out full convolution aggregation point by point to obtain feature aggregation point coordinates representing each neighborhood information and feature aggregation point high-dimensional feature vectors;
s3: constructing an attention mechanism module and endowing the sensing capability of the full convolution characteristic overlapping area;
s4: and reversely solving a rigid body transformation matrix T through a rigid body transformation estimation module to realize the three-dimensional reconstruction of the complete scene.
2. The low-coincidence point cloud registration method according to claim 1, wherein the specific process of the step S2 is as follows:
s2-1: modeling point cloud registration problem as a source point cloud given under different coordinate systems
Figure 416001DEST_PATH_IMAGE001
With a target point cloud
Figure 77052DEST_PATH_IMAGE002
Solving the rotation matrix
Figure 847562DEST_PATH_IMAGE003
And translation vector
Figure 801611DEST_PATH_IMAGE004
To minimize the point pair error:
Figure 391992DEST_PATH_IMAGE005
wherein the content of the first and second substances,
Figure 27373DEST_PATH_IMAGE006
representing a cloud of sourcesPWith a target point cloudQThe same-name point pair in (2) exists only inPAndQin the region of the overlap of (a) and (b),
Figure 511444DEST_PATH_IMAGE007
is a set of all rotation matrices around the origin in the euclidean space,
Figure 574078DEST_PATH_IMAGE008
represents a real number;
the degree of overlap between two relative point clouds is defined as:
Figure 651755DEST_PATH_IMAGE009
wherein the content of the first and second substances,
Figure 887565DEST_PATH_IMAGE010
as a cloud of sourcesPTo the target point cloudQThe true value of (a) is rotationally transformed,
Figure 367088DEST_PATH_IMAGE011
as a cloud of sourcesPTo the target point cloudQThe true-value translation transformation of (1),
Figure 600623DEST_PATH_IMAGE012
in order to be the distance threshold value,
Figure 809273DEST_PATH_IMAGE013
representing a nearest neighbor operation;
s2-2: extracting sparse full-volume aggregation joint characteristics;
replacing the traditional voxel grid convolution with a sparse full convolution, the sparse convolution is defined as:
Figure 724139DEST_PATH_IMAGE014
wherein the content of the first and second substances,urepresenting the location of the input,
Figure 854906DEST_PATH_IMAGE015
the representation of the complex field is represented by a complex field,
Figure 587239DEST_PATH_IMAGE016
for the input dimension of a full convolutional neural network,
Figure 373930DEST_PATH_IMAGE017
for the output dimension of a full convolutional neural network,
Figure 482700DEST_PATH_IMAGE018
is defined as relative to the inputuIs selected from the group consisting of the bias of (c),
Figure 467973DEST_PATH_IMAGE019
in order to output the sparse convolution,
Figure 246574DEST_PATH_IMAGE020
to correspond to
Figure 645194DEST_PATH_IMAGE021
Relative offset of the convolution outputiThe input of the position(s) to be determined,
Figure 964180DEST_PATH_IMAGE022
representing the parameters of the convolution kernel, input
Figure 272801DEST_PATH_IMAGE020
If and only if
Figure 848401DEST_PATH_IMAGE023
Temporal presence definition, Source Point cloudPWith a target point cloudQGenerating characteristic aggregation point coordinates representing each neighborhood information through hierarchical convolution and downsampling
Figure 609684DEST_PATH_IMAGE024
And feature aggregation point high-dimensional vector
Figure 466782DEST_PATH_IMAGE025
。
3. The low-coincidence point cloud registration method according to claim 1 or 2, wherein the specific process of the step S3 is as follows:
s3-1: based on the characteristic aggregation point high-dimensional vector, the DGCNN is adopted to enhance local information so as to enable the local information to be enhanced
Figure 754543DEST_PATH_IMAGE025
Energy source point cloudPWith a target point cloudQThe integral perception is achieved;
s3-2: feature aggregation point high-dimensional vector passing through self-attention module
Figure 140525DEST_PATH_IMAGE025
The method comprises the steps of dynamically detecting whether corresponding aggregation points with similar feature spaces exist in a target point cloud by adopting a cross attention module transducer, and explicitly mining the confidence degree of whether the aggregation points and the fields thereof exist in an overlapping region based on semantic information of sequence level
Figure 248159DEST_PATH_IMAGE026
；
S3-3: the overlapping confidence degree of the dynamic attention convolution network GAC and a single aggregation point and the neighborhood thereof is utilized, the convolution kernel is dynamically limited to the overlapping area obtained by prediction according to the confidence degree, and the occupation ratio of invalid features in a matching stage is reduced;
s3-4: feature aggregation point high-dimensional vector based on output
Figure 908947DEST_PATH_IMAGE025
High-dimensional information decoding is carried out by adopting hole sparse convolution, and reception field dynamic selection is carried out by utilizing a GAC network, so that effective information extraction is realized, and point-by-point characteristics are obtained
Figure 192161DEST_PATH_IMAGE027
Confidence with overlapping region
Figure 342520DEST_PATH_IMAGE028
。
4. The point cloud registration method with low coincidence degree according to claim 3, wherein the step S3-1 is implemented by:
convolving the DGCNN with a dynamic graph to aggregate point coordinates with features
Figure 78394DEST_PATH_IMAGE029
And its high-dimensional vector
Figure 372235DEST_PATH_IMAGE025
For inputting and enhancing the characteristic information of the aggregation point, the DGCNN core step edge convolution EdgeConv is defined as:
Figure 306693DEST_PATH_IMAGE030
wherein for the input
Figure 768898DEST_PATH_IMAGE031
Constructing a neighborhood set
Figure 851124DEST_PATH_IMAGE032
，
Figure 853715DEST_PATH_IMAGE033
，hIs composed of a plurality of layers of sensing bases,
Figure 845941DEST_PATH_IMAGE031
which represents the global information, is,
Figure 603682DEST_PATH_IMAGE034
representing local neighborhood information, the significance of the dynamic map update is
Figure 110887DEST_PATH_IMAGE031
No longer limited to the Euclidean space, but extends to the feature space, so in the hierarchical edge convolution, each time the edge convolution is carried out,
Figure 854852DEST_PATH_IMAGE031
the space definition of the system is dynamically changed, and the feature neighborhood is reconstructed, so that the receptive field covers the whole point cloud, and meanwhile, the sparse characteristic is kept;
Figure 91798DEST_PATH_IMAGE035
representing the mutual information of each edge in the feature space structure,
Figure 630227DEST_PATH_IMAGE036
represents a multilayer sensing base structure;
Figure 624728DEST_PATH_IMAGE037
is composed of
Figure 532903DEST_PATH_IMAGE031
At the final output convolution result, RELU represents the ramp activation function.
5. The low-coincidence point cloud registration method according to claim 3, wherein the specific process of the step S3-2 is as follows:
s3-2-1: processing a kernel model Transformer in NLP by adopting a natural language, modeling the point cloud into an information sequence with continuous characteristics, and aggregating the characteristic points into high-dimensional vectors
Figure 499722DEST_PATH_IMAGE025
And carrying out information interaction so as to mine spatial overlapping information, wherein the kernel operation of the Transformer module is defined as:
Figure 536948DEST_PATH_IMAGE038
wherein the indexQueryKey valueKeyInformation, informationValueThe equivalence is regarded as a database searching operation, is embodied into a point cloud registration problem,Query，Key，Valuehigh-dimensional vector of feature aggregation point
Figure 81062DEST_PATH_IMAGE025
，
Figure 166830DEST_PATH_IMAGE039
Is the vector dimension of Query;
s3-2-2: target point cloudQTo source point cloudPThe information flow of (1), then Query is
Figure 112789DEST_PATH_IMAGE040
Key is
Figure 320916DEST_PATH_IMAGE041
The Transformer operation is regarded as firstly calculating cosine similarity in a feature space, taking an index vector Query and a Key value vector Key as dot products, and taking vector dimensions
Figure 962113DEST_PATH_IMAGE042
Making a scale parameter;
s3-2-3: normalizing the obtained similarity degree value into an addition belonging to 0-1 through a softmax functionAnd probability Value distribution of 1, and multiplying with information Value corresponding to Key Value Key respectively to obtain cross Attention information Attention (Query, Key, Value) of index Query and all Key Value Key, namely target point cloudQCloud to sourcePThe cross information stream of (2); and isPToQThe reverse information flow calculation process follows the same calculation rule;
s3-2-4: the overlapping area information is implicit in the cross attention information, so that the source point clouds are respectively alignedPWith a target point cloudQThe aggregation points of (a) perform cross attention information flow calculation for one time respectively:
Figure 976206DEST_PATH_IMAGE043
aggregating the obtained attention information and the original feature into a high-dimensional vector
Figure 714354DEST_PATH_IMAGE044
Splicing, cat [ | ] non-calculation in the above formula]Representing the vector splicing operation on the feature dimension, and outputting the obtained mixed features as the confidence coefficient of a point-by-point overlapping region of aggregation points through a multi-layer perceptron MLP
Figure 31066DEST_PATH_IMAGE045
。
6. The low-coincidence point cloud registration method according to claim 3, wherein the specific process of the step S3-3 is as follows:
s3-3-1: the high-dimensional vectors of the feature aggregation points are obtained by network forward propagation
Figure 45378DEST_PATH_IMAGE046
And overlap region confidence
Figure 535265DEST_PATH_IMAGE047
The method adopts a dynamic attention network mechanism GAC to dynamically limit the receptive field of a convolution kernel in a predicted overlapping region, and reduces non-zero probabilityExtracting pollution point-by-point characteristics from invalid geometric information of the overlapping area;
s3-3-2: the polymerization point at this time
Figure 65603DEST_PATH_IMAGE048
Using undirected graphs
Figure 677850DEST_PATH_IMAGE049
It is shown that,
Figure 355956DEST_PATH_IMAGE050
respectively representing the vertices and non-directional edges of the graph structure,
Figure 852797DEST_PATH_IMAGE051
the number of undirected graph vertices; order to
Figure 362276DEST_PATH_IMAGE052
To represent
Figure 20790DEST_PATH_IMAGE053
Neighborhood vertex of
Figure 186192DEST_PATH_IMAGE054
The set of components is composed of a plurality of groups,
Figure 345778DEST_PATH_IMAGE053
representation versus undirected graphGThe vertex of (1);
s3-3-3: reduced symbols, use
Figure 850709DEST_PATH_IMAGE055
Representing a point cloudPOrQThe feature of (2) is aggregated into a high-dimensional vector,brepresenting a characteristic dimension of a current incoming GAC network, the GAC network and
Figure 476862DEST_PATH_IMAGE056
selecting probability for prior information, selecting aggregation point in overlapped region, and utilizing GACThe defined graph convolution operation aggregates the coordinates and high-dimensional features of the overlapping area again, namely the graph convolution operation is performed only on the overlapping area:
Figure 958921DEST_PATH_IMAGE057
wherein the weight is
Figure 531985DEST_PATH_IMAGE058
The calculation is as follows:
Figure 16056DEST_PATH_IMAGE059
wherein the content of the first and second substances,
Figure 78690DEST_PATH_IMAGE060
in order to be a multi-layer perceptron MLP,
Figure 156367DEST_PATH_IMAGE061
，
Figure 657756DEST_PATH_IMAGE062
on behalf of the operation of the network,
Figure 934016DEST_PATH_IMAGE063
is a central point
Figure 370814DEST_PATH_IMAGE064
Is determined by the point of the neighborhood of the point,
Figure 794842DEST_PATH_IMAGE065
and
Figure 975288DEST_PATH_IMAGE066
are respectively a central point
Figure 106055DEST_PATH_IMAGE067
And adjacent neighbor point
Figure 339852DEST_PATH_IMAGE068
The characteristics of the respective intermediate layer(s),
Figure 126543DEST_PATH_IMAGE069
is an offset; therefore, the graph convolution operation is performed by simultaneously utilizing Euclidean space coordinate information and feature space distance information in the neighborhood for aggregation, and the definition of the neighborhood of the feature space is different in each forward propagation, so that the dynamic limitation of the receptive field is realized.
7. The low-coincidence point cloud registration method of claim 3, wherein the specific process of the step S3-4 is as follows:
s3-4-1: will be provided with
Figure 641838DEST_PATH_IMAGE070
And
Figure 220586DEST_PATH_IMAGE071
performing dimension splicing operation to obtain a feature map of the polymerization point features and the confidence coefficient of the overlapping region;
s3-4-2: utilizing sparse hole convolution to carry out up-sampling operation on the characteristic graph, wherein parameters, namely convolution kernel size and sliding step length, are selected and down-sampled consistently to recover and input point cloudP、QThe confidence of point-by-point overlapping region with the same output point cloud size and output simultaneously
Figure 999187DEST_PATH_IMAGE072
。
8. The low-coincidence point cloud registration method according to claim 1 or 2, wherein the specific process of the step S4 is as follows:
s4-1: selecting a certain number of candidate matching points by adopting a probability sorting method based on point-by-point overlapping confidence coefficients, and obtaining an accurate corresponding relation of the homonymous point pairs in a high-dimensional feature space by using a robust outlier rejection method RANSAC;
s4-2: based on the corresponding relation of the homonymous point pairs, Singular Value Decomposition (SVD) is utilized to reversely solve pose transformation information of the rotation matrix and the translation matrix to form a rigid transformation matrix T;
s4-3: and (3) utilizing the relative pose transformation information to act on the target point cloud, transforming the target point cloud to a source point cloud coordinate system, and then carrying out point cloud density equalization operation to obtain a complete three-dimensional model.
9. The method of registering point clouds of low coincidence degree of claim 7, wherein in the step S3-4-2, the output is up-sampledP、QPoint by point characteristics
Figure 397807DEST_PATH_IMAGE027
Dimension is 32-dimensional compact features, and simultaneously, the confidence coefficient of the point-by-point overlapping region is output
Figure 716793DEST_PATH_IMAGE073
。
CN202111437345.XA 2021-11-30 2021-11-30 Low-coincidence point cloud registration method Active CN113838109B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202111437345.XA CN113838109B (en) 2021-11-30 2021-11-30 Low-coincidence point cloud registration method

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202111437345.XA CN113838109B (en) 2021-11-30 2021-11-30 Low-coincidence point cloud registration method

Publications (2)

Publication Number Publication Date
CN113838109A true CN113838109A (en) 2021-12-24
CN113838109B CN113838109B (en) 2022-02-15

Family

ID=78971944

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202111437345.XA Active CN113838109B (en) 2021-11-30 2021-11-30 Low-coincidence point cloud registration method

Country Status (1)

Country Link
CN (1) CN113838109B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114004871A (en) * 2022-01-04 2022-02-01 山东大学 Point cloud registration method and system based on point cloud completion
CN114937122A (en) * 2022-06-16 2022-08-23 黄冈强源电力设计有限公司 Rapid three-dimensional model reconstruction method for cement fiberboard house
CN115063459A (en) * 2022-08-09 2022-09-16 苏州立创致恒电子科技有限公司 Point cloud registration method and device and panoramic point cloud fusion method and system
CN115631221A (en) * 2022-11-30 2023-01-20 北京航空航天大学 Low-overlapping-degree point cloud registration method based on consistency sampling

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110634161A (en) * 2019-08-30 2019-12-31 哈尔滨工业大学（深圳） Method and device for quickly and accurately estimating pose of workpiece based on point cloud data
CN112150523A (en) * 2020-09-24 2020-12-29 中北大学 Three-dimensional point cloud registration method with low overlapping rate
CN113160293A (en) * 2021-05-13 2021-07-23 南京信息工程大学 Complex scene ground station point cloud automatic registration method based on feature probability

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110634161A (en) * 2019-08-30 2019-12-31 哈尔滨工业大学（深圳） Method and device for quickly and accurately estimating pose of workpiece based on point cloud data
CN112150523A (en) * 2020-09-24 2020-12-29 中北大学 Three-dimensional point cloud registration method with low overlapping rate
CN113160293A (en) * 2021-05-13 2021-07-23 南京信息工程大学 Complex scene ground station point cloud automatic registration method based on feature probability

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
ADRIEN GRESSIN 等: "Towards 3D lidar point cloud registration improvement using optimal neighborhood knowledge", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》 *
CONGCONG WEN 等: "Airborne LiDAR point cloud classification with global-local graph attention convolution neural network", 《ISPRS JOURNAL OF PHOTOGRAMMETRY AND REMOTE SENSING》 *
HAOZHE CHENG 等: "PTANet: Triple Attention Network for point cloud semantic segmentation", 《ENGINEERING APPLICATIONS OF ARTIFICIAL INTELLIGENCE》 *
李强: "基于多约束八叉树和多重特征的点云配准算法", 《中国优秀博硕士学位论文全文数据库(硕士) 信息科技辑(月刊)》 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114004871A (en) * 2022-01-04 2022-02-01 山东大学 Point cloud registration method and system based on point cloud completion
CN114937122A (en) * 2022-06-16 2022-08-23 黄冈强源电力设计有限公司 Rapid three-dimensional model reconstruction method for cement fiberboard house
CN115063459A (en) * 2022-08-09 2022-09-16 苏州立创致恒电子科技有限公司 Point cloud registration method and device and panoramic point cloud fusion method and system
CN115063459B (en) * 2022-08-09 2022-11-04 苏州立创致恒电子科技有限公司 Point cloud registration method and device and panoramic point cloud fusion method and system
CN115631221A (en) * 2022-11-30 2023-01-20 北京航空航天大学 Low-overlapping-degree point cloud registration method based on consistency sampling
CN115631221B (en) * 2022-11-30 2023-04-28 北京航空航天大学 Low-overlapping-degree point cloud registration method based on consistency sampling

Also Published As

Publication number Publication date
CN113838109B (en) 2022-02-15

Similar Documents

Publication Publication Date Title
CN113838109B (en) Low-coincidence point cloud registration method
Xie et al. Point clouds learning with attention-based graph convolution networks
Tong et al. Polynomial fitting algorithm based on neural network
Wu et al. Object-compositional neural implicit surfaces
Chen et al. HAPGN: Hierarchical attentive pooling graph network for point cloud segmentation
CN113191387A (en) Cultural relic fragment point cloud classification method combining unsupervised learning and data self-enhancement
Cui et al. Image steganography based on foreground object generation by generative adversarial networks in mobile edge computing with Internet of Things
CN115311502A (en) Remote sensing image small sample scene classification method based on multi-scale double-flow architecture
Zhan et al. FA-ResNet: Feature affine residual network for large-scale point cloud segmentation
Li et al. Multi-view-based siamese convolutional neural network for 3D object retrieval
CN115984339A (en) Double-pipeline point cloud completion method based on geometric feature refining and confrontation generation network
Min et al. Geometry guided network for point cloud registration
Tian et al. SAR object classification using the DAE with a modified triplet restriction
Zhou et al. Retrieval and localization with observation constraints
Li et al. A new algorithm of vehicle license plate location based on convolutional neural network
Stypułkowski et al. Representing point clouds with generative conditional invertible flow networks
Wang et al. Two-stage vSLAM loop closure detection based on sequence node matching and semi-semantic autoencoder
Zheng et al. Instance-Aware Spatial-Frequency Feature Fusion Detector for Oriented Object Detection in Remote Sensing Images
Song 3D virtual reality implementation of tourist attractions based on the deep belief neural network
Chen et al. 3D point cloud generation reconstruction from single image based on image retrieval
Yu et al. A deep neural network using double self-attention mechanism for als point cloud segmentation
Xiao et al. Multi-dimensional graph interactional network for progressive point cloud completion
Chen et al. Recognition and Classification of High Resolution Remote Sensing Image based on Convolutional Neural Network
Zao et al. Topology-Guided Road Graph Extraction From Remote Sensing Images
Wang et al. Quantitative Evaluation of Plant and Modern Urban Landscape Spatial Scale Based on Multiscale Convolutional Neural Network

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