Improved point cloud registration method for complex curved surface component

Abstract

The invention provides an improved point cloud registration method for a complex curved surface component, and relates to the technical field of computer vision and robots. The method comprises the steps of fixing the position of the complex curved surface component, performing measurement from two different visual angles by using a binocular structured light measurement system, and collecting surface information of the complex curved surface component; de-noising the source point cloud and the target point cloud according to local statistical characteristics of each point and neighboring points around the point; screening feature points, extracting feature point features of the source point cloud and the target point cloud, and determining feature corresponding point pairs; performing point cloud coarse registration according to the feature corresponding point pairs; acquiring local information of each point in the source point cloud and the target point cloud; determining a corresponding point set according to the local information and bidirectional search; and constructing a target function, and setting an iteration termination condition according to the corresponding point set to complete point cloud precise registration. According to the invention, precision registration of point clouds of different visual angles of a complex curved surface component can be realized, and the speed and precision of point cloud registration are improved.

Description

Technical Field

[0001] The invention relates to the fields of computer vision and robotics technology, and in particular to an improved point cloud registration method for complex curved surface components. Background Art

[0002] Although point cloud registration can be achieved in current related technologies, there are defects such as insufficient robustness in data with a large amount of noise, outliers or asymmetry, high requirements for the initial pose, and easy to fall into local optimality, and poor robustness to outliers and noise. It is difficult to achieve accurate registration of point clouds from different perspectives of complex curved surface components, resulting in a decrease in the speed and accuracy of point cloud registration.

[0003] The information disclosed in the background technology section of this application is only intended to deepen the understanding of the general background technology of this application, and should not be regarded as an admission or any form of suggestion that the information constitutes the prior art already known to those skilled in the art. Summary of the invention

[0004] The present invention provides an improved point cloud registration method for complex curved surface components, which can solve the technical problem that it is difficult to achieve accurate registration of point clouds of complex curved surface components from different viewing angles.

[0005] According to a first aspect of the present invention, there is provided an improved point cloud registration method for complex curved surface components, comprising: fixing the position of the complex curved surface component, measuring from two different perspectives using a binocular structured light measurement system, and collecting surface information of the complex curved surface component, wherein the surface information is stored in the form of a point cloud model, with one of the two different perspectives being used as a source point cloud and the other being used as a target point cloud; denoising the source point cloud and the target point cloud according to the local statistical characteristics of each point and its surrounding neighboring points; screening feature points, extracting feature point features of the source point cloud and the target point cloud, and determining feature corresponding point pairs; performing point cloud coarse registration according to the feature corresponding point pairs; obtaining local information of each point in the source point cloud and the target point cloud, wherein the local information includes a normal, a curvature and a covariance matrix; determining a corresponding point set according to the local information and a two-way search; constructing an objective function, setting an iteration termination condition according to the corresponding point set, and completing point cloud fine registration.

[0006] According to the present invention, the source point cloud and the target point cloud are denoised according to the local statistical characteristics of each point and its surrounding neighboring points, including: establishing a kd-tree for each point cloud, querying the k neighboring points of each point in the respective point cloud, obtaining the least squares fitting plane of each point and its k neighboring points, wherein k is a positive integer; obtaining the distance from each point to the least squares fitting plane; setting a distance threshold; when the distance from a point to the least squares fitting plane exceeds the distance threshold, determining that the point is a noise point and removing it.

[0007] According to the present invention, feature points are screened, feature point features of the source point cloud and the target point cloud are extracted, and feature corresponding point pairs are determined, including: obtaining the arithmetic mean of the angles between the normal vector of each point and the normal vectors of its k nearest neighboring points; obtaining the average value μ and variance σ of the angles between the normal vectors of all points and the normal vectors of their k nearest neighboring points; taking points whose arithmetic mean values ​​are greater than μ-2σ and less than μ+2σ as feature points; obtaining an extended feature histogram of each feature point; determining feature point features based on the extended feature histogram; establishing a kd-tree for the feature point features, performing nearest neighbor search on the source point cloud on the target point cloud, and obtaining feature corresponding point pairs.

[0008] According to the present invention, determining the feature point features according to the extended feature histogram includes: according to the formula Determine the feature FPFH′(p) of feature point p, where SPFH′(p) is the extended feature histogram of feature point p, p si is the i-th neighbor point of feature point p, SPFH′(p si ) is the extended feature histogram of the i-th neighboring point of feature point p, ω i is the Euclidean distance from the feature point to the i-th neighbor point of the feature point, k is the number of neighbor points, i≤k, and i and k are both positive integers.

[0009] According to the present invention, a point cloud rough registration is performed based on the feature corresponding point pairs, including: setting a maximum number of iterations, randomly selecting 3 from the feature corresponding point pairs in each iteration, and obtaining a rigid body transformation matrix; obtaining the total average distance between the source point cloud and the target point cloud; according to the formula ||q i ′-Tp j ′||< Get a rigid distance constraint, where q i ′、p j ' is the corresponding point in the feature corresponding point pair, T is the rigid body transformation matrix, is the total average distance between the source point cloud and the target point cloud; determines whether the remaining feature corresponding point pairs satisfy the rigid distance constraint, and if so, they are the correct matching point pairs, wherein the correct matching point pairs are the inliers; when the repeated iterations reach the maximum number of iterations, the rigid body transformation matrix with the largest number of inliers is taken as the correct coarse registration.

[0010] According to the present invention, the corresponding point set is determined based on the local information and the bidirectional search, including: using kd-tree to distribute in the source point cloud and the target point cloud to search for the nearest neighbor of each point, respectively, to obtain the corresponding point set of the source point cloud and the corresponding point set of the target point cloud; in the corresponding point set of the source point cloud, for each group of corresponding points, p is taken out. i ″The corresponding q j″, search q again in the corresponding point set of the target point cloud j The corresponding p v ″, the corresponding points satisfying i=v are grouped into a candidate corresponding point set; the unit normal vector and the point curvature of each group of corresponding points in the candidate corresponding point set are obtained; according to each group of corresponding points in the candidate corresponding point set, the unit normal vector and the point curvature, the distance condition, the normal angle constraint and the curvature constraint are obtained; when the distance condition, the normal angle constraint and the curvature constraint are all satisfied, the corresponding point set is determined.

[0011] According to the present invention, according to each group of corresponding points in the candidate corresponding point set, the unit normal vector and the point curvature, a distance condition, a normal angle constraint and a curvature constraint are obtained, including: according to the formula Get the distance condition, normal angle constraint and curvature constraint, where ∈ d is the preset distance, ∈ θ is the preset normal angle, ∈ σ is the preset curvature, T is the rigid body transformation matrix, p i ″、q j ″ is each group of corresponding points in the candidate corresponding point set, is the unit normal vector of each set of corresponding points in the candidate corresponding point set, is the point curvature of each set of corresponding points in the candidate corresponding point set.

[0012] According to the present invention, an objective function is constructed, and an iteration termination condition is set according to the corresponding point set to complete the point cloud precise registration, including: obtaining the inverse of the source point cloud covariance matrix and the inverse of the target point cloud covariance matrix; obtaining the rotation matrix and the translation matrix through iterative least squares method; obtaining the objective function and the error of the corresponding points after rigid body transformation according to the corresponding point set, the source point cloud, the target point cloud, the inverse of the source point cloud covariance matrix, the inverse of the target point cloud covariance matrix, the rotation matrix, the translation matrix and the target rigid body transformation matrix; averaging the errors of all corresponding points after rigid body transformation to obtain the average error of the corresponding points after rigid body transformation; when the iteration termination condition is met, that is, the average error of the corresponding points after rigid body transformation of two adjacent iteration processes is less than the error threshold or reaches the maximum number of iterations, the iteration is terminated to obtain the final target rigid body transformation matrix to complete the point cloud precise registration.

[0013] According to the present invention, according to the corresponding point set, the source point cloud, the target point cloud, the inverse of the source point cloud covariance matrix, the inverse of the target point cloud covariance matrix, the rotation matrix, the translation matrix and the target rigid body transformation matrix, the objective function and the error after the rigid body transformation of the corresponding point are obtained, including: according to the formula J i (i)=(q j-Tp i ) T (Ω i,P +RΩ i,Q R T )(q j -Tp i ), and obtain the objective function T * and p i With q j The error J after the rigid body transformation of the corresponding point i (i), where P is the source point cloud, Q is the target point cloud, and Γ C* is the corresponding point set, p i ∈P,q j ∈Q is Γ C* The corresponding points in is the unit normal vector of each set of corresponding points in the corresponding point set, T is the rigid body transformation matrix, is the transformation operator, is the information matrix, Ω i,P is the inverse of the source point cloud covariance matrix, is the inverse of the target point cloud covariance matrix, R is the rotation matrix, and t is the translation matrix.

[0014] According to a second aspect of the present invention, there is provided an improved point cloud registration device for complex curved surface components, comprising: a processor; a memory for storing processor executable instructions; wherein the processor is configured to call the instructions stored in the memory to execute the improved point cloud registration method for complex curved surface components.

[0015] According to a fourth aspect of the present invention, there is provided a computer-readable storage medium having computer program instructions stored thereon, wherein the computer program instructions, when executed by a processor, implement the improved point cloud registration method for complex curved surface components.

[0016] Technical effect: According to the present invention, by statistics of mean curvature and Gaussian curvature, point cloud features are extracted, which is helpful for correct matching of feature points, thereby improving the accuracy of the transformation matrix obtained by rough registration of point clouds. In the precise registration of point clouds, corresponding point pairs are obtained through symmetrical structures and local information is used to screen out more reliable corresponding point pairs. At the same time, the objective function is improved to introduce normal vector differences, and the form of information matrix is ​​derived through symmetrical structures, thereby improving the defects of the original algorithm, and the precise registration of point clouds of different viewing angles of complex curved surface components can be achieved, and the speed and accuracy of point cloud registration can be improved. When determining the features of feature points, the feature information can be fused by weighted averaging the features of the feature points themselves and the feature information of the surrounding neighboring points, thereby determining the features of the feature points, enriching the feature information, enhancing the robustness of the features, and facilitating the subsequent correct matching of feature point pairs. When obtaining rigid distance constraints, the rigid distance constraints can be obtained by the Euclidean distance between the corresponding point in the feature corresponding point pair and another corresponding point after transformation by the rigid body transformation matrix, which helps to filter out erroneous matches and improve the accuracy of registration. When obtaining distance conditions, normal angle constraints and curvature constraints, the distance conditions, normal angle constraints and curvature constraints can be obtained through the local information of the point cloud, thereby screening matching points, alleviating the algorithm's dependence on the initial pose, not easily falling into the local optimum, and improving the accuracy and robustness of point cloud registration. When obtaining the error after the rigid body transformation of the objective function and the corresponding point, the error after the rigid body transformation of the objective function and the corresponding point can be obtained based on the rotation matrix and the translation matrix, and by introducing the normal vector difference in the extra dimension and deriving the value of the information matrix from the symmetric structure, thereby enhancing the robustness and accuracy of the algorithm, and even if the initial pose deviation is large, high-precision results can be obtained.

[0017] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and not limiting of the present invention. Other features and aspects of the present invention will become more apparent from the following detailed description of exemplary embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other embodiments can be obtained based on these drawings without creative work.

[0019] Figure 1 A schematic diagram of a process flow of an improved point cloud registration method for a complex curved surface component according to an embodiment of the present invention is exemplarily shown;

[0020] Figure 2The basic framework diagram of the improved point cloud registration method for complex curved surface components according to an embodiment of the present invention is exemplarily shown. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0023] Figure 1 A schematic flow chart of an improved point cloud registration method for a complex curved surface component according to an embodiment of the present invention is exemplarily shown, the method comprising: step S101, fixing the position of the complex curved surface component, using a binocular structured light measurement system to measure from two different perspectives, and collecting surface information of the complex curved surface component, wherein the surface information is stored in a point cloud model, one of the two different perspectives is used as a source point cloud, and the other perspective is used as a target point cloud; step S102, based on the local statistical characteristics of each point and its surrounding neighboring points, the source point cloud and the target point cloud are compared. Point cloud denoising; step S103, screening feature points, extracting feature point features of the source point cloud and the target point cloud, and determining feature corresponding point pairs; step S104, performing point cloud coarse alignment according to the feature corresponding point pairs; step S105, obtaining local information of each point in the source point cloud and the target point cloud, wherein the local information includes normal, curvature and covariance matrix; step S106, determining corresponding point sets according to the local information and bidirectional search; step S107, constructing an objective function, setting iteration termination conditions according to the corresponding point sets, and completing point cloud precise alignment.

[0024] According to the improved point cloud registration method for complex curved surface components according to the embodiments of the present invention, point cloud features are extracted by statistics of mean curvature and Gaussian curvature, which helps to correctly match feature points, thereby improving the accuracy of the transformation matrix obtained by coarse point cloud registration. In fine point cloud registration, corresponding point pairs are obtained through symmetrical structures and local information is used to screen out more reliable corresponding point pairs. At the same time, the objective function is improved to introduce normal vector differences, and the form of information matrix is ​​derived through symmetrical structures, thereby improving the defects of the original algorithm, and realizing accurate registration of point clouds of complex curved surface components from different perspectives, thereby improving the speed and accuracy of point cloud registration.

[0025] According to one embodiment of the present invention, in step S101, the binocular structured light measurement system is composed of two industrial cameras and a projector. The projector projects grating stripes and performs measurement based on the triangulation principle. The three-dimensional data coordinates measured at each viewing angle are based on the coordinates of the binocular structured light measurement system itself.

[0026] According to one embodiment of the present invention, in step S102, the source point cloud and the target point cloud are denoised based on the local statistical characteristics of each point and its surrounding neighboring points, which can effectively remove noise points without mistakenly deleting valid point clouds, and is suitable for point cloud data of complex curved surface components.

[0027] According to one embodiment of the present invention, step S102 includes: establishing a kd-tree for each point cloud, querying the k nearest neighbor points of each point in the respective point cloud, obtaining the least squares fitting plane of each point and its k nearest neighbor points, where k is a positive integer; obtaining the distance from each point to the least squares fitting plane; setting a distance threshold; when the distance from the point to the least squares fitting plane exceeds the distance threshold, determining that the point is a noise point and removing it.

[0028] According to one embodiment of the present invention, the least squares fitting plane can fit the local geometric features of each point and its k neighboring points. The distance from each point to its corresponding least squares fitting plane can reflect the degree of deviation of the point relative to the least squares fitting plane. The distance threshold is used as a criterion for judging whether a point is a noise point. For each point, if the distance from each point to its corresponding least squares fitting plane exceeds the distance threshold, the point is judged as a noise point and removed, thereby achieving denoising of the point cloud data and retaining the original point cloud data features.

[0029] According to an embodiment of the present invention, in step S103, feature points are screened, feature point features of the source point cloud and the target point cloud are extracted, and feature corresponding point pairs are determined.

[0030] According to one embodiment of the present invention, step S103 includes: screening feature points, extracting feature point features of the source point cloud and the target point cloud, and determining feature corresponding point pairs, including: obtaining the arithmetic mean of the angle between the normal vector of each point and the normal vectors of its k nearest neighboring points; obtaining the average value μ and variance σ of the angle between the normal vectors of all points and the normal vectors of their k nearest neighboring points; taking the points whose arithmetic mean is greater than μ-2σ and less than μ+2σ as feature points; obtaining an extended feature histogram of each feature point; determining the feature point features based on the extended feature histogram; establishing a kd-tree for the feature point features, and performing a nearest neighbor search on the source point cloud to obtain feature corresponding point pairs.

[0031] According to one embodiment of the present invention, the arithmetic mean of the angle between the normal vector of each point and the normal vectors of its k nearest neighboring points is used to select points with large normal vector changes as feature points, that is, points with arithmetic mean values ​​greater than μ-2σ and less than μ+2σ are used as feature points. For each point, the simplified point feature histogram is expanded, and the expanded feature histogram of each feature point is obtained through the statistics of Gaussian curvature and mean curvature. The curvature measurement provides a deeper insight into the local shape of each point, which helps to extract feature information. A kd-tree is established for the obtained feature point features, and the source point cloud searches for the nearest neighbors on the target point cloud to obtain feature corresponding point pairs.

[0032] According to an embodiment of the present invention, determining the feature point features according to the extended feature histogram includes: determining the feature FPFH′(p) of the feature point p according to formula (1),

[0033]

[0034] Among them, SPFH′(p) is the extended feature histogram of feature point p, p si is the i-th neighbor point of feature point p, SPFH′(p si ) is the extended feature histogram of the i-th neighboring point of feature point p, ω i is the Euclidean distance from the feature point to the i-th neighbor point of the feature point, k is the number of neighbor points, i≤k, and i and k are both positive integers.

[0035] According to one embodiment of the present invention, in formula (1), It represents the weighted average of the extended feature histograms of all neighboring points of the feature point p, which represents the local features around the feature point. The extended feature histogram of feature point p is added with the weighted average result of the extended feature histograms of all neighboring points as the feature of feature point p.

[0036] In this way, the feature information can be fused by weighted averaging the feature of the feature point itself and the feature information of its surrounding neighboring points, thereby determining the feature of the feature point, enriching the feature information, enhancing the robustness of the feature, and facilitating the subsequent correct matching of feature point pairs.

[0037] According to an embodiment of the present invention, in step S104, a point cloud rough registration is performed based on the feature corresponding point pairs using rigid distance constraints and random sampling consistency.

[0038] According to an embodiment of the present invention, a point cloud rough registration is performed based on the feature corresponding point pairs, including: setting a maximum number of iterations, randomly selecting 3 from the feature corresponding point pairs in each iteration, and obtaining a rigid body transformation matrix; obtaining the total average distance between the source point cloud and the target point cloud; obtaining a rigid distance constraint according to formula (2),

[0039]

[0040] Among them, q i ′、p j ' is the corresponding point in the feature corresponding point pair, T is the rigid body transformation matrix, is the total average distance between the source point cloud and the target point cloud; determines whether the remaining feature corresponding point pairs satisfy the rigid distance constraint, and if so, they are the correct matching point pairs, wherein the correct matching point pairs are the inliers; when the repeated iterations reach the maximum number of iterations, the rigid body transformation matrix with the largest number of inliers is taken as the correct coarse registration.

[0041] According to one embodiment of the present invention, in formula (2), ||q i ′-Tp j ′|| represents the corresponding point q in the feature corresponding point pair i ′ and the corresponding point p after the rigid body transformation matrix transformation j The smaller the Euclidean distance, the better the matching of the corresponding points in the feature corresponding point pair. When the Euclidean distance is less than 3 times the total average distance between the source point cloud and the target point cloud, the point pair is correctly matched.

[0042] In this way, the rigid distance constraint can be obtained through the Euclidean distance between the corresponding point in the feature corresponding point pair and another corresponding point transformed by the rigid body transformation matrix, which helps to filter out erroneous matches and improve the accuracy of registration.

[0043] According to one embodiment of the present invention, in step S105, local information of each point in the source point cloud and the target point cloud is obtained, k nearest neighbor points may be searched through kd-tree, and normals, curvatures and covariance matrices may be calculated using the k nearest neighbor points.

[0044] According to one embodiment of the present invention, in step S106, the corresponding point set is determined based on the local information and two-way search. Compared with the traditional precise registration method in which one-way search is prone to one-to-many mismatches, not only is one-to-one matching achieved, but the local information of the point cloud is further used to filter the matching points.

[0045] According to an embodiment of the present invention, step S106 includes: using kd-tree to distribute in the source point cloud and the target point cloud to search for the nearest neighbor of each point, respectively, to obtain a corresponding point set of the source point cloud and a corresponding point set of the target point cloud; extracting p from each group of corresponding points in the corresponding point set of the source point cloud i ″The corresponding q j ″, search q again in the corresponding point set of the target point cloud j The corresponding p v ″, the corresponding points satisfying i=v are grouped into a candidate corresponding point set; the unit normal vector and the point curvature of each group of corresponding points in the candidate corresponding point set are obtained; according to each group of corresponding points in the candidate corresponding point set, the unit normal vector and the point curvature, the distance condition, the normal angle constraint and the curvature constraint are obtained; when the distance condition, the normal angle constraint and the curvature constraint are all satisfied, the corresponding point set is determined.

[0046] According to one embodiment of the present invention, obtaining a distance condition, a normal angle constraint and a curvature constraint according to each group of corresponding points in the candidate corresponding point set, the unit normal vector and the point curvature includes: obtaining the distance condition, the normal angle constraint and the curvature constraint according to formulas (3), (4) and (5),

[0047] ||q j ″-Tp i ″||>∈ d (3)

[0048]

[0049] Among them, ∈ d is the preset distance, ∈ θ is the preset normal angle, ∈ σ is the preset curvature, T is the rigid body transformation matrix, p i ″、q j ″ is each group of corresponding points in the candidate corresponding point set, is the unit normal vector of each set of corresponding points in the candidate corresponding point set, is the point curvature of each set of corresponding points in the candidate corresponding point set.

[0050] According to one embodiment of the present invention, in formula (3), ||q j ″-Tp i ″||>∈ d Represents the corresponding point q in the candidate corresponding point set j ″ and the corresponding point p after the rigid body transformation matrix transformation i The Euclidean distance between ″ is greater than the preset distance. In formula (4), Represents the corresponding point q in the candidate corresponding point set j ″ and the corresponding point p iThe normal angle between the unit normal vectors is less than the preset normal angle. In formula (5), Represents the corresponding point q in the candidate corresponding point set j ″ and the corresponding point p i The curvature of the point is less than the preset curvature.

[0051] In this way, the distance conditions, normal angle constraints and curvature constraints can be obtained through the local information of the point cloud, so as to screen the matching points, alleviate the algorithm's dependence on the initial pose, avoid falling into the local optimum, and improve the accuracy and robustness of the point cloud registration.

[0052] According to one embodiment of the present invention, in step S107, an objective function is constructed, a rigid body transformation that minimizes the objective function is calculated, and after the corresponding transformation is performed, the corresponding points are searched again, the objective function is constructed, and the rigid body transformation is calculated. The maximum number of iterations is set, and the iteration is terminated when the iteration termination condition or the maximum number of iterations is reached, and the rigid body transformation matrix calculated by the last iteration is used as the precise registration result of the point cloud.

[0053] According to one embodiment of the present invention, step S107 includes: obtaining the inverse of the source point cloud covariance matrix and the inverse of the target point cloud covariance matrix; obtaining the rotation matrix and the translation matrix through iterative least squares method; obtaining the objective function and the error after the rigid body transformation of the corresponding points according to the corresponding point set, the source point cloud, the target point cloud, the inverse of the source point cloud covariance matrix, the inverse of the target point cloud covariance matrix, the rotation matrix, the translation matrix and the target rigid body transformation matrix; averaging the errors after the rigid body transformation of all corresponding points to obtain the average error after the rigid body transformation of the corresponding points; when the iteration termination condition is met, that is, the average error after the rigid body transformation of the corresponding points in two adjacent iteration processes is less than the error threshold or reaches the maximum number of iterations, the iteration is terminated to obtain the final target rigid body transformation matrix, and the point cloud precise alignment is completed.

[0054] According to one embodiment of the present invention, according to the corresponding point set, the source point cloud, the target point cloud, the inverse of the source point cloud covariance matrix, the inverse of the target point cloud covariance matrix, the rotation matrix, the translation matrix and the target rigid body transformation matrix, the objective function and the error after the rigid body transformation of the corresponding point are obtained, including: obtaining the objective function T according to formulas (6), (7), (8) and (9) * and p i With q j The error J after the rigid body transformation of the corresponding point i (i)

[0055]

[0056] J i (i)=(q j -Tpi ) T (Ω i,P +RΩ i,Q R T )(q j -Tp i ) (9)

[0057] Get the objective function T * and p i With q j The error J after the rigid body transformation of the corresponding point i (i), where P is the source point cloud, Q is the target point cloud, and Γ C* is the corresponding point set, p i ∈P,q j ∈Q is Γ C* The corresponding points in is the unit normal vector of each set of corresponding points in the corresponding point set, T is the rigid body transformation matrix, is the transformation operator, is the information matrix, Ω i,P is the inverse of the source point cloud covariance matrix, is the inverse of the target point cloud covariance matrix, R is the rotation matrix, and t is the translation matrix.

[0058] According to one embodiment of the present invention, in formula (6), the objective function is to find an optimal rigid body transformation matrix so that the error between the transformed source point cloud and the target point cloud is minimized. According to one embodiment of the present invention, in formula (7), Indicates that p i Transformed by rigid body transformation matrix (including rotation matrix and translation vector). In formula (8), is the information matrix derived from the bidirectional symmetric structure. In formula (9), (q j -Tp i ) T (Ω i,P +RΩ i,Q R T )(q j -Tp i ) represents the error between the points in the source point cloud and the corresponding points in the target point cloud after rigid body transformation. This error is used to evaluate the quality of the current rigid body transformation matrix and is iteratively updated during the optimization process.

[0059] In this way, based on the rotation matrix and translation matrix, and by introducing normal vector differences in extra dimensions and deriving the value of the information matrix from the symmetric structure, the error after the rigid body transformation of the objective function and the corresponding points can be obtained, thereby enhancing the robustness and accuracy of the algorithm and obtaining high-precision results even if the initial pose deviation is large.

[0060] According to the improved point cloud registration method for complex curved surface components of the embodiment of the present invention, point cloud features are extracted by statistics of average curvature and Gaussian curvature, which is helpful for correct matching of feature points, thereby improving the accuracy of the transformation matrix obtained by rough point cloud registration. In point cloud fine registration, corresponding point pairs are obtained by symmetrical structure and more reliable corresponding point pairs are screened out by using local information. At the same time, the objective function is improved to introduce normal vector differences, and the form of information matrix is ​​derived by symmetrical structure, thereby improving the defects of the original algorithm, and the precision registration of point clouds of different viewing angles of complex curved surface components can be achieved, and the speed and accuracy of point cloud registration can be improved. When determining the features of feature points, the feature information can be fused by weighted average of the features of the feature points themselves and the feature information of the surrounding neighboring points, thereby determining the features of the feature points, enriching the feature information, enhancing the robustness of the features, and facilitating the subsequent correct matching of feature point pairs. When obtaining rigid distance constraints, the rigid distance constraints can be obtained by the Euclidean distance between the corresponding point in the feature corresponding point pair and another corresponding point after transformation by the rigid body transformation matrix, which helps to filter out wrong matches and improve the accuracy of registration. When obtaining distance conditions, normal angle constraints and curvature constraints, the distance conditions, normal angle constraints and curvature constraints can be obtained through the local information of the point cloud, thereby screening matching points, alleviating the algorithm's dependence on the initial pose, not easily falling into the local optimum, and improving the accuracy and robustness of point cloud registration. When obtaining the error after the rigid body transformation of the objective function and the corresponding point, the error after the rigid body transformation of the objective function and the corresponding point can be obtained based on the rotation matrix and the translation matrix, and by introducing the normal vector difference in the extra dimension and deriving the value of the information matrix from the symmetric structure, thereby enhancing the robustness and accuracy of the algorithm, and even if the initial pose deviation is large, high-precision results can be obtained.

[0061] Figure 2 The basic framework diagram of the improved point cloud registration method for complex curved surface components according to an embodiment of the present invention is exemplarily shown. The local surface data information of the complex curved surface workpiece is collected by a binocular structured light measurement system at different viewing angles. The source point cloud and the target point cloud are denoised based on the local statistical characteristics of each point in the source point cloud and the target point cloud and their surrounding neighboring points. Feature points are screened to extract the feature point features of the source point cloud and the target point cloud. Feature corresponding point pairs are determined based on kd-tree. Rough point cloud registration is performed using rigid distance constraints and random sampling consistency. The local information (normal, curvature, covariance) of each point in the source point cloud and the target point cloud is obtained. The corresponding point set is determined using the local information of the point and bidirectional search. The objective function is constructed and the point cloud is precisely registered using the iterative least squares method.

[0062] According to one embodiment of the present invention, there is provided an improved point cloud registration device for complex curved surface components, comprising: a processor; a memory for storing processor executable instructions; wherein the processor is configured to call the instructions stored in the memory to execute the improved point cloud registration method for complex curved surface components.

[0063] According to one embodiment of the present invention, a computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the improved point cloud registration method for complex curved surface components is implemented.

[0064] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0065] It should be understood by those skilled in the art that the embodiments of the present invention described above and shown in the accompanying drawings are only examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functional and structural principles of the present invention have been demonstrated and explained in the embodiments, and the embodiments of the present invention may be deformed or modified in any way without departing from the principles.

Claims

1. An improved point cloud registration method for complex curved surface components, characterized in that: include: The position of the complex curved surface component is fixed, and a binocular structured light measurement system is used to measure from two different perspectives to collect surface information of the complex curved surface component, wherein the surface information is stored in the form of a point cloud model, and one of the two different perspectives is used as the source point cloud, and the other perspective is used as the target point cloud; the source point cloud and the target point cloud are denoised according to the local statistical characteristics of each point and its surrounding neighboring points; feature points are screened, feature point features of the source point cloud and the target point cloud are extracted, and feature corresponding point pairs are determined; point cloud coarse alignment is performed according to the feature corresponding point pairs; local information of each point in the source point cloud and the target point cloud is obtained, wherein the local information includes normal, curvature and covariance matrix; corresponding point sets are determined according to the local information and bidirectional search; an objective function is constructed, and an iteration termination condition is set according to the corresponding point set to complete point cloud fine alignment.

2. The improved point cloud registration method for complex curved surface components according to claim 1, characterized in that: According to the local statistical characteristics of each point and its surrounding neighboring points, the source point cloud and the target point cloud are denoised, including: establishing a kd-tree for each point cloud, querying the k neighboring points of each point in the respective point cloud, obtaining the least squares fitting plane of each point and its k neighboring points, where k is a positive integer; obtaining the distance from each point to the least squares fitting plane; setting a distance threshold; when the distance from a point to the least squares fitting plane exceeds the distance threshold, determining that the point is a noise point and removing it.

3. The improved point cloud registration method for complex curved surface components according to claim 1, characterized in that: Screening feature points, extracting feature point features of the source point cloud and the target point cloud, and determining feature corresponding point pairs, including: obtaining the arithmetic mean of the angles between the normal vector of each point and the normal vectors of its k nearest neighboring points; obtaining the average value μ and variance σ of the angles between the normal vectors of all points and the normal vectors of their k nearest neighboring points; taking points whose arithmetic mean values ​​are greater than μ-2σ and less than μ+2σ as feature points; obtaining an extended feature histogram of each feature point; determining feature point features based on the extended feature histogram; establishing a kd-tree for the feature point features, performing nearest neighbor search of the source point cloud on the target point cloud, and obtaining feature corresponding point pairs.

4. The improved point cloud registration method for complex curved surface components according to claim 3, characterized in that: According to the extended feature histogram, determining the feature point features includes: according to the formula Determine the feature FPFH of feature point p ′ (p), where SPFH ′ (p) is the extended feature histogram of feature point p, p si is the i-th neighbor point of feature point p, SPFH ′ (p si ) is the extended feature histogram of the i-th neighboring point of feature point p, ω i is the Euclidean distance from the feature point to the i-th neighbor point of the feature point, k is the number of neighbor points, i≤k, and i and k are both positive integers.

5. The improved point cloud registration method for complex curved surface components according to claim 1, characterized in that: According to the feature corresponding point pairs, a point cloud rough registration is performed, including: setting a maximum number of iterations, randomly selecting 3 from the feature corresponding point pairs in each iteration, and obtaining a rigid body transformation matrix; obtaining the total average distance between the source point cloud and the target point cloud; according to the formula Get a rigid distance constraint, where q i ′ 、p j ′ is the corresponding point in the feature corresponding point pair, T is the rigid body transformation matrix, is the total average distance between the source point cloud and the target point cloud; determines whether the remaining feature corresponding point pairs satisfy the rigid distance constraint, and if so, they are the correct matching point pairs, wherein the correct matching point pairs are the inliers; when the repeated iterations reach the maximum number of iterations, the rigid body transformation matrix with the largest number of inliers is taken as the correct coarse registration.

6. The improved point cloud registration method for complex curved surface components according to claim 1, characterized in that: According to the local information and the bidirectional search, the corresponding point set is determined, including: using kd-tree to distribute in the source point cloud and the target point cloud to search for the nearest neighbor of each point, respectively, to obtain the corresponding point set of the source point cloud and the corresponding point set of the target point cloud; in the corresponding point set of the source point cloud, for each group of corresponding points, p is taken out. i ″The corresponding q j ″, search q again in the corresponding point set of the target point cloud j The corresponding p v ″, the corresponding points satisfying i=v are grouped into a candidate corresponding point set; the unit normal vector and the point curvature of each group of corresponding points in the candidate corresponding point set are obtained; according to each group of corresponding points in the candidate corresponding point set, the unit normal vector and the point curvature, the distance condition, the normal angle constraint and the curvature constraint are obtained; when the distance condition, the normal angle constraint and the curvature constraint are all satisfied, the corresponding point set is determined.

7. The improved point cloud registration method for complex curved surface components according to claim 6, characterized in that: According to each group of corresponding points in the candidate corresponding point set, the unit normal vector and the point curvature, a distance condition, a normal angle constraint and a curvature constraint are obtained, including: according to the formula Get the distance condition, normal angle constraint and curvature constraint, where ∈ d is the preset distance, ∈ θ is the preset normal angle, ∈ σ is the preset curvature, T is the rigid body transformation matrix, p i ″、q j ″ is each group of corresponding points in the candidate corresponding point set, is the unit normal vector of each set of corresponding points in the candidate corresponding point set, is the point curvature of each set of corresponding points in the candidate corresponding point set.

8. The improved point cloud registration method for complex curved surface components according to claim 1, characterized in that: Construct an objective function, set an iteration termination condition according to the corresponding point set, and complete the point cloud precise registration, including: obtaining the inverse of the source point cloud covariance matrix and the inverse of the target point cloud covariance matrix; obtaining the rotation matrix and the translation matrix through iterative least squares method; obtaining the objective function and the error of the corresponding points after rigid body transformation according to the corresponding point set, the source point cloud, the target point cloud, the inverse of the source point cloud covariance matrix, the inverse of the target point cloud covariance matrix, the rotation matrix, the translation matrix and the target rigid body transformation matrix; averaging the errors of all corresponding points after rigid body transformation to obtain the average error of the corresponding points after rigid body transformation; when the iteration termination condition is met, that is, the average error of the corresponding points after rigid body transformation of two adjacent iteration processes is less than the error threshold or reaches the maximum number of iterations, the iteration is terminated to obtain the final target rigid body transformation matrix and complete the point cloud precise registration.

9. The improved point cloud registration method for complex curved surface components according to claim 8, characterized in that: According to the corresponding point set, the source point cloud, the target point cloud, the inverse of the source point cloud covariance matrix, the inverse of the target point cloud covariance matrix, the rotation matrix, the translation matrix and the target rigid body transformation matrix, the objective function and the error after the rigid body transformation of the corresponding point are obtained, including: according to the formula Get the objective function T * and p i With q j The error J after the rigid body transformation of the corresponding point i (i), where P is the source point cloud and Q is the target point cloud. is the corresponding point set, p i ∈P,q j ∈Q is The corresponding points in is the unit normal vector of each set of corresponding points in the corresponding point set, T is the rigid body transformation matrix, is the transformation operator, is the information matrix, Ω i,P is the inverse of the source point cloud covariance matrix, is the inverse of the target point cloud covariance matrix, R is the rotation matrix, and t is the translation matrix.

10. A computer-readable storage medium, characterized in that: Computer program instructions are stored thereon, and when the computer program instructions are executed by a processor, the method according to any one of claims 1 to 9 is implemented.

Images