CN117351052A - Point cloud fine registration method based on feature consistency and spatial consistency - Google Patents

Point cloud fine registration method based on feature consistency and spatial consistency Download PDF

Info

Publication number
CN117351052A
CN117351052A CN202311335751.4A CN202311335751A CN117351052A CN 117351052 A CN117351052 A CN 117351052A CN 202311335751 A CN202311335751 A CN 202311335751A CN 117351052 A CN117351052 A CN 117351052A
Authority
CN
China
Prior art keywords
consistency
feature
point
point cloud
space
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
CN202311335751.4A
Other languages
Chinese (zh)
Other versions
CN117351052B (en
Inventor
曾慧
王瑄
江左
杨清港
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.)
Aerospace Science And Industry Group Intelligent Technology Research Institute Co ltd
Shunde Innovation School of University of Science and Technology Beijing
Original Assignee
Aerospace Science And Industry Group Intelligent Technology Research Institute Co ltd
Shunde Innovation School of University of Science and Technology Beijing
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 Aerospace Science And Industry Group Intelligent Technology Research Institute Co ltd, Shunde Innovation School of University of Science and Technology Beijing filed Critical Aerospace Science And Industry Group Intelligent Technology Research Institute Co ltd
Priority to CN202311335751.4A priority Critical patent/CN117351052B/en
Publication of CN117351052A publication Critical patent/CN117351052A/en
Application granted granted Critical
Publication of CN117351052B publication Critical patent/CN117351052B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

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
    • 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
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • 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)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Computational Linguistics (AREA)
  • Data Mining & Analysis (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Health & Medical Sciences (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Image Analysis (AREA)

Abstract

The invention provides a point cloud fine registration method based on feature consistency and space consistency, and belongs to the technical field of computer vision. The method comprises the following steps: processing the point pairs after the point cloud rough registration to obtain feature consistency and space consistency; on the basis of the feature consistency and the space consistency, selecting seed points by using the confidence coefficient; defining a k nearest neighbor feature space, centering on a seed point, searching nearest neighbor points in the k nearest neighbor feature space to form a cluster, and fusing feature consistency and space consistency with non-local attention to obtain a point cloud feature updating formula so as to obtain point cloud features containing long-distance information after updating the points in the cluster; generating a probability matching matrix through the updated point cloud characteristics, and obtaining a rotation vector and a translation matrix through singular value decomposition calculation based on the probability matching matrix to realize point cloud registration. By adopting the method and the device, the accuracy and the speed of point cloud registration can be improved.

Description

Point cloud fine registration method based on feature consistency and spatial consistency
Technical Field
The invention relates to the technical field of computer vision, in particular to a point cloud fine registration method based on feature consistency and space consistency.
Background
Active point cloudTarget point cloud->Wherein the points of the source point cloud and the target point cloud are different, namely N x ≠N y . The point cloud registration is to find the coincident point, i.e. the inner point, of the source point cloud and the target point cloud, and calculate the rotation matrix and translation vector between the inner points. In recent years, the point cloud fine registration method can be classified into a conventional method and a deep learning-based method.
The conventional method is represented by ICP and RANSAC, and an ICP variant algorithm and a RANSAC variant algorithm are continuously presented, but the common problems of the conventional method are that: only locally optimal solutions can be obtained and the computational efficiency is low.
With the progress of the deep learning technology in recent years, more and more students put their eyes on the deep learning method. In 2019, wang proposed a DCP algorithm to realize rapid point cloud registration through a DGCNN network and an attention mechanism, but could not solve the problem of outlier rejection in point cloud registration. In 2021, bai proposed a PointDSC method, which integrates spatial consistency into feature extraction, so as to implement outlier rejection, but ignores geometric features between point clouds, resulting in low accuracy of point cloud registration.
Disclosure of Invention
The embodiment of the invention provides a point cloud fine registration method based on feature consistency and space consistency, which can improve the accuracy and speed of point cloud registration. The technical scheme is as follows:
in one aspect, a point cloud fine registration method based on feature consistency and spatial consistency is provided, and the method is applied to electronic equipment and comprises the following steps:
processing the point pairs after the point cloud rough registration to obtain feature consistency and space consistency;
on the basis of the feature consistency and the space consistency, selecting seed points by using the confidence coefficient;
defining a k nearest neighbor feature space, centering on a seed point, searching nearest neighbor points in the k nearest neighbor feature space to form a cluster, and fusing feature consistency and space consistency with non-local attention to obtain a point cloud feature updating formula so as to obtain point cloud features containing long-distance information after updating the points in the cluster;
generating a probability matching matrix through the updated point cloud characteristics, and obtaining a rotation vector and a translation matrix through singular value decomposition calculation based on the probability matching matrix to realize point cloud registration.
Further, processing the point pairs after the point cloud coarse registration to obtain feature consistency and spatial consistency includes:
the normalized value of the characteristic difference of the defined point pair is alpha, the alpha value reflects the characteristic consistency of the point pair, and the smaller the alpha value is, the more the characteristics tend to be consistent, and the specific formula is as follows:
Δf n ＝||g(x n )-g(y n )||
α n ∈α，n＝1，2，...，N
wherein g (·) represents a dynamic graph convolutional neural network; x is x n And y n From s n Decomposed to form x n 、y n And c n Respectively representing an nth source point cloud, an nth target point cloud and an nth initial point pair, wherein the initial point pair is obtained by rough point cloud registration, and N represents the total number of the point pairs, namely: representing a 6-dimensional real space, ">Representing a 3-dimensional real space; Δf n Representing the feature difference of the nth point pair; alpha n Representing feature consistency of the nth point pair; the euclidean distance is represented by |· |; [] + Non-negative operation, denoted alpha n A value of 0 or more;
defining the difference of the space distance between every two groups of point pairs as beta, wherein the beta value reflects the space consistency, the smaller the beta value is, the more the spatial position characteristics between the point pairs are matched, and the Euclidean distance between the i-th group of point pairs is set as d i The Euclidean distance difference d between the ith group of point pairs and the jth group of point pairs ij The method comprises the following steps:
d ij ＝||d i -d j ||
spatial correspondence beta between the ith and jth sets of point pairs ij The method comprises the following steps:
β ij ∈β，i＝1，2，...，N，j＝＝1，2，...，N。
further, the selecting the seed points by using the confidence on the basis of the feature consistency and the space consistency comprises the following steps:
defining seed confidence C as:
C＝α+β
wherein the alpha value reflects the feature consistency of the point pairs and the beta value reflects the spatial consistency;
and sorting the point pairs according to the confidence coefficient C value from large to small, and selecting the point pairs with p% of the top confidence coefficient ranking as seed points, wherein p is a constant.
Further, defining a k-nearest neighbor feature space, centering on a seed point, searching for nearest neighbor points in the k-nearest neighbor feature space to form a cluster, fusing feature consistency and spatial consistency with non-local attention, and obtaining a point cloud feature updating formula to obtain a point cloud feature containing long-distance information after updating the points in the cluster, wherein the method comprises the following steps:
defining a k-nearest neighbor feature space, searching for a nearest neighbor point in the k-nearest neighbor feature space by taking a seed point as a center, and recording indexes of the seed point and the nearest neighbor point as index, wherein the index is a two-dimensional matrix with the size of [ n ] seed ，n k ]，n seed Is the number of seed points, n k The number of adjacent points taking the seed point as the center; the seed points and the adjacent points form clusters, if the points are in the clusters, the seed points are regarded as inner points, and if the points are out of the clusters, the seed points are regarded as outliers;
fusing the feature consistency and the space consistency with non-local attention to obtain a point cloud feature updating formula in the k neighbor feature space, wherein the point cloud feature updating formula comprises the following steps:
f on the right side of the equal sign is a point pair characteristic obtained by performing one-layer convolution on an initial point pair coordinate obtained by coarse registration; f on the left side of the equal sign is the updated feature, MLP (·) is the model of the multi-layer perceptron, softmax (·) is the activation function, χ represents feature similarity,representing vector cross-multiplication, h (·) is a linear mapping function, α' is the expansion of α, i.e.:
wherein N is c Representing feature dimensions in a feature space;
and obtaining the point cloud characteristics containing long-distance information after the point in the cluster is updated through the obtained point cloud characteristic updating formula.
Further, the generating a probability matching matrix through the updated point cloud features, and obtaining a rotation vector and a translation matrix through singular value decomposition calculation based on the probability matching matrix, wherein the achieving the point cloud registration includes:
carrying out L2 normalization and non-negative operation on the updated point cloud characteristics, and calculating main characteristic values of the point cloud characteristics to obtain a probability matching matrix;
based on the probability matching matrix, the rotation vector and the translation matrix are obtained through singular value decomposition calculation, and point cloud registration is achieved.
Further, performing L2 normalization and non-negative operation on the updated features, and calculating the main feature values of the updated features to obtain a probability matching matrix, where the obtaining includes:
generating a probability matching matrix e through the updated characteristic information:
e＝L{m}
wherein m is obtained by normalizing the updated point cloud characteristic f output by the k nearest neighbor characteristic space, and L { · } represents calculating a main characteristic vector by a power iteration method, [ · }] + Non-negative operation, meaning that the value of m is non-negative.
Further, the rotation vector R and the translation matrix t calculated by singular value decomposition are expressed as:
in one aspect, an electronic device is provided, which includes a processor and a memory, where the memory stores at least one instruction, and the at least one instruction is loaded and executed by the processor to implement the above-mentioned point cloud fine registration method based on feature consistency and spatial consistency.
In one aspect, a computer readable storage medium is provided, in which at least one instruction is stored, and the at least one instruction is loaded and executed by a processor to implement the above-mentioned point cloud fine registration method based on feature consistency and spatial consistency.
The technical scheme provided by the embodiment of the invention has the beneficial effects that at least:
(1) In the aspect of point cloud registration, aiming at the problem that the traditional method is easy to fall into a local optimal solution, the embodiment of the invention provides a point cloud fine registration method based on feature consistency and space consistency, and the method can rapidly and accurately give out a rotation vector and a translation matrix to obtain a global optimal solution.
(2) Aiming at the low overlapping rate condition that the quantity of source point clouds and target point clouds is different in point cloud registration, the problem that outliers are difficult to process by the conventional algorithm is solved; thus, by processing the point pairs after the point cloud rough registration, the point cloud registration under the condition of low overlapping rate can be completed with higher accuracy and faster speed.
(3) Aiming at the problem that in point cloud registration, the prior algorithm ignores topology information among point clouds, the embodiment of the invention acquires the geometry information among the point clouds by introducing a dynamic graph convolution neural network; by fusing non-local attention, the point cloud characteristics containing long-distance information are acquired, and the accuracy of point cloud registration is greatly improved.
Drawings
In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly described below, and it is apparent that the drawings in the following description are only some embodiments of the present invention, and other drawings may be obtained according to these drawings without inventive effort for a person skilled in the art.
Fig. 1 is a schematic flow chart of a point cloud fine registration method based on feature consistency and spatial consistency according to an embodiment of the present invention;
fig. 2 is a schematic diagram of a point cloud registration network structure according to an embodiment of the present invention;
FIG. 3 (a) is a schematic diagram of feature consistency provided by an embodiment of the present invention;
FIG. 3 (b) is a schematic diagram of spatial uniformity provided by an embodiment of the present invention;
FIG. 4 is a schematic diagram of a feature consistency network structure according to an embodiment of the present invention;
fig. 5 is a schematic diagram of a k-nearest neighbor feature space structure provided by an embodiment of the present invention;
FIG. 6 is a schematic diagram of registration results on a 3D Patch public dataset provided by an embodiment of the present invention;
fig. 7 is a schematic structural diagram of an electronic device according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the embodiments of the present invention will be described in further detail with reference to the accompanying drawings.
As shown in fig. 1 and fig. 2, the embodiment of the invention provides a point cloud fine registration method based on feature consistency and space consistency, and after the method is used for rough registration of point clouds, point pairs obtained by rough registration generally contain mismatching points and outliers and are easy to fall into a locally optimal solution; by the method, outliers can be removed, registration accuracy is improved, and a global optimal solution is obtained. The method may be implemented by an electronic device, which may be a terminal or a server, the method comprising:
s101, processing point pairs after point cloud rough registration to obtain feature consistency and space consistency; the method specifically comprises the following steps:
a1, defining the normalized value of the characteristic difference of the point pair as alpha, wherein the alpha value reflects the characteristic consistency of the point pair, and the smaller the alpha value is, the more consistent the characteristic is, and the specific formula is as follows:
Δf n ＝||g(x n )-g(y n )||
α n ∈α，n＝1，2，...，N
wherein g (·) represents a dynamic graph convolutional neural network (Dynamic Graph Convolutional Neural Network, DGCNN); x is x n And y n From c n Decomposed to form x n 、y n And c n Respectively representing an nth source point cloud, an nth target point cloud and an nth initial point pair, wherein N represents the total number of the point pairs, namely: representing a 6-dimensional real space, ">Representing a 3-dimensional real space, wherein an initial point pair is obtained by rough point cloud registration; Δf n Representing the feature difference of the nth point pair; alpha n Representing feature consistency of the nth point pair; the euclidean distance is represented by |· |; [] + Non-negative operation, denoted alpha n A value of 0 or more;
a2, defining the difference of the space distance between every two groups of point pairs as beta, reflecting the space consistency of beta values, and setting the Euclidean distance between the i-th group of point pairs as d as the smaller the beta value is, the more the spatial position features between the point pairs are matched i The Euclidean distance difference d between the ith group of point pairs and the jth group of point pairs ij The method comprises the following steps:
d ij ＝||d i -d j ||
spatial correspondence beta between the ith and jth sets of point pairs ij The method comprises the following steps:
β ij ∈β，i＝1，2，...，N，j＝＝1，2，...，N。
as shown in FIG. 3, FIG. 3 (a) is a feature consistency illustrationThe graph, c1 and c2 are correct point pairs, c3 is an incorrect point pair, Δf 1 、Δf 2 And Δf 3 Representing the characteristic difference between the point pairs, the characteristic difference is |Deltaf 1 -Δf 2 |＜|Δf 2 -Δf 3 I (I); FIG. 3 (b) is a schematic view of spatial consistency, c1 and c2 are correct point pairs, c3 is an incorrect point pair, and d1, d2, d3 and d4 represent the Euclidean distance between points, then there is |d1-d2| < |d3-d4|.
FIG. 4 is a schematic diagram of a DGCNN network, specifically, a network input source point cloud x, in a feature consistency calculation n And target point cloud y n Performing edge convolution operation on the source point cloud and the target point cloud respectively, and extracting four layers of edge convolution, wherein the number of four layers of convolution output channels is [16, 16, 32, 64 sequentially]And adding the convolution outputs of each layer, and obtaining a final characteristic value through a multi-layer perceptron and a maximum pooling layer, wherein the convolution kernel of each layer is 1 multiplied by 1. The source point cloud characteristic value and the target point cloud characteristic value are subjected to difference to obtain a characteristic difference value delta f n ，Δf n After normalization, the feature consistency alpha is obtained.
S102, selecting seed points by using confidence on the basis of feature consistency and space consistency; the method specifically comprises the following steps:
b1, defining seed confidence coefficient C as follows:
C＝α+β
the alpha value reflects the feature consistency of the point pair, the beta value reflects the space consistency, and the larger the confidence coefficient C value of the point pair is, the more the feature consistency and the space consistency are met by the point pair, and the greater the possibility that the point pair is a correct matching point is;
and B2, sorting the point pairs according to the confidence C value from large to small, and selecting the point pairs with p percent (for example, 10 percent) of the top confidence ranking as seed points, wherein p is a constant.
S103, defining a k neighbor feature space, taking a seed point as a center, searching neighbor points in the k neighbor feature space to form a cluster, and fusing feature consistency and space consistency with non-local attention to obtain a point cloud feature updating formula so as to obtain point cloud features containing long-distance information after point updating in the cluster; the method specifically comprises the following steps:
c1, defining a k neighbor feature space, searching for neighbor points in the k neighbor feature space by taking a seed point as a center, and recording indexes of the seed point and neighbor points thereof as index, wherein index is a two-dimensional matrix with the size of [ n ] seed ，n k ]Wherein n is seed Is the number of seed points, n k The number of adjacent points taking the seed point as the center; the seed points and the adjacent points form clusters, if the points are in the clusters, the seed points are regarded as inner points, and if the points are out of the clusters, the seed points are regarded as outliers;
and C2, fusing the feature consistency and the space consistency with non-local attention, and obtaining a point cloud feature updating formula in the k neighbor feature space, wherein the point cloud feature updating formula is as follows:
f on the right side of the equal sign is a point cloud position feature (simply referred to as a point cloud feature) obtained by performing one-layer convolution on a coordinate by an initial point obtained by rough registration; f on the left side of the equal sign is the updated feature, MLP (·) is the model of the multi-layer perceptron, softmax (·) is the activation function, χ represents feature similarity,representing vector cross-multiplication, h (·) is a linear mapping function, α' is the expansion of α, i.e.:
wherein N is c Representing feature dimensions in a feature space;
in order to better understand the point cloud feature update formula described in this embodiment, a classical non-local attention model needs to be described, where the classical non-local attention model may be expressed as:
the non-local attention model has the ability to contact context information.
The feature update formula in this embodiment fuses feature consistency and spatial consistency with non-local attention, so as to obtain long-distance features combining context information in feature space, and fuse the long-distance features with original features, thereby completing feature update.
As shown in fig. 5, the k-nearest neighbor feature space network inputs a point cloud feature f,f is convolved to obtain the features phi and +.>After transposition, θ, < > is obtained>Performing matrix multiplication on phi and theta to obtain feature similarity χ; after χ and space consistency beta are fused, the whole is taken as a weight and added to the feature consistency alpha', updated features f are obtained through a multi-layer perceptron MLP, the new features are fused with feature information and space information, and the receptive field is increased.
And C3, obtaining the point cloud characteristics containing long-distance information after the point in the cluster is updated through the obtained point cloud characteristic updating formula. Each point pair has corresponding point cloud characteristics, and the point cloud characteristics of the point pairs are ordered according to index, namely, the points in the clusters are screened out, and the number of the points in the clusters is n seed ×n k The method comprises the steps of carrying out a first treatment on the surface of the The point pairs not in index will be discarded with their point cloud features, thus completing outlier culling.
S104, generating a probability matching matrix through the updated point cloud characteristics, and obtaining a rotation vector and a translation matrix through singular value decomposition calculation based on the probability matching matrix to realize point cloud registration; the method specifically comprises the following steps:
d1, carrying out L2 normalization and non-negative operation on the updated point cloud characteristics, and calculating main characteristic values of the point cloud characteristics to obtain a probability matching matrix;
in this embodiment, the probability matching matrix e is generated by the updated feature information:
e＝L{m}
wherein m is obtained by normalizing the updated point cloud characteristic f output by the k nearest neighbor characteristic space, and L { · } represents calculating a main characteristic vector by a power iteration method, [ · }] + Non-negative operation, meaning that the value of m is non-negative.
And D2, based on the probability matching matrix, obtaining a rotation vector and a translation matrix through singular value decomposition calculation, and realizing point cloud registration.
In this embodiment, the rotation vector R and the translation matrix t calculated by singular value decomposition are expressed as:
in order to verify the effect of the point cloud fine registration method based on feature consistency and spatial consistency provided in the present embodiment, the present embodiment uses the following evaluation indexes:
(1) Recall ratio RR: recall is one of the important metrics that measure the performance of the registration algorithm. The method is used for evaluating the identification capability of the algorithm on the correctly matched point pairs, the recall rate is defined as the ratio of the correctly matched point pairs to the total correctly matched point pairs, the recall rate has a value range of 0 to 1, and the larger the value is, the higher the robustness and accuracy of the algorithm are, and the correct matched point pairs can be better identified.
(2) Rotation error RE: the Rotation Error (Rotation Error) is used to measure the Rotation matrix accuracy in the point cloud transformation matrix, and is defined as follows:
wherein,is the predicted rotation matrix value, R * Is a rotation matrix true value, arccos is an inverse cosine function, tr is a trace operation.
(3) Translation error TE: the translation error (Translation Error) is used to measure the translation vector accuracy in the point cloud transformation matrix, defined as follows:
wherein,is the predicted translational vector value, t * Is the translation vector true value.
In order to verify the performance of the point cloud fine registration method based on feature consistency and space consistency provided by the embodiment of the invention, in the embodiment, a public data set 3DMatch is adopted to compare with a classical algorithm in recent years; the recall RR, rotation error RE, and translation error TE were used as evaluation criteria, and the point cloud registration results are shown in table 1. As shown in fig. 6, the source point cloud does not fully coincide with the target point cloud; as can be seen from table 1 and fig. 6, the point cloud registration result provided by the embodiment of the invention is better.
TABLE 1 Point cloud registration results
In summary, the point cloud fine registration method based on feature consistency and spatial consistency provided by the embodiment of the invention has at least the following beneficial effects:
(1) In the aspect of point cloud registration, aiming at the problem that the traditional method is easy to fall into a local optimal solution, the embodiment of the invention provides a point cloud fine registration method based on feature consistency and space consistency, and the method can rapidly and accurately give out a rotation vector and a translation matrix to obtain a global optimal solution.
(2) Aiming at the low overlapping rate condition that the quantity of source point clouds and target point clouds is different in point cloud registration, the problem that outliers are difficult to process by the conventional algorithm is solved; thus, by processing the point pairs after the point cloud rough registration, the point cloud registration under the condition of low overlapping rate can be completed with higher accuracy and faster speed.
(3) In point cloud registration, the topology information problem between point clouds is ignored by the conventional algorithm, and the embodiment of the invention acquires the geometrical information between the point clouds by introducing a dynamic graph convolution neural network (wherein the geometrical information of the nth point in the source point cloud is g (x n ) The geometric information of the nth point in the target point cloud is g (y n ) By fusing non-local attention, the point cloud characteristics containing long-distance information are acquired, and the accuracy of point cloud registration is greatly improved.
Fig. 7 is a schematic structural diagram of an electronic device 600 according to an embodiment of the present invention, where the electronic device 600 may have a relatively large difference due to different configurations or performances, and may include one or more processors (central processing units, CPU) 601 and one or more memories 602, where at least one instruction is stored in the memories 602, and the at least one instruction is loaded and executed by the processors 601 to implement the above-mentioned point cloud fine registration method based on feature consistency and spatial consistency.
In an exemplary embodiment, a computer readable storage medium, such as a memory comprising instructions executable by a processor in a terminal to perform the above-described point cloud fine registration method based on feature consistency and spatial consistency is also provided. For example, the computer readable storage medium may be ROM, random Access Memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, etc.
It will be understood by those skilled in the art that all or part of the steps for implementing the above embodiments may be implemented by hardware, or may be implemented by a program for instructing relevant hardware, where the program may be stored in a computer readable storage medium, and the storage medium may be a read-only memory, a magnetic disk or an optical disk, etc.
The foregoing description of the preferred embodiments of the invention is not intended to limit the invention to the precise form disclosed, and any such modifications, equivalents, and alternatives falling within the spirit and scope of the invention are intended to be included within the scope of the invention.

Claims (7)

1. The point cloud fine registration method based on the feature consistency and the space consistency is characterized by comprising the following steps of:
processing the point pairs after the point cloud rough registration to obtain feature consistency and space consistency;
on the basis of the feature consistency and the space consistency, selecting seed points by using the confidence coefficient;
defining a k nearest neighbor feature space, centering on a seed point, searching nearest neighbor points in the k nearest neighbor feature space to form a cluster, and fusing feature consistency and space consistency with non-local attention to obtain a point cloud feature updating formula so as to obtain point cloud features containing long-distance information after updating the points in the cluster;
generating a probability matching matrix through the updated point cloud characteristics, and obtaining a rotation vector and a translation matrix through singular value decomposition calculation based on the probability matching matrix to realize point cloud registration.
2. The point cloud fine registration method based on feature consistency and spatial consistency according to claim 1, wherein the processing the point pairs after the point cloud coarse registration to obtain feature consistency and spatial consistency comprises:
the normalized value of the characteristic difference of the defined point pair is alpha, the alpha value reflects the characteristic consistency of the point pair, and the smaller the alpha value is, the more the characteristics tend to be consistent, and the specific formula is as follows:
Δf n ＝‖g(x n )-g(y n )‖
wherein g (·) represents a dynamic graph convolutional neural network; x is x n And y n From c n Decomposed to form x n 、y n And c n Respectively representing an nth source point cloud, an nth target point cloud and an nth initial point pair, wherein the initial point pair is obtained by rough point cloud registration, and N represents the total number of the point pairs, namely: representing a 6-dimensional real space, ">Representing a 3-dimensional real space; Δf n Representing the feature difference of the nth point pair; alpha n Representing feature consistency of the nth point pair; II indicates Euclidean distance; [] + Non-negative operation, denoted alpha n A value of 0 or more;
defining the difference of the space distance between every two groups of point pairs as beta, wherein the beta value reflects the space consistency, the smaller the beta value is, the more the spatial position characteristics between the point pairs are matched, and the Euclidean distance between the i-th group of point pairs is set as d i The Euclidean distance difference d between the ith group of point pairs and the jth group of point pairs ij The method comprises the following steps:
d ij ＝||d i -d j ||
spatial correspondence beta between the ith and jth sets of point pairs ij The method comprises the following steps:
3. the point cloud fine registration method based on feature consistency and spatial consistency according to claim 1, wherein the selecting the seed points by using the confidence level based on the feature consistency and the spatial consistency comprises:
defining seed confidence C as:
C＝α+β
wherein the alpha value reflects the feature consistency of the point pairs and the beta value reflects the spatial consistency;
and sorting the point pairs according to the confidence coefficient C value from large to small, and selecting the point pairs with p% of the top confidence coefficient ranking as seed points, wherein p is a constant.
4. The method for fine registration of point clouds based on feature consistency and spatial consistency according to claim 1, wherein defining a k-nearest neighbor feature space, centering on a seed point, searching for its nearest neighbor point in the k-nearest neighbor feature space to form a cluster, fusing feature consistency and spatial consistency with non-local attention, and obtaining a point cloud feature update formula to obtain a point cloud feature containing long-distance information after updating the points in the cluster comprises:
defining a k-nearest neighbor feature space, searching for a nearest neighbor point in the k-nearest neighbor feature space by taking a seed point as a center, and recording indexes of the seed point and the nearest neighbor point as index, wherein the index is a two-dimensional matrix with the size of [ n ] seed ,n k ]，n seed Is the number of seed points, n k The number of adjacent points taking the seed point as the center; the seed points and the adjacent points form clusters, if the points are in the clusters, the seed points are regarded as inner points, and if the points are out of the clusters, the seed points are regarded as outliers;
fusing the feature consistency and the space consistency with non-local attention to obtain a point cloud feature updating formula in the k neighbor feature space, wherein the point cloud feature updating formula comprises the following steps:
f on the right side of the equal sign is a point pair characteristic obtained by performing one-layer convolution on an initial point pair coordinate obtained by coarse registration; f on the left side of the equal sign is the updated feature, MLP (·) is the model of the multi-layer perceptron, softmax (·) is the activation function, χ represents feature similarity,representing vector cross-multiplication, h (·) is a linear mapping function, α' is the expansion of α, i.e.:
wherein N is c Representing feature dimensions in a feature space;
and obtaining the point cloud characteristics containing long-distance information after the point in the cluster is updated through the obtained point cloud characteristic updating formula.
5. The point cloud fine registration method based on feature consistency and space consistency according to claim 1, wherein the generating a probability matching matrix by updated point cloud features, obtaining a rotation vector and a translation matrix by singular value decomposition calculation based on the probability matching matrix, and realizing point cloud registration comprises:
carrying out L2 normalization and non-negative operation on the updated point cloud characteristics, and calculating main characteristic values of the point cloud characteristics to obtain a probability matching matrix;
based on the probability matching matrix, the rotation vector and the translation matrix are obtained through singular value decomposition calculation, and point cloud registration is achieved.
6. The method for fine registration of point clouds based on feature consistency and spatial consistency according to claim 5, wherein the step of performing L2 normalization and non-negative operation on the updated features and calculating main feature values thereof to obtain a probability matching matrix comprises:
generating a probability matching matrix e through the updated characteristic information:
e＝L{m}
wherein m is defined by k neighborThe updated point cloud characteristic f of the feature space output is normalized, and L { · } represents that a power iteration method is used for calculating a main characteristic vector, [ ·] + Non-negative operation, meaning that the value of m is non-negative.
7. The point cloud fine registration method based on feature consistency and spatial consistency according to claim 6, wherein the rotation vector R and the translation matrix t calculated by singular value decomposition are expressed as:
CN202311335751.4A 2023-10-16 2023-10-16 Point cloud fine registration method based on feature consistency and spatial consistency Active CN117351052B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202311335751.4A CN117351052B (en) 2023-10-16 2023-10-16 Point cloud fine registration method based on feature consistency and spatial consistency

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202311335751.4A CN117351052B (en) 2023-10-16 2023-10-16 Point cloud fine registration method based on feature consistency and spatial consistency

Publications (2)

Publication Number Publication Date
CN117351052A true CN117351052A (en) 2024-01-05
CN117351052B CN117351052B (en) 2024-09-20

Family

ID=89359016

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202311335751.4A Active CN117351052B (en) 2023-10-16 2023-10-16 Point cloud fine registration method based on feature consistency and spatial consistency

Country Status (1)

Country Link
CN (1) CN117351052B (en)

Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110176746A1 (en) * 2008-10-03 2011-07-21 Universite Joseph Fourier Grenoble 1 Method for registering a set of points in images
CN107590829A (en) * 2017-09-18 2018-01-16 西安电子科技大学 A kind of seed point pick-up method for being applied to the intensive cloud data registration of various visual angles
CN110322507A (en) * 2019-06-04 2019-10-11 东南大学 A method of based on depth re-projection and Space Consistency characteristic matching
CN112396641A (en) * 2020-11-17 2021-02-23 中国矿业大学(北京) Point cloud global registration method based on congruent two-baseline matching
US20220084221A1 (en) * 2019-02-11 2022-03-17 Siemens Aktiengesellschaft An apparatus and a method for performing a data driven pairwise registration of three-dimensional point clouds
CN114588636A (en) * 2020-12-07 2022-06-07 辉达公司 Graphics processing unit for detecting cheating using neural networks
CN114937066A (en) * 2022-06-09 2022-08-23 重庆理工大学 Point cloud registration system and method based on cross offset features and space consistency
CN115116013A (en) * 2021-03-19 2022-09-27 上海交通大学 Online dense point cloud semantic segmentation system and method integrating time sequence features

Patent Citations (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20110176746A1 (en) * 2008-10-03 2011-07-21 Universite Joseph Fourier Grenoble 1 Method for registering a set of points in images
CN107590829A (en) * 2017-09-18 2018-01-16 西安电子科技大学 A kind of seed point pick-up method for being applied to the intensive cloud data registration of various visual angles
US20220084221A1 (en) * 2019-02-11 2022-03-17 Siemens Aktiengesellschaft An apparatus and a method for performing a data driven pairwise registration of three-dimensional point clouds
CN110322507A (en) * 2019-06-04 2019-10-11 东南大学 A method of based on depth re-projection and Space Consistency characteristic matching
CN112396641A (en) * 2020-11-17 2021-02-23 中国矿业大学(北京) Point cloud global registration method based on congruent two-baseline matching
CN114588636A (en) * 2020-12-07 2022-06-07 辉达公司 Graphics processing unit for detecting cheating using neural networks
CN115116013A (en) * 2021-03-19 2022-09-27 上海交通大学 Online dense point cloud semantic segmentation system and method integrating time sequence features
CN114937066A (en) * 2022-06-09 2022-08-23 重庆理工大学 Point cloud registration system and method based on cross offset features and space consistency

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
PORNANAN RAKTRAKULTHUM等: "Vehicle Classification in Congested Traffic Based on 3D Point Cloud Using SVM and KNN", 《2017 9TH INTERNATIONAL CONFERENCE ON INFORMATION TECHNOLOGY AND ELECTRICAL ENGINEERING》, 31 December 2017 (2017-12-31), pages 1 - 6 *
陈立: "建筑物三维点云边界特征提取算法研究", 《中国优秀硕士学位论文全文数据库基础科学辑》, no. 06, 15 June 2020 (2020-06-15), pages 005 - 175 *

Also Published As

Publication number Publication date
CN117351052B (en) 2024-09-20

Similar Documents

Publication Publication Date Title
CN111291139B (en) Knowledge graph long-tail relation completion method based on attention mechanism
CN108665491B (en) Rapid point cloud registration method based on local reference points
WO2022095356A1 (en) Transfer learning method for image classification, related device, and storage medium
CN113807422B (en) Weighted graph convolutional neural network scoring prediction model integrating multi-feature information
CN110378423A (en) Feature extracting method, device, computer equipment and storage medium
Xu et al. Iterative manifold embedding layer learned by incomplete data for large-scale image retrieval
Nie et al. MMFN: Multimodal information fusion networks for 3D model classification and retrieval
CN108399268A (en) A kind of increment type isomery figure clustering method based on game theory
CN111797269A (en) Multi-view three-dimensional model retrieval method based on multi-level view associated convolutional network
WO2022267955A1 (en) Post-fusion multi-view clustering method and system based on local maximum alignment
CN116071331A (en) Workpiece surface defect detection method based on improved SSD algorithm
CN114119690A (en) Point cloud registration method based on neural network reconstruction Gaussian mixture model
CN114782503A (en) Point cloud registration method and system based on multi-scale feature similarity constraint
KR20210035017A (en) Neural network training method, method and apparatus of processing data based on neural network
CN117876447A (en) Three-dimensional point cloud registration method based on micro-surface fusion and alignment
CN117351052B (en) Point cloud fine registration method based on feature consistency and spatial consistency
Chen et al. R2-MLP: Round-roll MLP for multi-view 3D object recognition
Lian et al. SORCNet: robust non-rigid shape correspondence with enhanced descriptors by Shared Optimized Res-CapsuleNet
CN116705192A (en) Drug virtual screening method and device based on deep learning
Lakhili et al. Rigid and non-rigid 3D shape classification based on 3D Hahn moments neural networks model
Guan et al. View-based 3D model retrieval by joint subgraph learning and matching
Wang et al. Image matching via the local neighborhood for low inlier ratio
Hao et al. Cascaded geometric feature modulation network for point cloud processing
Mittal Ensemble of transfer learnt classifiers for recognition of cardiovascular tissues from histological images
CN111401385A (en) Similarity calculation method for image local topological structure feature descriptors

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