WO2020168620A1 - 平面几何一致性检测方法、计算机设备、及存储介质 - Google Patents

平面几何一致性检测方法、计算机设备、及存储介质 Download PDF

Info

Publication number
WO2020168620A1
WO2020168620A1 PCT/CN2019/081424 CN2019081424W WO2020168620A1 WO 2020168620 A1 WO2020168620 A1 WO 2020168620A1 CN 2019081424 W CN2019081424 W CN 2019081424W WO 2020168620 A1 WO2020168620 A1 WO 2020168620A1
Authority
WO
WIPO (PCT)
Prior art keywords
rotation
distance
plane
normal
vector
Prior art date
Application number
PCT/CN2019/081424
Other languages
English (en)
French (fr)
Inventor
邢自然
杨丹
Original Assignee
曜科智能科技(上海)有限公司
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 曜科智能科技(上海)有限公司 filed Critical 曜科智能科技(上海)有限公司
Publication of WO2020168620A1 publication Critical patent/WO2020168620A1/zh
Priority to US17/383,419 priority Critical patent/US20210357477A1/en

Links

Images

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/21Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
    • G06F18/213Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
    • G06F18/2133Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on naturality criteria, e.g. with non-negative factorisation or negative correlation
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V20/00Scenes; Scene-specific elements
    • G06V20/60Type of objects
    • G06V20/64Three-dimensional objects
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/22Indexing; Data structures therefor; Storage structures
    • G06F16/2228Indexing structures
    • G06F16/2237Vectors, bitmaps or matrices
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/20Information retrieval; Database structures therefor; File system structures therefor of structured data, e.g. relational data
    • G06F16/22Indexing; Data structures therefor; Storage structures
    • G06F16/2228Indexing structures
    • G06F16/2255Hash tables
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10Complex mathematical operations
    • G06F17/16Matrix or vector computation, e.g. matrix-matrix or matrix-vector multiplication, matrix factorization
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00Pattern recognition
    • G06F18/20Analysing
    • G06F18/24Classification techniques
    • G06F18/241Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2413Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on distances to training or reference patterns
    • G06F18/24147Distances to closest patterns, e.g. nearest neighbour classification
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T7/00Image analysis
    • G06T7/30Determination of transform parameters for the alignment of images, i.e. image registration
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06VIMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00Arrangements for image or video recognition or understanding
    • G06V10/70Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/751Comparing pixel values or logical combinations thereof, or feature values having positional relevance, e.g. template matching
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06TIMAGE DATA PROCESSING OR GENERATION, IN GENERAL
    • G06T2207/00Indexing scheme for image analysis or image enhancement
    • G06T2207/10Image acquisition modality
    • G06T2207/10016Video; Image sequence

Definitions

  • This application relates to the field of image processing technology, in particular to a plane geometric consistency detection method, computer equipment, and storage medium.
  • Plane is the main feature used in artificial image scenes and has been fully utilized in many surveying and mapping methods.
  • 3D mapping technology provides detailed and accurate information about the scene, and can be used for localization and mapping (SLAM), augmented reality, video stabilization, 3D reconstruction and other fields because it can be applied to inexpensive sensors. It is becoming popular and used by people day by day.
  • NDT Normal Distribution Transform
  • ICP Itera. Closest Point
  • the purpose of this application is to provide a plane geometric consistency detection method, computer equipment, and storage medium to solve the problems in the prior art.
  • this application provides a plane geometric consistency detection method, in which each plane is uniquely represented in a coordinate system by its normal vector and the distance between the origin of the coordinate system where the plane is located and the plane;
  • the method is used to obtain a consistent conversion relationship from a first plane set composed of first planes in a first coordinate system to a second plane set composed of second planes in a second coordinate system, and the conversion relationship passes through the target
  • the rotation matrix and the target translation matrix are expressed; the method includes: obtaining a first normal vector set and a first distance set composed of each first normal vector and each first distance representing each first plane, and representing each A second normal vector set and a second distance set composed of each second normal vector and each second distance of the second plane;
  • obtaining the target rotation matrix includes: obtaining each pair of first normal vectors in the first normal vector set , A pair of second normal vectors that match the angle of the pair of first normal vectors in the second normal vector set, and calculate the first rotation matrix converted from each pair of first normal vectors to the pair of
  • the first rotation matrix is defined by a pair of first normal vectors and another first normal vector perpendicular to them, a pair of second normal vectors matching the included angle and another second normal vector perpendicular to them.
  • the mapping relationship between normal vectors is obtained.
  • each pair of first normal vectors is not parallel; the first normal vector of each parallel first plane and the second normal vector of each parallel second plane are set to be the same.
  • the included angle matching means that the deviation between the included angle between a pair of first normal vectors and the included angle between a pair of second normal vectors is within a preset range.
  • the clustering method includes: inserting each element in the first target category or the second target category into a hash table, and constructing its corresponding index according to each element; or,
  • the index of each element is established based on the data it contains and the maximum value it can reach.
  • the rotation vector is expressed as a vector extending along a rotation axis vector direction and having a length representing its rotation angle; the construction of the first hash table inserted into the rotation vector includes: dividing the three-dimensional angle space into Is a number of cubic cells with a size ⁇ 3 , ⁇ ⁇ (0, 2 ⁇ ); establish a three-dimensional hash table with a size of N 3 as the first hash table, And by calculating the index To insert each rotation vector in the rotation vector set into each cell; where r is the rotation vector.
  • the construction of the second hash table with the difference in insertion distance includes: dividing the three-dimensional space into multiple sizes of A cell in the form of a cube, th d ⁇ (0,maxD); establish a three-dimensional hash table of size N 3 as the second hash table, And by calculating the index Insert each distance difference in the distance difference set into the second hash table; where d is the distance difference, maxD is the maximum distance, and th d is the cell size threshold.
  • the first/second projection matrix is calculated by two orthogonal matrices obtained by performing singular value decomposition on the first/second rotation matrix.
  • the method further includes: calculating the scores of the obtained multiple target rotation matrices, wherein the scores are related to the number of non-parallel plane groups used to calculate the target rotation matrix, and each The number of parallel planes in a plane group; where the target rotation matrix with the best score is the rotation matrix that meets the maximum geometric consistency of the conversion from the first plane set to the second plane set.
  • this application provides a computer device, including: a processor and a memory; the memory is used to store a computer program; the processor is used to run the computer program to achieve plane geometry Consistency detection method; wherein each plane is uniquely represented in a coordinate system by its normal vector and the distance between the origin of the coordinate system where the plane is located and the plane; the method is used to obtain each first plane in the first coordinate system A consistent conversion relationship from the first plane set composed to the second plane set composed of the second planes in the second coordinate system, and the conversion relationship is expressed by the principle of the target rotation matrix and the target translation matrix; the plane geometry
  • the consistency detection method includes: obtaining a first normal vector set and a first distance set composed of each first normal vector representing each first plane and each first distance, and a first normal vector set and a first distance set composed of each second plane representing each second plane.
  • a second normal vector set and a second distance set composed of a normal vector and each second distance; obtaining the target rotation matrix includes: obtaining each pair of first normal vectors in the first normal vector set and the pair of first normal vectors A pair of second normal vectors whose included angles match between the vectors in the second normal vector set, calculate the first rotation matrix converted from each pair of first normal vectors to a pair of second normal vectors whose included angles match, and obtain each The first projection matrix of the first rotation matrix in the special orthogonal group SO(3) space; each first projection matrix obtained is converted into a rotation vector through the inverse transformation of the Rodriguez rotation formula to obtain a rotation vector set; Each element in each rotation vector set is clustered to obtain the first target classification with the most elements.
  • a pair of first normal vectors and a pair of first normal vectors and a pair of first normal vectors are used as the basis for calculating the rotation vector for each element in the first target classification.
  • the two normal vectors are respectively placed in a first normal vector sequence and a second normal vector sequence; the second rotation matrix converted from the first normal vector sequence to the second normal vector sequence is calculated, and each second rotation matrix is obtained
  • the first rotation matrix is defined by a pair of first normal vectors and another first normal vector perpendicular to them, a pair of second normal vectors matching the included angle and another second normal vector perpendicular to them.
  • the mapping relationship between normal vectors is obtained.
  • each pair of first normal vectors is not parallel; the first normal vector of each parallel first plane and the second normal vector of each parallel second plane are set to be the same.
  • the included angle matching means that the deviation between the included angle between a pair of first normal vectors and the included angle between a pair of second normal vectors is within a preset range.
  • the clustering includes: inserting each element in the first target category or the second target category into a hash table, and constructing its corresponding index according to each element.
  • the index of each element is established based on the data it contains and the maximum value it can reach.
  • the rotation vector is represented as a vector extending along a rotation axis vector direction and having a length representing its rotation angle; the construction of the first hash table inserted into the rotation vector includes: dividing the three-dimensional angle space into Is a number of cubic cells with a size ⁇ 3 , ⁇ ⁇ (0, 2 ⁇ ); establish a three-dimensional hash table with a size of N 3 as the first hash table, And by calculating the index To insert each rotation vector in the rotation vector set into each cell; where r is the rotation vector.
  • the construction of the second hash table with the difference in insertion distance includes: dividing the three-dimensional space into multiple sizes of A cell in the form of a cube, th d ⁇ (0,maxD); establish a three-dimensional hash table of size N 3 as the second hash table, And by calculating the index Insert each distance difference in the distance difference set into the second hash table; where d is the distance difference, maxD is the maximum distance, and th d is the cell size threshold.
  • the first/second projection matrix is calculated by two orthogonal matrices obtained by performing singular value decomposition on the first/second rotation matrix.
  • the method further includes: calculating the scores of the obtained multiple target rotation matrices, wherein the scores are related to the number of non-parallel plane groups used to calculate the target rotation matrix, and each The number of parallel planes in a plane group; where the target rotation matrix with the best score is the rotation matrix that meets the maximum geometric consistency of the conversion from the first plane set to the second plane set.
  • the present application provides a computer-readable storage medium storing a computer program that executes the method when the computer program is run by a processor.
  • the target rotation matrix includes: obtaining each pair of first normal vectors in the first normal vector set and a pair of second normal vectors matching the angle between the pair of first normal vectors in the second normal vector set, calculating the first rotation matrix and Its first projection matrix; convert the obtained first projection matrix into a rotation vector to obtain a rotation vector set; cluster each element in each rotation vector set to obtain the element in the first target category with the most elements and obtain the first The normal vector sequence and the second normal vector sequence; calculate the second rotation matrix converted from the first normal vector sequence to the second normal vector sequence and use its second projection matrix as the target rotation matrix; in addition, pass the first and second The distance set obtains the distance difference set, the clustering obtains the second target classification, and the first and second distances of the elements and the corresponding second normal vectors are used to calculate the target translation matrix; the solution of this application achieve
  • FIG. 1 shows a schematic flow chart of a method for detecting plane geometric consistency in an embodiment of this application.
  • Fig. 2 shows a schematic structural diagram of a computer device in an embodiment of this application.
  • the first plane set in the first coordinate system F A is expressed as The plane contained Normal vector And from the origin of the first coordinate system to the plane The distance of is uniquely expressed so that it is located at The point p A on satisfies:
  • each plane in F B Can be made by satisfying the above formula Unique definition.
  • FIG. 1 a schematic flow chart of the method for detecting plane geometric consistency in an embodiment of the present application is shown. Indeed, this method is used to obtain maximum correspondence between the R BA and F B F A rigid body motion mapping, T BA.
  • the method includes:
  • Step S101 Obtain the first normal vector set and the first distance set composed of the first normal vectors representing the first planes and the first distances, and the second normal vector sums representing the second planes Each second distance is composed of a second normal vector set and a second distance set.
  • the first plane be expressed as The corresponding first plane set is expressed as The second plane is represented as The corresponding second plane set is expressed as The first normal vector is expressed as The second normal vector is expressed as
  • Step S102 Obtain the target rotation matrix R BA .
  • the characteristics of rigid body motion that is, maintaining the norm and intersection of vectors, are used to calculate the rotation matrix between the plane sets in the two coordinate systems.
  • the rotation matrix between the plane sets in the two coordinate systems.
  • step S102 may specifically include:
  • Step S121 Obtain each pair of first normal vectors in the first normal vector set and a pair of second normal vectors that match the angle between the pair of first normal vectors in the second normal vector set.
  • n i and n j is the normal vector at different planes F A.
  • Step S122 Calculate the first rotation matrix converted from each pair of first normal vectors to the pair of second normal vectors whose included angles match, and obtain each first rotation matrix in the special orthogonal group SO(3) space The first projection matrix.
  • n k is a vector perpendicular to them constructed according to n i and n j
  • n r is a vector perpendicular to them constructed from n p and n q ; the reason for construction is because the rotation matrix is a 3 ⁇ 3 matrix
  • each normal vector is a 3 ⁇ 1 vector.
  • R BA is an orthonormal matrix, it falls into the Lie group SO(3), and It is not necessarily in SO(3), therefore, in this embodiment, the SVD method Project to SO(3) to get R BA . Since U and V must be orthogonal matrices after SVD decomposition, R BA is also an orthogonal matrix.
  • the R BA obtained here is only a preliminary result, and its calculation method can be optimized in terms of the amount of calculation and can be further optimized for consistency.
  • N max ⁇ N A ,N B ⁇ .
  • N 2 the number of rotation schemes is significantly reduced to N 2 .
  • the threshold th ⁇ >0 is introduced as the acceptable angular noise threshold to establish the criterion for the matching angle; so that for ⁇ i,j ⁇ PG A , the angle value is ( ⁇ i,j -th ⁇ , ⁇ i , PG B within j + th ⁇ ) in the range of ⁇ p, q are considered ⁇ i, j.
  • the search time for possible matches of ⁇ i,j will be reduced to O(logN 2 ); in addition, in the worst case corresponding to the time complexity of O(N 4 ), for example, most plane perpendicular to each other, to find the total time complexity for all pairs of matching angle between PG a PG B average and reduced to O (N 2 logN).
  • the matching set M is defined as the following formulas (14) ⁇ (15):
  • Step S123 Convert each obtained first projection matrix into a rotation vector through the inverse transformation of the Rodriguez rotation formula to obtain a rotation vector set;
  • Step S124 Cluster each element in each rotation vector set to obtain the first target classification with the most elements, which will be a pair of first normal vectors on which each element in the first target classification is calculated, that is, the rotation vector And a pair of second normal vectors are respectively placed in a first normal vector sequence and a second normal vector sequence;
  • Step S125 Calculate the second rotation matrix converted from the first normal vector sequence to the second normal vector sequence, and obtain the second projection matrix of each second rotation matrix in the special orthogonal group SO(3) space as the result The target rotation matrix.
  • R is transformed into a three-dimensional rotation vector with the rotation axis (vector) as the direction and the angle as the length, so as to obtain a more compact R representation through the axis and the angle .
  • the S function is also involved to flip the coordinate sign of r to ensure its uniqueness
  • a hash table can be used to solve this problem.
  • the three-dimensional angle space is divided into a plurality of cube units with a size of ⁇ 3 to store the rotation vector of M, where ⁇ ⁇ (0, 2 ⁇ ). Therefore, a first hash table of size N 3 is generated, where N is given by the number of cells in each rotation axis.
  • a three-dimensional hash table of size N 3 is established as the first hash table, where:
  • the cell Cr with the largest number of elements n c (ie the first target classification) is calculated according to formula (2) with The best rotation scheme between is used as the target rotation matrix.
  • V A and V B with a size of 3 ⁇ n c .
  • V A and V B in equations (24) and (25) are both 3*n c matrices; thus in (26), and equation ( 2)
  • equation ( 1) The same principle as (11) is that according to Transform to find the rotation matrix from V A to V B
  • the principle of reuse is similar to formulas (27) and (28) of formulas (12) and (13) to obtain the target rotation matrix R BA :
  • Step S103 Obtain the translation matrix T BA .
  • step S103 acquiring translation matrix T BA between F A and F B can be derived from the following detailed description:
  • Step S131 Subtract each first distance in the first distance set from each second distance in the second distance set to obtain each distance difference and form a distance difference set.
  • Step S132 Cluster the elements in the distance difference set to obtain the second target category with the most elements.
  • Step S133 According to the first distance and the second distance as the basis for calculating the distance difference for each element in the second target classification, combined with each of the second distances as the basis for calculating the distance difference and the second distance belonging to the same plane.
  • the normal vector is calculated to obtain the target translation matrix.
  • the distance difference set Let the first distance be expressed as The second distance is expressed as The element in the distance difference is expressed as:
  • each distance difference in the distance difference set can also be inserted into a second hash table.
  • the calculation method of the index includes:
  • maxD and th d respectively represent the defined maximum distance and the size threshold of the cell storing the element, and the cell may be.
  • the second normal vector that belongs to the same plane can be detected and the transposed matrix It can also be obtained by obtaining T BA .
  • each planegroup is equivalent to only participating in the calculation once, thereby reducing the amount of calculation.
  • each target rotation matrix will be voted to calculate a score (score), and the target rotation matrix with the highest score is selected as the rotation matrix with the best consistency; the score is the definition To be related to the number of all planes used to calculate the corresponding target rotation matrix, all planes include the first plane set under F A All first planes in And the second plane set under F B All second planes in
  • the target rotation matrix R i is calculated based on m plane groups that are not parallel to each other, where each plane group has r j parallel planes, where each plane group can be a group In parallel Composition, it can also be a set of parallel Composition; R i 's score is related to m and r j .
  • the score calculation method of R i may be:
  • each plane group is composed of n planes, it will take O(n 2 ) time to find a matching plane according to the content obtained by the aforementioned translation matrix T BA . Therefore, the total time to calculate the score of the rotation matrix R i is
  • FIG. 2 a schematic diagram of the structure of the computer device 200 in the embodiment is shown.
  • the computer device 200 includes a memory 201 and a processor 202.
  • the memory is used to store computer programs
  • the processor is configured to run the computer program to execute, for example, the plane geometric consistency detection method in the embodiment of FIG. 1, which specifically includes:
  • the second normal vector set and the second distance set are respectively composed of distances;
  • obtaining the target rotation matrix includes: obtaining each pair of first normal vectors in the first normal vector set, and the pair of first normal vectors in the second normal vector Concentrate a pair of second normal vectors with matching angles, calculate the first rotation matrix from each pair of first normal vectors to a pair of second normal vectors with matching angles, and obtain the special The first projection matrix in the orthogonal group SO(3) space; each first projection matrix obtained is converted into a rotation vector by the inverse transformation of the Rodriguez rotation formula to obtain a rotation vector set; for each rotation vector set The elements are clustered to obtain the first target category with the most elements, and a pair of first normal vectors and a pair of second normal vectors that are the calculation basis for each element in the first target category, namely the rotation vector, are placed
  • each pair of first normal vectors is not parallel; the first normal vector of each parallel first plane and the second normal vector of each parallel second plane are set to be the same.
  • the included angle matching means that the deviation between the included angle between a pair of first normal vectors and the included angle between a pair of second normal vectors is within a preset range.
  • the clustering includes: inserting each element in the first target category or the second target category into a hash table, and constructing its corresponding index according to each element.
  • the index of each element is established based on the data it contains and the maximum value it can reach.
  • the rotation vector is expressed as a vector extending along a rotation axis vector direction and having a length representing its rotation angle; the construction of the first hash table inserted into the rotation vector includes: dividing the three-dimensional angle space into Is a number of cubic cells with a size of ⁇ 3, ⁇ (0,2 ⁇ ); establish a three-dimensional hash table with a size of N 3 as the first hash table, And by calculating the index To insert each rotation vector in the rotation vector set into each cell; where r is the rotation vector.
  • the construction method of the second hash table with the difference in insertion distance includes: dividing the three-dimensional space into a plurality of cube-shaped cells with a size of thd 3 , th d ⁇ (0, maxD); the establishment size is The three-dimensional hash table of N 3 serves as the second hash table, And by calculating the index Insert each distance difference in the distance difference set into the second hash table; where d is the distance difference, maxD is the maximum distance, and th d is the cell size threshold.
  • the first/second projection matrix is calculated by two orthogonal matrices obtained by performing singular value decomposition on the first/second rotation matrix, and "/" means and/or, that is, In other words, the first projection matrix is obtained by calculating two orthogonal matrices obtained by performing singular value decomposition on the first rotation matrix, and/or, the second projection matrix is calculated by performing singular values on the second rotation matrix The two orthogonal matrices obtained by decomposition are calculated.
  • the method further includes: calculating the scores of the obtained multiple target rotation matrices, wherein the scores are related to the number of non-parallel plane groups used to calculate the target rotation matrix, and each The number of parallel planes in a plane group; where the target rotation matrix with the best score is the rotation matrix that meets the maximum geometric consistency of the conversion from the first plane set to the second plane set.
  • the memory 201 may include, but is not limited to, high-speed random access memory and non-volatile memory.
  • non-volatile memory For example, one or more disk storage devices, flash memory devices, or other non-volatile solid state storage devices.
  • the processor 202 may be a general-purpose processor, including a central processing unit (Central Processing Unit, CPU for short), a network processor (Network Processor, NP for short), etc.; it may also be a digital signal processor (Digital Signal Processing, DSP for short) and an application specific integrated circuit. (ApplicationSpecificIntegratedCircuit, ASIC for short), Field-ProgrammableGateArray (FPGA for short) or other programmable logic devices, discrete gates or transistor logic devices, discrete hardware components.
  • CPU Central Processing Unit
  • NP Network Processor
  • DSP Digital Signal Processing
  • ASIC Application SpecificIntegratedCircuit
  • FPGA Field-ProgrammableGateArray
  • the computer-readable storage medium may include, but is not limited to, a floppy disk, Optical disk, CD-ROM (compact disk-read only memory), magneto-optical disk, ROM (read only memory), RAM (random access memory), EPROM (erasable programmable read only memory), EEPROM (electrically erasable Except programmable read-only memory), magnetic or optical cards, flash memory, or other types of media/machine-readable media suitable for storing machine-executable instructions.
  • the computer-readable storage medium may be a product that is not connected to a computer device, or a component that has been connected to a computer device for use.
  • the computer program is a routine, program, object, component, data structure, etc. that performs a specific task or implements a specific abstract data type.
  • the plane geometry consistency detection method, computer equipment, and storage medium of this application realize the acquisition of the first normal vector set, the first distance set, the second normal vector set, and the second distance set;
  • the target rotation matrix includes: obtaining each pair of first normal vectors in the first normal vector set and a pair of second normal vectors matching the angle between the pair of first normal vectors in the second normal vector set, and calculating the first rotation matrix And its first projection matrix; convert the obtained first projection matrix into a rotation vector to obtain a rotation vector set; cluster each element in each rotation vector set to obtain the element in the first target classification with the most elements A normal vector sequence and a second normal vector sequence; calculate the second rotation matrix from the first normal vector sequence to the second normal vector sequence and use its second projection matrix as the target rotation matrix; in addition, pass the first and second The two distance sets obtain the distance difference set, the clustering obtains the second target classification, and the first and second distances of the elements and the corresponding second normal vectors are used to calculate the target translation matrix; the solution of this application achieves high consistency and Fast and

Abstract

本申请的平面几何一致性检测方法、计算机设备、及存储介质,方法通过获得第一和第二法向量集、及第一和第二距离集;获取每对第一法向量及其夹角匹配的一对第二法向量,计算第一旋转矩阵及其第一投影矩阵;将第一投影矩阵转换为旋转向量组成旋转向量集;对旋转向量集中各元素进行聚类,根据最多元素的第一目标分类中元素生成第一和第二法向量序列;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵并以其第二投影矩阵作为目标旋转矩阵;另外,通过第一和第二距离集得到距离差集,聚类得到第二目标分类并以其中元素的第一、第二距离和对应第二法向量,计算得到目标平移矩阵;本申请方案实现高一致性且快速的不同坐标系下平面映射。

Description

平面几何一致性检测方法、计算机设备、及存储介质 技术领域
本申请涉及图像处理技术领域,尤其涉及平面几何一致性检测方法、计算机设备、及存储介质。
背景技术
平面是人造图像场景使用的主要特征,在许多测绘方法中已经被充分利用。3D映射技术提供了关于该场景的详细和准确的信息,并且由于能够应用于廉价传感器而可被用于定位与地图构建(SLAM)、增强现实、视频稳像、3D重建和其他领域的能力,正日益为人们所流行和使用。
然而,在考虑其精度、速度和鲁棒性时,平面的配准仍然具有挑战性。许多平面配准方法,通过利用各种特征,例如点、边缘和平面等,需要花费大量时间来检索帧之间的相对几何约束。因此,这种方法需要选择关键帧结构来实现实时性的表现。
此外,一些算法采用投影关联方法,如NDT(Normal Distribution Transform)或ICP(Itera.Closest Point)等,需要良好的初始预测并导出局部最优解。
因此,这些方法高度依赖于初始化来检索良好的几何约束。
发明内容
鉴于以上所述现有技术的缺点,本申请的目的在于提供平面几何一致性检测方法、计算机设备、及存储介质,解决现有技术中的问题。
为实现上述目标及其他相关目标,本申请提供一种平面几何一致性检测方法,每个平面在一坐标系下通过其法向量、及平面所在坐标系原点距该平面的距离唯一表示;所述方法用于获得由第一坐标系下各第一平面组成的第一平面集向由第二坐标系下各第二平面组成的第二平面集转换的一致的转换关系,所述转换关系通过目标旋转矩阵及目标平移矩阵表示;所述方法包括:获得由表示各第一平面的各第一法向量和各第一距离分别组成的第一法向量集及第一距离集、以及由表示每个第二平面的各第二法向量和各第二距离分别组成的第二法向量集及第二距离集;获取所述目标旋转矩阵,包括:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量,计算每对第一法向量向其夹角匹配的一对第二法向量转换的第一旋转矩阵,并获取每个所述第一旋转矩阵在特殊正交群SO(3)空间中的第一投影矩阵;通过罗德里格斯旋转公式的逆变换将得到的各第一投影矩 阵转换为旋转向量,得到旋转向量集;对各旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类,将作为该第一目标分类中每个元素即旋转向量的计算依据的一对第一法向量和一对第二法向量分别置于一第一法向量序列和第二法向量序列中;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵,并获取每个所述第二旋转矩阵在特殊正交群SO(3)空间中的第二投影矩阵作为所述目标旋转矩阵;获取所述目标平移矩阵,包括:将第一距离集中的每个第一距离与第二距离集中的每个第二距离相减,得到各距离差并组成距离差集;将距离差集中的各元素进行聚类,以获得拥有最多元素的第二目标分类;根据作为第二目标分类中的每个元素即距离差的计算依据的第一距离和第二距离,结合与每个所述作为距离差计算依据的第二距离属于同一平面的第二法向量,计算得到所述目标平移矩阵。
于一实施例中,所述第一旋转矩阵是通过一对第一法向量连同垂直它们的另一第一法向量、与夹角匹配的一对第二法向量连同垂直它们的另一第二法向量间的映射关系得到的。
于一实施例中,所述每对第一法向量间不平行;平行的各第一平面的第一法向量、和平行的各第二平面的第二法向量设为同一个。
于一实施例中,所述夹角匹配指的是一对第一法向量间夹角和一对第二法向量间夹角的偏差在预设范围内。
于一实施例中,所述聚类方式包括:将第一目标分类或第二目标分类中的各元素插入哈希表中,并根据每个元素构建其对应索引;或者,
于一实施例中,每个元素的索引是根据其包含数据、及其所能达到的最大取值来建立的。
于一实施例中,所述旋转向量表示为沿一旋转轴向量方向延伸并具有表示其旋转角度的长度的向量;插入旋转向量的第一哈希表的构建方式包括:将三维角度空间分为多个尺寸为∈ 3的立方体形式的单元格,∈∈(0,2π);建立尺寸为N 3的三维哈希表作为第一哈希表,
Figure PCTCN2019081424-appb-000001
且通过计算索引
Figure PCTCN2019081424-appb-000002
以将旋转向量集中各旋转向量插入各单元格中;其中,r为旋转向量。
于一实施例中,插入距离差的第二哈希表的构建方式包括:将三维空间分为多个尺寸为
Figure PCTCN2019081424-appb-000003
的立方体形式的单元格,th d∈(0,maxD);建立尺寸为N 3的三维哈希表作为第二哈希表,
Figure PCTCN2019081424-appb-000004
且通过计算索引
Figure PCTCN2019081424-appb-000005
将距离差集中的各距离差插入第二哈希表中;其中,d为距离差,maxD为最大距离,th d为单元格的尺寸阈值。
于一实施例中,所述第一/第二投影矩阵是通过对第一/第二旋转矩阵进行奇异值分解得到的两个正交矩阵计算得到的。
于一实施例中,所述方法还包括:计算所获取的多个目标旋转矩阵的得分,其中,该得 分相关于用于计算目标旋转矩阵的互不平行的各平面组的数量、以及构成每个平面组的平行平面的数量;其中,得分最佳的目标旋转矩阵为符合第一平面集向第二平面集转换的最大几何一致性的旋转矩阵。
为实现上述目标及其他相关目标,本申请提供一种计算机设备,包括:处理器及存储器;所述存储器,用于存储计算机程序;所述处理器,用于运行所述计算机程序以实现平面几何一致性检测方法;其中,每个平面在一坐标系下通过其法向量、及平面所在坐标系原点距该平面的距离唯一表示;所述方法用于获得由第一坐标系下各第一平面组成的第一平面集向由第二坐标系下各第二平面组成的第二平面集转换的一致的转换关系,所述转换关系通过目标旋转矩阵及目标平移矩阵表示的原理;所述平面几何一致性检测方法包括:获得由表示各第一平面的各第一法向量和各第一距离分别组成的第一法向量集及第一距离集、以及由表示每个第二平面的各第二法向量和各第二距离分别组成的第二法向量集及第二距离集;获取所述目标旋转矩阵,包括:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量,计算每对第一法向量向其夹角匹配的一对第二法向量转换的第一旋转矩阵,并获取每个所述第一旋转矩阵在特殊正交群SO(3)空间中的第一投影矩阵;通过罗德里格斯旋转公式的逆变换将得到的各第一投影矩阵转换为旋转向量,得到旋转向量集;对各旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类,将作为该第一目标分类中每个元素即旋转向量的计算依据的一对第一法向量和一对第二法向量分别置于一第一法向量序列和第二法向量序列中;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵,并获取每个所述第二旋转矩阵在特殊正交群SO(3)空间中的第二投影矩阵作为所述目标旋转矩阵;获取所述目标平移矩阵,包括:将第一距离集中的每个第一距离与第二距离集中的每个第二距离相减,得到各距离差并组成距离差集;将距离差集中的各元素进行聚类,以获得拥有最多元素的第二目标分类;根据作为第二目标分类中的每个元素即距离差的计算依据的第一距离和第二距离,结合与每个所述作为距离差计算依据的第二距离属于同一平面的第二法向量,计算得到所述目标平移矩阵。
于一实施例中,所述第一旋转矩阵是通过一对第一法向量连同垂直它们的另一第一法向量、与夹角匹配的一对第二法向量连同垂直它们的另一第二法向量间的映射关系得到的。
于一实施例中,所述每对第一法向量间不平行;平行的各第一平面的第一法向量、和平行的各第二平面的第二法向量设为同一个。
于一实施例中,所述夹角匹配指的是一对第一法向量间夹角和一对第二法向量间夹角的偏差在预设范围内。
于一实施例中,所述聚类包括:将第一目标分类或第二目标分类中的各元素插入哈希表中,并根据每个元素构建其对应索引。
于一实施例中,每个元素的索引是根据其包含数据、及其所能达到的最大取值来建立的。
于一实施例中,所述旋转向量表示为沿一旋转轴向量方向延伸并具有表示其旋转角度的长度的向量;插入旋转向量的第一哈希表的构建方式包括:将三维角度空间分为多个尺寸为∈ 3的立方体形式的单元格,∈∈(0,2π);建立尺寸为N 3的三维哈希表作为第一哈希表,
Figure PCTCN2019081424-appb-000006
且通过计算索引
Figure PCTCN2019081424-appb-000007
以将旋转向量集中各旋转向量插入各单元格中;其中,r为旋转向量。
于一实施例中,插入距离差的第二哈希表的构建方式包括:将三维空间分为多个尺寸为
Figure PCTCN2019081424-appb-000008
的立方体形式的单元格,th d∈(0,maxD);建立尺寸为N 3的三维哈希表作为第二哈希表,
Figure PCTCN2019081424-appb-000009
且通过计算索引
Figure PCTCN2019081424-appb-000010
将距离差集中的各距离差插入第二哈希表中;其中,d为距离差,maxD为最大距离,th d为单元格的尺寸阈值。
于一实施例中,所述第一/第二投影矩阵是通过对第一/第二旋转矩阵进行奇异值分解得到的两个正交矩阵计算得到的。
于一实施例中,所述方法还包括:计算所获取的多个目标旋转矩阵的得分,其中,该得分相关于用于计算目标旋转矩阵的互不平行的各平面组的数量、以及构成每个平面组的平行平面的数量;其中,得分最佳的目标旋转矩阵为符合第一平面集向第二平面集转换的最大几何一致性的旋转矩阵。
为实现上述目标及其他相关目标,本申请提供一种计算机可读存储介质,存储有计算机程序,所述计算机程序被处理器运行时执行所述的方法。
如上所述,本申请的平面几何一致性检测方法、计算机设备、及存储介质,实现获得由第一法向量集、第一距离集、第二法向量集、及第二距离集;获取所述目标旋转矩阵,包括:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量,计算第一旋转矩阵及其第一投影矩阵;将得到的第一投影矩阵转换为旋转向量,得到旋转向量集;对各旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类中元素得到第一法向量序列和第二法向量序列;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵并以其第二投影矩阵作为所述目标旋转矩阵;另外,通过第一、第二距离集得到距离差集,聚类得到第二目标分类并以其中元素的第一、第二距离和对应第二法向量,计算得到所述目标平移矩阵;本申请的方案实现高一致性且快速的不同坐标系下平面映射。
附图说明
图1显示为本申请实施例中的平面几何一致性检测方法的流程示意图。
图2显示为本申请实施例中的计算机设备的结构示意图。
具体实施方式
以下通过特定的具体实例说明本申请的实施方式,本领域技术人员可由本说明书所揭露的内容轻易地了解本申请的其他优点与功效。本申请还可以通过另外不同的具体实施方式加以实施或应用,本说明书中的各项细节也可以基于不同观点与应用,在没有背离本申请的精神下进行各种修饰或改变。需说明的是,在不冲突的情况下,本申请中的实施例及实施例中的特征可以相互组合。
假设第一坐标系F A下的第一平面集表示为
Figure PCTCN2019081424-appb-000011
其中包含的平面
Figure PCTCN2019081424-appb-000012
可以由其法向量
Figure PCTCN2019081424-appb-000013
和从第一坐标系原点到平面
Figure PCTCN2019081424-appb-000014
的距离唯一地表示,使得位于
Figure PCTCN2019081424-appb-000015
上的点p A满足:
Figure PCTCN2019081424-appb-000016
and
Figure PCTCN2019081424-appb-000017
F A和F B之间刚体运动的转换关系满足f BA=(R BA,T BA),R BA为F A转换到F B的旋转矩阵,T BA为平移矩阵;相应的,在F A和F B观察到的物理平面的平面参数满足:
Figure PCTCN2019081424-appb-000018
Figure PCTCN2019081424-appb-000019
其中,式(3)的推导过程为:
Figure PCTCN2019081424-appb-000020
同样,F B中的每个平面
Figure PCTCN2019081424-appb-000021
可以由满足上式子的
Figure PCTCN2019081424-appb-000022
唯一定义。
因此,对应两个平面集
Figure PCTCN2019081424-appb-000023
Figure PCTCN2019081424-appb-000024
要实现在它们之间建立最大几何一致性的转换关系的问题,可以转化为以下的约束优化问题:
Figure PCTCN2019081424-appb-000025
从而找寻刚体映射关系f BA以令以下定义的集合
Figure PCTCN2019081424-appb-000026
的基数最大化。
Figure PCTCN2019081424-appb-000027
如图1所示,展示本申请实施例中的平面几何一致性检测方法的流程示意图。实际上,该方法也就是用于获取F A和F B间刚体运动映射的最大一致性的R BA,T BA
所述方法包括:
步骤S101:获得由表示各第一平面的各第一法向量和各第一距离分别组成的第一法向量集及第一距离集、以及由表示每个第二平面的各第二法向量和各第二距离分别组成的第二法向量集及第二距离集。
承前所述,设第一平面表示为
Figure PCTCN2019081424-appb-000028
相应的第一平面集表示为
Figure PCTCN2019081424-appb-000029
第二平面表示为
Figure PCTCN2019081424-appb-000030
相应的第二平面集表示为
Figure PCTCN2019081424-appb-000031
第一法向量表示为
Figure PCTCN2019081424-appb-000032
第二法向量表示为
Figure PCTCN2019081424-appb-000033
步骤S102:获取所述目标旋转矩阵R BA
在一实施例中,利用刚体运动的特点,即保持向量的范数和交积,来计算两坐标系下平面集之间的旋转矩阵。换言之,通过从平面之间的角度生成旋转建议(proposal),我们能够从中提取最大旋转一致性。为此,首先创建轮换建议,并且通过忽略不可能的旋转建议并采用如下所述的基于投票(例如聚类)的测量机制,能有效地测量一致性。
相应的,步骤S102具体可以包括:
步骤S121:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量。
在一实施例中,假设给定大小为N A的第一平面集
Figure PCTCN2019081424-appb-000034
根据成对第一法向量间的夹角,则存在
Figure PCTCN2019081424-appb-000035
个夹角θ i,j
计算为:
Figure PCTCN2019081424-appb-000036
从而,根据计算每个夹角构成对应第一平面集的第一角组PG A
Figure PCTCN2019081424-appb-000037
and
Figure PCTCN2019081424-appb-000038
从式(8)可知,n i和n j是F A下不同平面的法向量。
相似的,设θ p,q是属于第二平面集
Figure PCTCN2019081424-appb-000039
中一平面的第二法向量n p和另一平面的第二法向量n q间的夹角,可以得到基于
Figure PCTCN2019081424-appb-000040
Figure PCTCN2019081424-appb-000041
个θ p,q,构成第二角组PG B
步骤S122:计算每对第一法向量向其夹角匹配的一对第二法向量转换的第一旋转矩阵,并获取每个所述第一旋转矩阵在特殊正交群SO(3)空间中的第一投影矩阵。
具体的,利用两个平面集的成对角度组PG A、PG B间存在角度匹配的关系来找到平面集之间的映射关系。即在对PG A的角θ i,j和PG B的θ p,q间,利用两组法线生成旋转方案的最小解为:
Figure PCTCN2019081424-appb-000042
Figure PCTCN2019081424-appb-000043
Figure PCTCN2019081424-appb-000044
Figure PCTCN2019081424-appb-000045
R BA=[U·V′](13)
其中,n k是根据n i和n j构建的垂直于它们的一个向量,n r是由n p和n q构建的垂直于它们的一个向量;构建原因是因为旋转矩阵是3×3的矩阵,而每个法向量是3×1的向量,在有了n i和n j,n p和n q之后,还需要分别再得到一个对应坐标系下的位置相关的向量(可以是再一个法向量)才能构建3×3的矩阵,而本实施例中采用获取每对法向量的垂直向量的方式,属于计算量较小的R BA极小代数解。
另外,在本实施例中,由于R BA是标准正交矩阵,落在李群SO(3),而
Figure PCTCN2019081424-appb-000046
未必在SO(3)中,因此,本实施例中通过SVD方法将
Figure PCTCN2019081424-appb-000047
投影到SO(3)中,得到R BA。由于进行SVD分解后U,V一定是正交矩阵,所以R BA也是正交矩阵。
此处获得的R BA,只是一个初步的结果,其的计算方法能在在计算量上得到优化,且能得到进一步的一致性优化。
具体来讲,设N=max{N A,N B}。总共存在
Figure PCTCN2019081424-appb-000048
可能方案。在此,我们提出了一个有效的测量机制,在平均情况下,旋转方案数显著减少到N 2
由于刚体运动的特点,在F A和F B中保持了两个平面之间的夹角。因此,不需要遍历这两个角度组的所有元素。此外,平行平面被忽略,因为它们具有相似的法向量对旋转建议集没有贡献,例如所选取的成对第一法向量间不平行,相应夹角匹配的一对第二法向量也就不会 平行;而平行的多个第一法向量或平行的多个第二法向量在参与旋转方案的计算时可以视为同一个。
可选的,引入阈值thθ>0作为可接受的角度噪声阈值,建立对所述匹配角度的判别标准;使得对于θ i,j∈PG A,角度值在(θ i,j-thθ,θ i,j+thθ)范围内的PG B中的θ p,q被认为是θ i,j。因此,通过搜索算法,例如二进制搜索,θ i,j的可能匹配的查找时间将减少到O(logN 2);此外,在对应时间复杂度O(N 4)的最坏情况下,例如大多数平面彼此垂直,找到PG A和PG B之间的所有成对匹配角的总时间复杂度将平均减少到O(N 2logN)。
然后,将匹配集M定义为下式(14)~(15):
Figure PCTCN2019081424-appb-000049
θ i,j∈PG Ap,q∈PG B   (15)
而关于进一步的一致性优化,见以下步骤:
步骤S123:通过罗德里格斯旋转公式的逆变换将得到的各第一投影矩阵转换为旋转向量,得到旋转向量集;
步骤S124:对各旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类,将作为该第一目标分类中每个元素即旋转向量的计算依据的一对第一法向量和一对第二法向量分别置于一第一法向量序列和第二法向量序列中;
步骤S125:计算第一法向量序列向第二法向量序列转换的第二旋转矩阵,并获取每个所述第二旋转矩阵在特殊正交群SO(3)空间中的第二投影矩阵作为所述目标旋转矩阵。
举例说明,为了减小搜索空间,根据Rodrigues公式的逆,将R变换为以旋转轴(向量)为方向、以角度为长度的三维旋转矢量,从而得到通过轴和角的更紧凑的R的表示。
SO(3)中的旋转矩阵R应该满足R TR=RR T=I(I为单位矩阵)和det(R)=1,这是正交矩阵的特性,并且旋转向量r可以计算如下:
Figure PCTCN2019081424-appb-000050
Figure PCTCN2019081424-appb-000051
其中,为旋转轴的向量表示,A根据罗德里格斯公式可以推导得到,r 11+r 22+r 33是旋转矩阵对角线上的分量;由于cosθ=(r 11+r 22+r 33-1)/2,因此c即为cosθ;而ρ的模即为sinθ,因此,s=sinθ。
如果s=0且c=1,则r=0;如果s=0且c=-1,则令v等于R+I的一个非0列。
然后,得到
Figure PCTCN2019081424-appb-000052
在本实施例中,还涉及S功能来翻转r的坐标符号以确保其唯一性;
Figure PCTCN2019081424-appb-000053
更普遍的是,if s≠0,则得到:
Figure PCTCN2019081424-appb-000054
θ=arctan(s,c),andr=uθ(20)
从而将匹配集M转换为与r相关:
Figure PCTCN2019081424-appb-000055
即所述旋转向量集。
现在仍然需要找到最一致的旋转矩阵,可以采用聚类的原理;据此,有许多有效的方法来提取这种最大一致性,例如使用哈希表或执行二进制搜索来对各旋转向量进行聚类,取其中含有元素(即旋转向量)最多的分类进行目标旋转矩阵的重建。
在一实施例中,可以选用哈希表来解决这个问题。
为了对旋转矢量进行聚类分析,将三维角度空间划分为多个大小为∈ 3的立方体单元来存储M的旋转矢量,其中∈∈(0,2π)。因此,生成大小为N 3的第一哈希表,其中N由每个旋转轴上的中的单元格数给出。
建立尺寸为N 3的三维哈希表作为第一哈希表,其中:
Figure PCTCN2019081424-appb-000056
且通过计算索引:
Figure PCTCN2019081424-appb-000057
以将旋转向量集中各旋转向量插入各单元格中;其中,r为旋转向量。
在得到哈希表之后,具有最大元素数n c的单元格Cr(即第一目标分类),根据公式(2)再计算出
Figure PCTCN2019081424-appb-000058
Figure PCTCN2019081424-appb-000059
之间的最佳旋转方案,作为目标旋转矩阵。
具体的,根据Cr,我们得到大小为3×n c的V A和V B。具体的,由于每个法向量为3×1的向量,则式(24),(25)中的V A和V B均为3*n c的矩阵;从而在(26)中,与式(2)和(11)原理相同的是,根据
Figure PCTCN2019081424-appb-000060
变形以求V A到V B的旋转矩阵
Figure PCTCN2019081424-appb-000061
再利用原理类似于公式(12)、(13)的公式(27)、(28)得到目标旋转矩阵R BA
Figure PCTCN2019081424-appb-000062
Figure PCTCN2019081424-appb-000063
Figure PCTCN2019081424-appb-000064
Figure PCTCN2019081424-appb-000065
R BA=[U·V′](28)
步骤S103:获取所述平移矩阵T BA
具体的,对于每个刚性物体,在刚体运动下,其上任意两点之间的距离不会随时间变化。通过比较两个平面集的距离集
Figure PCTCN2019081424-appb-000066
我们可以发现满足大多数平面关系的距离有一定的差异。因此,步骤S103的获取F A和F B之间的平移矩阵T BA可以从下面详细描述的流程导出:
步骤S131:将第一距离集中的每个第一距离与第二距离集中的每个第二距离相减,得到各距离差并组成距离差集。
步骤S132:将距离差集中的各元素进行聚类,以获得拥有最多元素的第二目标分类。
步骤S133:根据作为第二目标分类中的每个元素即距离差的计算依据的第一距离和第二距离,结合与每个所述作为距离差计算依据的第二距离属于同一平面的第二法向量,计算得到所述目标平移矩阵。
举例来说,定义距离差集为
Figure PCTCN2019081424-appb-000067
设第一距离表示为
Figure PCTCN2019081424-appb-000068
第二距离表示为
Figure PCTCN2019081424-appb-000069
Figure PCTCN2019081424-appb-000070
中的元素即距离差表示为:
Figure PCTCN2019081424-appb-000071
与前述实施例中相似的是,也可以将距离差集中的各个距离差插入一个第二哈希表中。
具体的,根据所计算的索引将
Figure PCTCN2019081424-appb-000072
中的所有元素(即距离差)插入第二哈希表中;所述索引的计算方法包括:
Figure PCTCN2019081424-appb-000073
其中,maxD和th d分别表示所定义的最大距离和存放元素的单元格的尺寸阈值,所述单元格可以是。最后,选择包含最多元素的单元格C d(即第二目标分类)来计算平移矩阵T BA
进一步的,从C d取出至少3个元素{d i1,j1,d i2,j2,…,d im,jm},根据式(3),得到:
Figure PCTCN2019081424-appb-000074
变形为:
Figure PCTCN2019081424-appb-000075
由于d im,jm可检测而为已知,而对应与
Figure PCTCN2019081424-appb-000076
同属于一个平面的第二法向量可以检测,令其转置矩阵
Figure PCTCN2019081424-appb-000077
也可以求得,即可求得T BA
基于上述具有
Figure PCTCN2019081424-appb-000078
个单元格的第二哈希表及O(N 2)个平移方案(proposal)。因此,上述平移矩阵的获取过程的时间复杂度为O(N 2)。
此外,如前述实施例所述,对于互相平行的平面,因为它们的法向量是平行的,所以我们把互相平行的平面(例如各平行的第一平面,各平行的第二平面)看成一个平面组plane group。这样在计算旋转方案(proposal)的时候每一planegroup相当与只参与了一次计算,从而能减少计算量。
然而,在哈希表聚类分析后,会对的每一个目标旋转矩阵会投票计算一个得分(score),以选出得分最高的目标旋转矩阵作为最佳一致性的旋转矩阵;该score是定义为相关于用于计算对应目标旋转矩阵的所有平面的个数,所有平面即包括F A下的第一平面集
Figure PCTCN2019081424-appb-000079
中的所有第一平面
Figure PCTCN2019081424-appb-000080
以及F B下第二平面集
Figure PCTCN2019081424-appb-000081
中的所有第二平面
Figure PCTCN2019081424-appb-000082
由于之前计算第一旋转矩阵的时候忽略了平行的平面,因此,在计算score的时候要重新加回原来被忽略掉的平行的平面。
举例来讲,设目标旋转矩阵R i是基于m个互不平行的平面组(plane group)计算得到的,其中每个平面组有r j个平行的平面,其中,各个平面组可以是一组平行的
Figure PCTCN2019081424-appb-000083
构成,也可以是一组平行的
Figure PCTCN2019081424-appb-000084
构成;R i的score相关于m和r j
在一实施例中,R i的score计算方式可以是:
Figure PCTCN2019081424-appb-000085
假设:
Figure PCTCN2019081424-appb-000086
由于每个平面组由n个平面组成,则根据前述平移矩阵T BA获取的内容,会需要O(n 2)时间来找到匹配的平面。因此,计算旋转矩阵R i的分数的总时间为
Figure PCTCN2019081424-appb-000087
因此,计算最佳k次旋转的所有分数总共需要O(k*N 2)时间。
然后,具有最佳分数的旋转矩阵表示最大几何一致性。
如图2所示,展示实施例中计算机设备200的结构示意图。
所述计算机设备200包括:存储器201及处理器202。
所述存储器,用于存储计算机程序;
所述处理器,用于运行所述计算机程序,以执行例如图1实施例中的平面几何一致性检测方法,即具体包括:
获得由表示各第一平面的各第一法向量和各第一距离分别组成的第一法向量集及第一距离集、以及由表示每个第二平面的各第二法向量和各第二距离分别组成的第二法向量集及第二距离集;获取所述目标旋转矩阵,包括:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量,计算每对第一法向量向其夹角匹配的一对第二法向量转换的第一旋转矩阵,并获取每个所述第一旋转矩阵在特殊正交群SO(3)空间中的第一投影矩阵;通过罗德里格斯旋转公式的逆变换将得到的各第一投影矩阵转换为旋转向量,得到旋转向量集;对各旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类,将作为该第一目标分类中每个元素即旋转向量的计算依据的一对第一法向量和一对第二法向量分别置于一第一法向量序列和第二法向量序列中;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵,并获取每个所述第二旋转矩阵在特殊正交群SO(3)空间中的第二投影矩阵作为所述目标旋转矩阵;获取所述目标平移矩阵,包括:将第一距离集中的每个第一距离与第二距离集中的每个第二距离相减,得到各距离差并组成距离差集;将距离差集中的各元素进行聚类,以获得拥有最多元素的第二目标分类;根据作为第二目标分类中的每个元素即距离差的计算依据的第一距离和第二距离,结合与每个所述作为距离差计算依据的第二距离属于同一平面的第二法向量,计算得到所述目标平移矩阵。
于一实施例中,所述每对第一法向量间不平行;平行的各第一平面的第一法向量、和平行的各第二平面的第二法向量设为同一个。
于一实施例中,所述夹角匹配指的是一对第一法向量间夹角和一对第二法向量间夹角的偏差在预设范围内。
于一实施例中,所述聚类包括:将第一目标分类或第二目标分类中的各元素插入哈希表中,并根据每个元素构建其对应索引。
于一实施例中,每个元素的索引是根据其包含数据、及其所能达到的最大取值来建立的。
于一实施例中,所述旋转向量表示为沿一旋转轴向量方向延伸并具有表示其旋转角度的长度的向量;插入旋转向量的第一哈希表的构建方式包括:将三维角度空间分为多个尺寸为∈3 的立方体形式的单元格,∈∈(0,2π);建立尺寸为N 3的三维哈希表作为第一哈希表,
Figure PCTCN2019081424-appb-000088
且通过计算索引
Figure PCTCN2019081424-appb-000089
以将旋转向量集中各旋转向量插入各单元格中;其中,r为旋转向量。
于一实施例中,插入距离差的第二哈希表的构建方式包括:将三维空间分为多个尺寸为thd 3的立方体形式的单元格,th d∈(0,maxD);建立尺寸为N 3的三维哈希表作为第二哈希表,
Figure PCTCN2019081424-appb-000090
且通过计算索引
Figure PCTCN2019081424-appb-000091
将距离差集中的各距离差插入第二哈希表中;其中,d为距离差,maxD为最大距离,th d为单元格的尺寸阈值。
于一实施例中,所述第一/第二投影矩阵是通过对第一/第二旋转矩阵进行奇异值分解得到的两个正交矩阵计算得到的,“/”表示和/或,也就是说,所述第一投影矩阵是通过对第一旋转矩阵进行奇异值分解得到的两个正交矩阵计算得到的,和/或,所述第二投影矩阵是通过对第二旋转矩阵进行奇异值分解得到的两个正交矩阵计算得到的。
于一实施例中,所述方法还包括:计算所获取的多个目标旋转矩阵的得分,其中,该得分相关于用于计算目标旋转矩阵的互不平行的各平面组的数量、以及构成每个平面组的平行平面的数量;其中,得分最佳的目标旋转矩阵为符合第一平面集向第二平面集转换的最大几何一致性的旋转矩阵。
在一些实施例中,所述存储器201可能包括但不限于高速随机存取存储器、非易失性存储器。例如一个或多个磁盘存储设备、闪存设备或其他非易失性固态存储设备。
所述处理器202可以是通用处理器,包括中央处理器(CentralProcessingUnit,简称CPU)、网络处理器(NetworkProcessor,简称NP)等;还可以是数字信号处理器(DigitalSignalProcessing,简称DSP)、专用集成电路(ApplicationSpecificIntegratedCircuit,简称ASIC)、现场可编程门阵列(Field-ProgrammableGateArray,简称FPGA)或者其他可编程逻辑器件、分立门或者晶体管逻辑器件、分立硬件组件。
另外,前述方法实施例(如图1中的方法实施例)中所涉及的各种计算机程序可以装载在计算机可读存储介质中,所述计算机可读存储介质可包括,但不限于,软盘、光盘、CD-ROM(紧致盘-只读存储器)、磁光盘、ROM(只读存储器)、RAM(随机存取存储器)、EPROM(可擦除可编程只读存储器)、EEPROM(电可擦除可编程只读存储器)、磁卡或光卡、闪存、或适于存储机器可执行指令的其他类型的介质/机器可读介质。所述计算机可读存储介质可以是未接入计算机设备的产品,也可以是已接入计算机设备使用的部件。
在具体实现上,所述计算机程序为执行特定任务或实现特定抽象数据类型的例程、程序、对象、组件、数据结构等等。
综上所述,本申请的平面几何一致性检测方法、计算机设备、及存储介质,实现获得由第一法向量集、第一距离集、第二法向量集、及第二距离集;获取所述目标旋转矩阵,包括:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量,计算第一旋转矩阵及其第一投影矩阵;将得到的第一投影矩阵转换为旋转向量,得到旋转向量集;对各旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类中元素得到第一法向量序列和第二法向量序列;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵并以其第二投影矩阵作为所述目标旋转矩阵;另外,通过第一、第二距离集得到距离差集,聚类得到第二目标分类并以其中元素的第一、第二距离和对应第二法向量,计算得到所述目标平移矩阵;本申请的方案实现高一致性且快速实时的不同坐标系下平面映射。
上述实施例仅例示性说明本申请的原理及其功效,而非用于限制本申请。任何熟悉此技术的人士皆可在不违背本申请的精神及范畴下,对上述实施例进行修饰或改变。因此,举凡所属技术领域中具有通常知识者在未脱离本申请所揭示的精神与技术思想下所完成的一切等效修饰或改变,仍应由本申请的权利要求所涵盖。

Claims (21)

  1. 一种平面几何一致性检测方法,其特征在于,每个平面在一坐标系下通过其法向量、及平面所在坐标系原点距该平面的距离唯一表示;所述方法用于获得由第一坐标系下各第一平面组成的第一平面集向由第二坐标系下各第二平面组成的第二平面集转换的一致的转换关系,所述转换关系通过目标旋转矩阵及目标平移矩阵表示;所述方法包括:
    获得由表示各第一平面的各第一法向量和各第一距离分别组成的第一法向量集及第一距离集、以及由表示每个第二平面的各第二法向量和各第二距离分别组成的第二法向量集及第二距离集;
    获取所述目标旋转矩阵,包括:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量,计算每对第一法向量向其夹角匹配的一对第二法向量转换的第一旋转矩阵,并获取每个所述第一旋转矩阵在特殊正交群SO(3)空间中的第一投影矩阵;通过罗德里格斯旋转公式的逆变换将得到的各第一投影矩阵转换为旋转向量,得到旋转向量集;对旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类,将作为该第一目标分类中每个元素即旋转向量的计算依据的一对第一法向量和一对第二法向量分别置于一第一法向量序列和第二法向量序列中;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵,并获取每个所述第二旋转矩阵在特殊正交群SO(3)空间中的第二投影矩阵作为所述目标旋转矩阵;
    获取所述目标平移矩阵,包括:将第一距离集中的每个第一距离与第二距离集中的每个第二距离相减,得到各距离差并组成距离差集;将距离差集中的各元素进行聚类,以获得拥有最多元素的第二目标分类;根据作为第二目标分类中的每个元素即距离差的计算依据的第一距离和第二距离,结合与每个所述作为距离差计算依据的第二距离属于同一平面的第二法向量,计算得到所述目标平移矩阵。
  2. 根据权利要求1所述的方法,其特征在于,所述第一旋转矩阵是通过一对第一法向量连同垂直它们的另一第一法向量、与夹角匹配的一对第二法向量连同垂直它们的另一第二法向量间的映射关系得到的。
  3. 根据权利要求1所述的方法,其特征在于,所述每对第一法向量间不平行;平行的各第一平面的第一法向量、和平行的各第二平面的第二法向量设为同一个。
  4. 根据权利要求1或3所述的方法,其特征在于,所述夹角匹配指的是一对第一法向量间夹角和一对第二法向量间夹角的偏差在预设范围内。
  5. 根据权利要求1所述的方法,其特征在于,所述聚类方式包括:将第一目标分类或第 二目标分类中的各元素插入哈希表中,并根据每个元素构建其对应索引。
  6. 根据权利要求5所述的方法,其特征在于,每个元素的索引是根据其包含数据、及其所能达到的最大取值来建立的。
  7. 根据权利要求6所述的方法,其特征在于,所述旋转向量表示为沿一旋转轴向量方向延伸并具有表示其旋转角度的长度的向量;插入旋转向量的第一哈希表的构建方式包括:
    将三维角度空间分为多个尺寸为∈ 3的立方体形式的单元格,∈∈(0,2π);
    建立尺寸为N 3的三维哈希表作为第一哈希表,
    Figure PCTCN2019081424-appb-100001
    且通过计算索引
    Figure PCTCN2019081424-appb-100002
    以将旋转向量集中各旋转向量插入各单元格中;其中,r为旋转向量。
  8. 根据权利要求6所述的方法,其特征在于,插入距离差的第二哈希表的构建方式包括:
    将三维空间分为多个尺寸为
    Figure PCTCN2019081424-appb-100003
    的立方体形式的单元格,th d∈(0,maxD);
    建立尺寸为N 3的三维哈希表作为第二哈希表,
    Figure PCTCN2019081424-appb-100004
    且通过计算索引
    Figure PCTCN2019081424-appb-100005
    将距离差集中的各距离差插入第二哈希表中;其中,d为距离差,maxD为最大距离,th d为单元格的尺寸阈值。
  9. 根据权利要求1所述的方法,其特征在于,所述第一/第二投影矩阵是通过对第一/第二旋转矩阵进行奇异值分解得到的两个正交矩阵计算得到的。
  10. 根据权利要求3所述的方法,其特征在于,还包括:
    计算所获取的多个目标旋转矩阵的得分,其中,该得分相关于用于计算目标旋转矩阵的互不平行的各平面组的数量、以及构成每个平面组的平行平面的数量;
    其中,得分最佳的目标旋转矩阵为符合第一平面集向第二平面集转换的最大几何一致性的旋转矩阵。
  11. 一种计算机设备,其特征在于,包括:处理器及存储器;
    所述存储器,用于存储计算机程序;
    所述处理器,用于运行所述计算机程序以实现平面几何一致性检测方法;其中,每个平面在一坐标系下通过其法向量、及平面所在坐标系原点距该平面的距离唯一表示;所述方法用于获得由第一坐标系下各第一平面组成的第一平面集向由第二坐标系下各第二平面组成的第二平面集转换的一致的转换关系,所述转换关系通过目标旋转矩阵及目标平移矩阵表示的原理;
    所述平面几何一致性检测方法包括:
    获得由表示各第一平面的各第一法向量和各第一距离分别组成的第一法向量集及第一距离集、以及由表示每个第二平面的各第二法向量和各第二距离分别组成的第二法向量集及第二距离集;
    获取所述目标旋转矩阵,包括:获取第一法向量集中的每对第一法向量、和该对第一法向量在第二法向量集中夹角匹配的一对第二法向量,计算每对第一法向量向其夹角匹配的一对第二法向量转换的第一旋转矩阵,并获取每个所述第一旋转矩阵在特殊正交群SO(3)空间中的第一投影矩阵;通过罗德里格斯旋转公式的逆变换将得到的各第一投影矩阵转换为旋转向量,得到旋转向量集;对各旋转向量集中的各元素进行聚类,以获得拥有最多元素的第一目标分类,将作为该第一目标分类中每个元素即旋转向量的计算依据的一对第一法向量和一对第二法向量分别置于一第一法向量序列和第二法向量序列中;计算第一法向量序列向第二法向量序列转换的第二旋转矩阵,并获取每个所述第二旋转矩阵在特殊正交群SO(3)空间中的第二投影矩阵作为所述目标旋转矩阵;
    获取所述目标平移矩阵,包括:将第一距离集中的每个第一距离与第二距离集中的每个第二距离相减,得到各距离差并组成距离差集;将距离差集中的各元素进行聚类,以获得拥有最多元素的第二目标分类;根据作为第二目标分类中的每个元素即距离差的计算依据的第一距离和第二距离,结合与每个所述作为距离差计算依据的第二距离属于同一平面的第二法向量,计算得到所述目标平移矩阵。
  12. 根据权利要求11所述的计算机设备,其特征在于,所述第一旋转矩阵是通过一对第一法向量连同垂直它们的另一第一法向量、与夹角匹配的一对第二法向量连同垂直它们的另一第二法向量间的映射关系得到的。
  13. 根据权利要求11所述的计算机设备,其特征在于,所述每对第一法向量间不平行;平行的各第一平面的第一法向量、和平行的各第二平面的第二法向量设为同一个。
  14. 根据权利要求11或13所述的计算机设备,其特征在于,所述夹角匹配指的是一对第一法向量间夹角和一对第二法向量间夹角的偏差在预设范围内。
  15. 根据权利要求11所述的计算机设备,其特征在于,所述聚类包括:将第一目标分类或第二目标分类中的各元素插入哈希表中,并根据每个元素构建其对应索引。
  16. 根据权利要求15所述的计算机设备,其特征在于,每个元素的索引是根据其包含数据、及其所能达到的最大取值来建立的。
  17. 根据权利要求16所述的计算机设备,其特征在于,所述旋转向量表示为沿一旋转轴向量方向延伸并具有表示其旋转角度的长度的向量;插入旋转向量的第一哈希表的构建 方式包括:
    将三维角度空间分为多个尺寸为∈ 3的立方体形式的单元格,∈∈(0,2π);
    建立尺寸为N 3的三维哈希表作为第一哈希表,
    Figure PCTCN2019081424-appb-100006
    且通过计算索引
    Figure PCTCN2019081424-appb-100007
    以将旋转向量集中各旋转向量插入各单元格中;其中,r为旋转向量。
  18. 根据权利要求16所述的计算机设备,其特征在于,插入距离差的第二哈希表的构建方式包括:
    将三维空间分为多个尺寸为th d 3的立方体形式的单元格,th d∈(0,maxD);
    建立尺寸为N 3的三维哈希表作为第二哈希表,
    Figure PCTCN2019081424-appb-100008
    且通过计算索引
    Figure PCTCN2019081424-appb-100009
    将距离差集中的各距离差插入第二哈希表中;其中,d为距离差,maxD为最大距离,
    Figure PCTCN2019081424-appb-100010
    为单元格的尺寸阈值。
  19. 根据权利要求11所述的计算机设备,其特征在于,所述第一/第二投影矩阵是通过对第一/第二旋转矩阵进行奇异值分解得到的两个正交矩阵计算得到的。
  20. 根据权利要求13所述的计算机设备,其特征在于,所述方法还包括:
    计算所获取的多个目标旋转矩阵的得分,其中,该得分相关于用于计算目标旋转矩阵的互不平行的各平面组的数量、以及构成每个平面组的平行平面的数量;
    其中,得分最佳的目标旋转矩阵为符合第一平面集向第二平面集转换的最大几何一致性的旋转矩阵。
  21. 一种计算机可读存储介质,其特征在于,存储有计算机程序,所述计算机程序被处理器运行时执行如权利要求1至10中任一项所述的方法。
PCT/CN2019/081424 2019-02-19 2019-04-04 平面几何一致性检测方法、计算机设备、及存储介质 WO2020168620A1 (zh)

Priority Applications (1)

Application Number Priority Date Filing Date Title
US17/383,419 US20210357477A1 (en) 2019-02-19 2021-07-22 Plane-geometry consistency detection method, computer device, storage medium

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN201910125336.3A CN111582293B (zh) 2019-02-19 2019-02-19 平面几何一致性检测方法、计算机设备、及存储介质
CN201910125336.3 2019-02-19

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US17/383,419 Continuation US20210357477A1 (en) 2019-02-19 2021-07-22 Plane-geometry consistency detection method, computer device, storage medium

Publications (1)

Publication Number Publication Date
WO2020168620A1 true WO2020168620A1 (zh) 2020-08-27

Family

ID=72126053

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2019/081424 WO2020168620A1 (zh) 2019-02-19 2019-04-04 平面几何一致性检测方法、计算机设备、及存储介质

Country Status (3)

Country Link
US (1) US20210357477A1 (zh)
CN (1) CN111582293B (zh)
WO (1) WO2020168620A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115496865B (zh) * 2022-11-18 2023-03-14 南京智程信息科技有限公司 相似模型查找方法、装置及存储介质

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104143210A (zh) * 2014-07-31 2014-11-12 哈尔滨工程大学 一种多尺度法向特征点云配准方法
CN104395932A (zh) * 2012-06-29 2015-03-04 三菱电机株式会社 对准数据的方法
US20150085979A1 (en) * 2012-05-01 2015-03-26 Universitat Bern Image distortion correction and robust phantom detection
CN105488535A (zh) * 2015-12-05 2016-04-13 乔付 一种三维点云匹配方法
CN106871928A (zh) * 2017-01-18 2017-06-20 北京工业大学 基于李群滤波的捷联惯性导航初始对准方法
CN108779984A (zh) * 2016-03-16 2018-11-09 索尼公司 信号处理设备和信号处理方法

Family Cites Families (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9875542B2 (en) * 2012-01-23 2018-01-23 Nec Corporation Camera calibration device, camera calibration method, and camera calibration program
GB2559157A (en) * 2017-01-27 2018-08-01 Ucl Business Plc Apparatus, method and system for alignment of 3D datasets
CN107564089B (zh) * 2017-08-10 2022-03-01 腾讯科技(深圳)有限公司 三维图像处理方法、装置、存储介质和计算机设备
CN107680159B (zh) * 2017-10-16 2020-12-08 西北工业大学 一种基于投影矩阵的空间非合作目标三维重建方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20150085979A1 (en) * 2012-05-01 2015-03-26 Universitat Bern Image distortion correction and robust phantom detection
CN104395932A (zh) * 2012-06-29 2015-03-04 三菱电机株式会社 对准数据的方法
CN104143210A (zh) * 2014-07-31 2014-11-12 哈尔滨工程大学 一种多尺度法向特征点云配准方法
CN105488535A (zh) * 2015-12-05 2016-04-13 乔付 一种三维点云匹配方法
CN108779984A (zh) * 2016-03-16 2018-11-09 索尼公司 信号处理设备和信号处理方法
CN106871928A (zh) * 2017-01-18 2017-06-20 北京工业大学 基于李群滤波的捷联惯性导航初始对准方法

Also Published As

Publication number Publication date
US20210357477A1 (en) 2021-11-18
CN111582293B (zh) 2023-03-24
CN111582293A (zh) 2020-08-25

Similar Documents

Publication Publication Date Title
Sattler et al. Image Retrieval for Image-Based Localization Revisited.
CN110705574B (zh) 定位方法及装置、设备、存储介质
Lee et al. Topology preserving relaxation labeling for nonrigid point matching
Choudhary et al. Visibility probability structure from sfm datasets and applications
Gardner et al. Measuring distance between unordered sets of different sizes
JP2010272092A (ja) 画像検索装置およびその方法
Cui et al. Progressive large-scale structure-from-motion with orthogonal MSTs
WO2020168620A1 (zh) 平面几何一致性检测方法、计算机设备、及存储介质
Dong et al. Probability driven approach for point cloud registration of indoor scene
Frahm et al. Geo-registered 3D models from crowdsourced image collections
CN112651408B (zh) 一种基于点对变换特征的三维局部表面描述方法和系统
Arandjelovic et al. Efficient image retrieval for 3D structures.
Maybank A probabilistic definition of salient regions for image matching
Marron et al. Speculation on the generality of the backward stepwise view of pca
JP4188632B2 (ja) 距離画像の統合方法及び距離画像統合装置
Zhang et al. Accelerating Globally Optimal Consensus Maximization in Geometric Vision
Wang et al. Robust image orientation based on relative rotations and tie points
Wu et al. Second-order configuration of local features for geometrically stable image matching and retrieval
Hornáček et al. Extracting vanishing points across multiple views
Yang et al. Estimating the fundamental matrix based on the end-to-end convolutional network
Zhang et al. A Visual SLAM Approach to Deeply Explore the Spatial Constraints in Indoor Environment Based on the Manhattan Hypothesis
Gherardi et al. Automatic camera orientation and structure recovery with Samantha
Xiao et al. Learning two-view stereo matching
US8867865B2 (en) Point set matching with outlier detection
Li et al. New environmental line feature-based vision navigation: Design and analysis

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19916331

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19916331

Country of ref document: EP

Kind code of ref document: A1