A point cloud registration method based on curved surface feature region constraint

Abstract

The application discloses a point cloud registration method based on a curved surface characteristic region constraint, comprising a point cloud downsampling module, a characteristic point extraction module, a characteristic point description and characteristic matching module, and a registration module based on a curved surface characteristic constraint region, and is used in a registration task with noise and unordered point cloud data. The point cloud downsampling module is subjected to denoising treatment by adopting a denoising algorithm based on KD-tree. The characteristic point extraction module adopts an extraction detection method based on a curved surface change index. The characteristic description and characteristic matching module uses a fast point feature histogram and a random sampling consistency matching algorithm to complete initial pose transformation. The registration module based on the curved surface characteristic constraint region adopts a method of adding a curved surface characteristic region constraint, accelerates the matching rate of icp characteristic point pairs by a KD-tree algorithm, completes point cloud registration, and improves the accuracy of point cloud registration. The application can be applied to model registration tasks in a complex part digitization detection process.

Description

Technical Field

[0001] This invention belongs to the field of product digital inspection, specifically involving a point cloud registration method based on surface feature region constraints. Background Technology

[0002] With the rapid development of digital design and manufacturing technologies, aerospace companies are placing higher demands on the quality of parts manufacturing. Since the precision of parts manufacturing depends on the accuracy of error detection, digital error detection technology is also facing greater requirements. Currently, traditional parts inspection methods include template inspection and coordinate measuring machine (CMM) inspection. Template inspection, being manual, is easily affected by subjective factors, and its accuracy and speed are no longer sufficient to meet the requirements of modern industrial manufacturing. With the widespread adoption of CMMs in industrial manufacturing, 3D digital inspection technology has attracted considerable attention. While CMMs can perform high-precision measurements on complex parts, they can only measure a single key inspection point and cannot obtain complete information about the surface of the part, resulting in low efficiency and high cost. Therefore, an efficient parts inspection solution is needed to ensure the accurate shape of parts.

[0003] Currently, CAD model-based digital inspection technology is receiving increasing attention and is gradually becoming the main direction of future digital inspection. CAD model-based digital inspection technology mainly refers to registering the CAD design model with the scanned point cloud data of the workpiece to unify the coordinates, then calculating the deviation between the two models to compare the product's processing accuracy for manufacturing and assembly. The steps generally include data acquisition, registration, and deviation calculation and analysis. Among these, the point cloud registration technology between the scanned dataset and the design model's point cloud dataset is the foundation and core of everything. Summary of the Invention

[0004] To address the challenges of quality inspection for critical structural components in weaponry due to their complex structures, a point cloud registration method based on surface feature region constraints is proposed. This method mainly includes a point cloud downsampling module, a feature point extraction module, a feature point description and feature matching module, and a registration module based on surface feature constraint regions. It solves the problems of slow feature point extraction speed, slow point cloud registration speed, and low accuracy in disordered point cloud sets, providing technical support for digital precision measurement.

[0005] Furthermore, the point cloud downsampling module establishes topological set relationships by constructing k-neighborhoods of the point cloud, calculates normal vectors and curvature to estimate feature information of the part surface, and uses a KD-tree-based denoising algorithm to smooth and denoise the point cloud, thereby reducing the dimensionality of the point cloud.

[0006] Furthermore, the feature point extraction module includes the following steps:

[0007] (1) Choose any point p(x, y, z) T Choose point p inside a sphere with center r and radius r. i (x i y i , z i ) T As its neighborhood point set.

[0008] (2) Calculate the covariance matrix C at point p and perform principal component analysis.

[0009] CV = DV

[0010] Where V is the matrix of eigenvectors and D is the diagonal matrix of eigenvalues ​​of C.

[0011] (3) Use the Hotling transform to transform all neighborhood points into a new coordinate system formed by the feature vectors in V, forming a new set of neighborhood points.

[0012]

[0013] (4) Let X be the component of the new neighborhood point on the x-axis, and let Y be the component of the new neighborhood point on the y-axis. Calculate the surface change index based on X and Y.

[0014] X = {x1, x2, ..., x} n}

[0015] Y = {y1, y2, ..., y} n}

[0016]

[0017] Index of Surface Variations In reality, it is the ratio of the lengths of the x-axis and y-axis of the projection of the neighboring points onto the new coordinate system, which reflects the surface changes of a local region with point p as the center and r as the radius.

[0018] Furthermore, the feature point description and feature point pair matching module mainly includes the following steps: After extracting feature points using a feature point extraction algorithm based on surface variation index, it is necessary to calculate the surface normal and the deviation angle between the normals in the neighboring regions of the feature points, and to parameterize the differences between neighboring points. Simultaneously, a fast point feature histogram is used to complete a complete description of the feature points. Finally, the feature matching task is completed using a random sampling consensus algorithm. The specific steps are as follows:

[0019] (1) Select sampling points from the source point cloud. Set a distance threshold d, and select points in the source point cloud whose distance is greater than d as sampling points to ensure that the fast point feature histograms of the selected points are different.

[0020] (2) Find the corresponding points of the sampling points in (1). Search for points in the target point cloud that have similar features to the sampling points in (1) and take them as the corresponding points of the sampling points respectively.

[0021] (3) Calculate the transformation matrix between the sampled point set and its corresponding point set. Using the transformation matrix, calculate the distance error between the transformed corresponding point sets, express the result using the Huber penalty function, and determine the current registration performance.

[0022]

[0023]

[0024] In the formula: m1 is a given value, l i This represents the distance difference of the i-th group after transformation.

[0025] In summary, this process involves finding the optimal solution for the transformation matrix. When the distance error function value is minimized, the final transformation matrix can be obtained.

[0026] Furthermore, the registration module based on surface feature constraint regions mainly establishes constraint regions, assigns different weights to different regions, prioritizes matching feature point pairs within the constraint regions as alignment references, and finally uses the KD-tree algorithm to first establish an index structure for each feature point and quickly search for the nearest point pair, thereby improving the registration efficiency of point clouds.

[0027] Experimental comparisons have shown that the point cloud registration method of this invention achieves significantly higher accuracy than other point cloud registration methods when processing point clouds of complex curved surface features. Attached Figure Description

[0028] Figure 1 This is an image of the source point cloud obtained by scanning with a scanner according to an embodiment of the present invention;

[0029] Figure 2 This is a target point cloud image converted from digital model CAD format according to an embodiment of the present invention;

[0030] Figure 3 , Figure 4 They are Figure 1 , Figure 2 The image after downsampling;

[0031] Figure 5 yes Figure 3 , Figure 4 The image after coarse registration in the middle image;

[0032] Figure 6 yes Figure 3 , Figure 4 The image after fine registration.

[0033] Figure 7 This is a schematic diagram illustrating the implementation principle of the present invention. Detailed Implementation

[0034] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. However, the embodiments are not intended to limit the present invention, but are only intended to better illustrate the present invention and demonstrate its effects.

[0035] This invention mainly addresses the registration of point cloud models of complex curved surface parts. During point cloud data acquisition, noise is inevitably present, which greatly affects the registration efficiency and accuracy of point clouds. To address this problem, this invention proposes a point cloud registration method based on surface feature region constraints, which mainly includes four steps: point cloud downsampling, feature point extraction, feature point description and feature matching, and registration based on surface feature constraint regions.

[0036] This embodiment illustrates the effects of the present invention by combining the point cloud model registration process of the radome. For example... Figure 1 The image shows the source point cloud model acquired by the time-scanning radome. Figure 2 The image shows a point cloud model converted from a radar dome digital model. Since the two point clouds were acquired using different methods, their coordinate systems are different, necessitating point cloud registration. This model is used as input to this invention, and the output is a rigid body transformation matrix. The steps are described in detail below.

[0037] The point cloud downsampling module establishes topological set relationships by constructing k-neighborhoods of the point cloud, calculates normal vectors and curvature to estimate the feature information of the part surface, and uses a KD-tree-based denoising algorithm to smooth and denoise the point cloud, reducing its dimensionality. The downsampling effect is as follows: Figure 3 , Figure 4 As shown.

[0038] The feature point extraction module extracts feature points using a surface transformation index-based extraction algorithm, thus accelerating registration efficiency. The specific steps are as follows:

[0039] (1) Choose any point p(x, y, z) T Choose point p inside a sphere with center r and radius r. i (x i y i , z i ) T As its neighborhood point set.

[0040] (2) Calculate the covariance matrix C at point p and perform principal component analysis.

[0041] CV = DV

[0042] Where V is the matrix of eigenvectors and D is the diagonal matrix of eigenvalues ​​of C.

[0043] (3) Use the Hotling transform to transform all neighborhood points into a new coordinate system formed by the feature vectors in V, forming a new set of neighborhood points.

[0044]

[0045] (4) Let X be the component of the new neighborhood point on the x-axis, and let Y be the component of the new neighborhood point on the y-axis. Calculate the surface change index based on X and Y.

[0046] X = {x1, x2, ..., x} n}

[0047] Y = {y1, y2, ..., y} n}

[0048]

[0049] Index of Surface Variations In reality, it is the ratio of the lengths of the x-axis and y-axis of the projection of the neighboring points onto the new coordinate system, which reflects the surface changes of a local region with point p as the center and r as the radius.

[0050] Feature point description and feature matching involve extracting feature points using a surface variation index-based feature point extraction algorithm, calculating the surface normals and the angles between them in the neighboring regions of each feature point, parameterizing the differences between neighboring points, and combining this with a fast point feature histogram to complete a comprehensive description of the feature points. Finally, a random sampling consensus algorithm is used to complete the feature matching task. The specific steps of feature matching are as follows:

[0051] (1) Select sampling points from the source point cloud. Set a distance threshold d, and select points in the source point cloud whose distance is greater than d as sampling points to ensure that the fast point feature histograms of the selected points are different.

[0052] (2) Find the corresponding points of the sampling points in (1). Search for points in the target point cloud that have similar features to the sampling points in (1) and take them as the corresponding points of the sampling points respectively.

[0053] (3) Calculate the transformation matrix between the sampled point set and its corresponding point set. Using the transformation matrix, calculate the distance error between the transformed corresponding point sets, express the result using the Huber penalty function, and determine the current registration performance.

[0054]

[0055]

[0056] In the formula: m1 is a given value, l i This represents the distance difference of the i-th group after transformation.

[0057] In summary, this process involves finding the optimal solution for the transformation matrix. The final transformation matrix can be obtained when the distance error function is minimized. In this example, the matching effect is as follows: Figure 5 As shown.

[0058] Registration based on surface feature constraints mainly involves establishing constraint regions, assigning different weights to different regions, prioritizing matching feature point pairs within the constraint regions as alignment references, and finally using the KD-tree algorithm to first establish an index structure for each feature point and quickly search for the nearest point pair, thereby improving the registration efficiency of the point cloud. The steps of the ICP algorithm based on KD-tree for quickly searching for the nearest point pair are as follows:

[0059] (1) Compare the query point with the value of the dimension determined after splitting. If it is less than or equal to the value of the split dimension, proceed to the left subtree branch; if it is greater than the value of the split dimension, proceed to the right subtree branch. This process is repeated until several leaf nodes of the binary search tree are reached, and the nearest neighbor similar point in the same subspace as the query point is found along the search path.

[0060] (2) Search the subspace of the leaf node on other search paths. If there is a closer point on the leaf node, jump to the leaf node in the subspace to find the nearest point on the line.

[0061] (3) Repeat steps (1) and (2) above until the search path is empty, then end the search.

[0062] (4) Determine the corresponding point pairs based on the KD-tree, search for the nearest point in the scanned point cloud A in the target point cloud B, and form corresponding point pairs;

[0063] (5) Calculate the centroids of the two new point sets:

[0064]

[0065]

[0066] (6) Calculate the objective function f(R) k T k The principle is to minimize the objective function value.

[0067]

[0068] In the formula R k For the optimal rotation matrix, T k It is a translation matrix.

[0069] (7) Transform the registered point cloud B coordinates to B a+1 :

[0070]

[0071] (8) Calculate the average distance between point pairs

[0072]

[0073]

[0074] In the formula This represents the average distance after k iterations and k+1 iterations.

[0075] (9) τ is the preset average distance error, when If the iteration ends, the iteration continues; otherwise, the iteration restarts.

[0076] In this example, the registration based on surface feature constraints is completed according to the above steps. The final registration and matching effect is as follows: Figure 6 As shown. Comparative verification shows that the root mean square error (RMSE) of the registration result calculated using the present invention is 0.125 mm, while the RMS error calculated using the classic ICP algorithm is 2.315 mm. The registration accuracy of the present invention is significantly lower.

[0077] The above-described embodiments are merely preferred embodiments provided to fully illustrate the present invention, and the scope of protection of the present invention is not limited thereto. Equivalent substitutions or modifications made by those skilled in the art based on the present invention are all within the scope of protection of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A point cloud registration method based on surface feature region constraints, characterized in that, It includes a point cloud downsampling module, a feature point extraction module, a feature point description and feature matching module, and a registration module based on surface feature constraint regions; The point cloud downsampling module establishes topological set relationships by constructing k-neighborhoods of the point cloud, calculates normal vectors and curvature to estimate feature information of the part surface, and uses a KD-tree-based denoising algorithm to smooth and denoise the point cloud, thereby reducing the dimensionality of the point cloud. The feature point extraction module includes the following steps: Step (1) Select any point Choose it as the center. Points inside a sphere with radius As its neighborhood point set; Step (2) Calculate the point covariance matrix at Perform principal component analysis. ； in, A matrix of eigenvectors, for The diagonal matrix of eigenvalues; Step (3) uses the Hotelling transform to convert all neighborhood points to In the new coordinate system formed by the eigenvectors, a new set of neighborhood points is formed; ； Step (4) Let For new neighborhood points in Components on the axis, let For new neighborhood points in Components on the axis, according to , Calculate the index of surface variation ; ； ； ； Index of Surface Variations In fact, it is the projection of the neighboring points into the new coordinate system. axis, The ratio of the lengths of the axes reflects the relationship between points. For the center of the ball, The surface variation in a local region with a radius; The feature point description and feature point pair matching module includes the following steps: after extracting feature points using a feature point extraction algorithm based on surface change index, the surface normal and the deviation angle between the normals in the neighboring region of the feature point are calculated, and the differences between neighboring points are parameterized. At the same time, a complete description of the feature points is completed by combining a fast point feature histogram. Finally, the feature matching task is completed by a random sampling consistency algorithm. The registration module based on surface feature constraint regions establishes constraint regions, assigns different weights to different regions, prioritizes matching feature point pairs within the constraint regions as alignment references, and finally uses the KD-tree algorithm to first establish an index structure for each feature point and quickly search for the nearest point pair, thereby improving the registration efficiency of point clouds.

2. The point cloud registration method based on surface feature region constraints as described in claim 1, characterized in that, The implementation steps of the feature point description and feature point pair matching module are as follows: Step (1) Select sampling points from the source point cloud; set a distance threshold. Select point clouds with a distance greater than 1000 from the source point cloud. The points are used as sampling points to ensure that the fast point feature histograms of the points are different; Step (2) Find the corresponding points of the sampling points in step (1); search for points in the target point cloud that have similar features to the sampling points in step (1), and use them as the corresponding points of the sampling points respectively; Step (3) Calculate the transformation matrix between the sampling point set and its corresponding point set; use the transformation matrix to find the distance error between the transformed corresponding point sets, use the Huber penalty function to represent the result, and judge the current registration performance; ； ； In the formula: It is a given value. Indicates the transformed number of... The distance difference between groups; When the distance error function value is minimized, the final transformation matrix is ​​obtained.

Images