Registration method for point cloud data of non-rotary body hood type structure

Abstract

The invention discloses a registration method for point cloud data of a non-rotary body hood structure, and the method comprises the steps: converting original point cloud data into a combination form of skeleton data and edge feature data, and carrying out the initial registration operation through the combination data of the skeleton data and the edge feature data, thereby obtaining an initial transformation matrix; then performing fine registration operation on the basis of initial registration to obtain a fine registration transformation matrix, and realizing registration of the original point cloud and the target point cloud; according to the method, rapid registration can be carried out on mass point cloud data of a non-rotary body head cover type structure.

Description

Technical Field

[0001] The invention belongs to the technical field of rapid registration of non-rotating point cloud data, and in particular relates to a registration method for point cloud data of a non-rotating head cover structure. Background Art

[0002] With the rapid development of advanced manufacturing technology, three-dimensional shape measurement technology has been widely used in aerospace, weapons and equipment, automobile and shipbuilding and other fields. As a key component of spacecraft, the radome can protect the normal operation of the guidance system of the aircraft under extreme conditions. The radome adopts a large volume, multi-surface, deep cavity and special-shaped structure. Therefore, when measuring the shape of such workpieces, in order to ensure the accuracy of the workpiece details, laser scanning or optical measurement technology is usually used to obtain massive point cloud data of the target workpiece. In actual engineering applications, it is necessary to align the measured data with the design model through the registration algorithm to analyze the geometric parameters of the measured workpiece. Point cloud registration is divided into two steps: coarse registration and fine registration. In recent years, many scholars have carried out a lot of research on point cloud algorithms. Common methods include random sampling consensus (RANSAC) method, statistical solution method, normal distribution transformation (NDT), iterative closest point (ICP) algorithm, etc.

[0003] However, since the surface of the radome workpiece is continuous and has no obvious geometric features, it is impossible to perform registration operations based on the feature points of the workpiece's point cloud data. The only way to calculate the transformation matrix is ​​to use all the point cloud data of the measured workpiece as the input of the algorithm. However, the amount of actual measured point cloud data is large, and the matching between the measured model and the design model data points in the existing registration algorithm results is poor. The operation time cost is high and cannot meet the actual engineering needs. Therefore, it is necessary to study a fast registration method for point cloud data of non-rotating head cover structures to solve the problem of slow registration speed and low accuracy of massive point cloud data under this application requirement. Summary of the invention

[0004] In view of this, the present invention provides a registration method for point cloud data of non-rotating head cover structures, which can quickly perform registration on massive point cloud data of non-rotating head cover structures.

[0005] The technical solution for implementing the present invention is as follows:

[0006] A registration method for point cloud data of non-rotating head cover structures, which converts the original point cloud data into a combination of skeleton data and edge feature data, uses the combined data of skeleton data and edge data to perform initial registration operation to obtain an initial transformation matrix, and then performs fine registration operation on the basis of the initial registration to obtain a fine registration transformation matrix to achieve registration of the original point cloud and the target point cloud.

[0007] Furthermore, the original point cloud data and the target point cloud data are preprocessed, the skeleton points and edge feature points of the two point cloud data are calculated respectively, and the skeleton points and edge feature points are used to replace the point cloud data for registration operation.

[0008] Furthermore, the preprocessing includes: performing voxel filtering on the original point cloud and extracting the central skeleton of the point cloud.

[0009] Furthermore, the preprocessing includes: extracting edge feature points, specifically: first fitting a least squares plane based on the k nearest neighbors and the target point, projecting the target point and its neighborhood points onto the plane, and finally constructing an edge point detection operator using the idea of ​​weighted equivalent force and the density parameters of the points, and extracting edge feature points by setting a threshold.

[0010] Furthermore, based on the PCA initial registration algorithm, the initial registration transformation matrix R0 and T0 of the original point cloud model are obtained, and the original point cloud data P source0 After the initial transformation, we get P' source0 ; Finally, P' source0 The ICP algorithm is used to iteratively solve the target0 The fine registration transformation matrix R and T are used to obtain the optimal transformation between the model and the target point cloud model. R0 is the rotation matrix of the initial registration transformation, T0 is the translation vector of the initial registration transformation, R is the rotation matrix of the fine registration transformation, and T is the translation vector of the fine registration transformation.

[0011] Beneficial effects:

[0012] 1. The method of the present invention first performs voxel filtering on the original point cloud data to be registered and the target point cloud data, and then performs skeleton extraction and edge feature extraction to obtain the skeleton points and edge feature points of the point cloud data, which are used as the coarse registration object. The skeleton point cloud data and edge feature points are used to replace the massive original point cloud data, which saves computing resources and improves computing efficiency. At the same time, compared with the existing method, it is a simplification with higher accuracy.

[0013] 2. The present invention uses the PCA initial registration algorithm to calculate the covariance matrix of two groups of point clouds composed of skeleton points and edge feature points, and then solves the rotation matrix and translation vector based on the eigenvector. The rotation matrix and translation vector are applied to the original point cloud data to perform rigid body transformation operations to achieve coarse registration, which can achieve rapid registration of point cloud data and pave the way for subsequent fine registration.

[0014] 3. The present invention uses the ICP iterative algorithm to perform precise registration operations on the original point cloud data and the target point cloud data, thereby achieving a registration effect with high accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 The figure is a flow chart of the method of the present invention.

[0016] Figure 2 The initial position relationship and preprocessing of the point cloud data of the embodiment of the present invention; (a) the initial position relationship of the model, (b) the voxel filtering result of the source point cloud model, (c) the skeleton points and edge feature points.

[0017] Figure 3 The key points and registration results of the target point cloud model of an embodiment of the present invention; (a) skeleton points and edge feature points of the target point cloud data, (b) the registration result of an embodiment of the present invention. DETAILED DESCRIPTION

[0018] The present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0019] The present invention provides a registration method for point cloud data of non-rotating headgear structures, such as Figure 1 As shown, the registration method of the present invention requires preprocessing the original point cloud data and the target point cloud data, respectively calculating the skeleton points and edge feature points of the two point cloud data, and replacing the point cloud data with the skeleton points and edge feature points for registration operation. The preprocessing is divided into the following steps, first of which voxel filtering is performed.

[0020] The main idea of ​​voxel filtering sampling is to divide the original point cloud into reasonable voxels and then extract voxel grid points. The steps of voxelization downsampling are as follows:

[0021] (1) According to the point cloud data coordinate set {P i (x i ,y i ,z i )i=1,2,...,n},Pi∈R 3 , calculate the maximum value x in the three coordinate directions of X, Y, and Z max ,y max 、z max and the minimum value x min ,y min 、z min .

[0022] (2) Create a minimum bounding box based on the maximum value of the point cloud coordinates, with a volume of v, a length, width and height of l×w×h, and a side length that satisfies the following relationship:

[0023]

[0024] (3) Set the side length r of the voxel grid and calculate the size of the voxel grid:

[0025]

[0026] The right side of the equation Indicates rounding down.

[0027] (4) Calculate the index d of each point cloud in the voxel grid:

[0028]

[0029] (5) Sort the elements in d from small to large, calculate the centroid of each voxel grid, and replace all the points in the grid with the centroid to obtain the filtered point cloud data.

[0030] Original point cloud set P source0 and the target point cloud set P target0 After the voxel filtering of formulas (1) to (3), the new point cloud set P is obtained. source , P target .

[0031] After completing voxel filtering, Figure 2 As shown in (b), the point cloud center skeleton is extracted. For a point cloud dataset Solving the local L1 median skeleton of the data set with regularization terms can be summarized by the following optimization formula:

[0032]

[0033] Where: —Input point set; —A set of points randomly sampled from Q, θ represents the Gaussian weight, and the number of points |I|<<|J|. After minimizing formula (4), X becomes a regular form and becomes the "skeleton point set" at the center of the input model.

[0034] The point cloud obtained by real scanning measurement is often uneven. In order to prevent the skeleton points from tending to the dense area of ​​the point cloud and improve the ability to resist wild points, a local density measurement weight parameter is introduced. For the input source point cloud data, the following formula is used to calculate each point P in the point cloud data j The local density weight parameter of :

[0035] d j =1+∑ j′∈J\{j} θ(||p j -p j′ ||) (5)

[0036] Where: h d ——Neighborhood radius parameter.

[0037] Therefore, the definition is the point set of the current iteration, and the point set of the next iteration is X k+1 The iteration relationship of the skeleton point set is:

[0038]

[0039] Where: —Input point set and sampling point set weights; —weights of the sampling point set and the neighborhood; σ i —Distribution metric defined by the covariance matrix; μ—weight parameter that controls the size of the “repulsion”, with a value range of [0,1 / 2).

[0040] After adding the local density weight, the skeleton points are less affected by the uneven distribution of the point cloud, and a smaller neighborhood radius h is selected. d To reduce the interference of a few wild points. source , P target After iterative operations of formulas (4) to (6), the points converge to the skeleton point set P source_S , P target_S .

[0041] After the skeleton points of the original point cloud and the target point cloud are extracted, the boundary feature points are extracted. First, the least square plane needs to be fitted based on the k nearest neighbors and the target point, and the target point and its neighborhood points are projected onto the plane. Finally, the edge point detection operator is constructed using the idea of ​​weighted equivalent force and the density parameter of the point, and the boundary points are extracted by setting the threshold.

[0042] P source , P target The target point p i With local point set Fit the least squares plane Ax+By+Cz+D=0 as the micro-cutting plane of the target point k neighborhood, and define a0=-A / C, a1=-B / C, a2=-D / C. The target point p i (x i ,y i ,z i ) on the micro-cutting plane. i ′(x i ′,y i ′,z i ′) can be obtained by the following formula:

[0043]

[0044] Similarly, the neighborhood point set The projection point of the point in the plane can also be solved by formula (7).

[0045] The principle of determining whether the target point pi in the point cloud model is a boundary feature point is as follows: Figure 2 As shown. The coordinate axis represents the fitted least square plane, and the projection point p of the target point on the planei ′ and the projection point of the neighboring point Composition vector n j (j=1,2,…,k), which is normalized to vector n′ j (j=1,2,…,k), by solving the vector n′ j The vector sum of F i , the modulus of the vector sum can be used to obtain the resultant force parameter ||F i ||. According to the principle of vector addition, when the target point p i When it is an internal point, the neighborhood points are distributed more evenly (such as Figure 2 (a)), ||F i || has a smaller value; when the target point is located at the boundary position (such as Figure 2 (c) shows that the distribution of neighborhood points has no adjacent points on one side, ||F i || has a large value, and is compared with the set threshold to detect whether the target point is a boundary point. The core idea of ​​this method is similar to the process of finding the resultant force. j Equivalent to F i The various components of force.

[0046] The equivalent force F at the target point with inverse distance weight i for:

[0047]

[0048] w ij is the introduced inverse distance weight parameter, the expression is:

[0049]

[0050] Where: w ij ——Force vector n′ j The weight of ij —Target point p in the point cloud model i The Euclidean distance to the jth neighboring point, i.e. d ij =|n j |;l—power parameter that controls the influence of distance, with a value of [0.5,3].

[0051] The density parameter is constructed based on the mean of the Euclidean distance between the target point and its neighborhood. The calculation formula of the target point density parameter is as follows:

[0052]

[0053] According to the equivalent force parameter with inverse distance weight ||F i || and density parameter ρ i Construct a weighted edge point detection operator:

[0054]

[0055] Where: C i ——Target point p in the point cloud i The edge detection operator.

[0056] In the process of boundary point extraction, a threshold parameter ε is set to detect boundary points. i When the value is greater than ε, the target point is judged as a boundary point. source , P target After the traversal operation of the detection operator of formula (11), the edge feature point set P is extracted. source_T , P target_T .

[0057] Original point cloud data P source With the target point cloud data P target After preprocessing, the corresponding skeleton points and edge feature points can be obtained, and the point cloud data can be processed equivalently. source0 =P source_S ∪P source_T , P target0 =P target_S ∪P target_T .

[0058] Finally, the calculation process of the equivalent data based on the PCA initial registration algorithm and the ICP fine registration algorithm is as follows:

[0059] First, calculate the center point S of the original point cloud and the center point T of the target point cloud:

[0060]

[0061] Where: N is the number of equivalent point sets of the original point cloud; M is the number of equivalent point sets of the target point cloud.

[0062] Then, calculate the covariance matrix C of the two sets of point clouds S and C T As shown below:

[0063]

[0064] By the covariance matrix C S and C T Singular value decomposition can be performed to obtain the eigenvectors corresponding to the two matrices as follows:

[0065]

[0066] Where: U S , U T ——A 3×3 matrix.

[0067] Each matrix consists of three eigenvectors, which can be regarded as the three main directions of the corresponding point cloud. Therefore, the initial transformation matrix can be obtained as follows:

[0068]

[0069] Where: R0——initial rotation matrix; T0——initial translation vector.

[0070] According to formulas (12)-(15), the initial transformation matrices R0 and T0 of the source point cloud model can be obtained, and the original point cloud data P source0 After the initial transformation, we get P' source0 Finally, P' source0 The ICP algorithm is used to iteratively solve the target0 The precise registration transformation matrix R and T are used to obtain the optimal transformation between the model and the target point cloud model. The process can be described by the following relationship:

[0071]

[0072] Figure 3 The key points and registration results of the target point cloud model of the embodiment of the present invention, wherein (a) are the skeleton points and edge feature points of the target point cloud data, and (b) is the registration result of the embodiment of the present invention.

[0073] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the protection scope of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A registration method for point cloud data of non-rotating headgear structures, characterized in that: The original point cloud data is converted into a combination of skeleton data and edge feature data. The combined data of skeleton data and edge data are used for initial registration operation to obtain the initial transformation matrix. Then, based on the initial registration, a fine registration operation is performed to obtain the fine registration transformation matrix to achieve registration of the original point cloud and the target point cloud.

2. The registration method according to claim 1, characterized in that: The original point cloud data and the target point cloud data are preprocessed, the skeleton points and edge feature points of the two point cloud data are calculated respectively, and the skeleton points and edge feature points are used to replace the point cloud data for registration operation.

3. The registration method according to claim 2, characterized in that: The preprocessing includes: performing voxel filtering on the original point cloud and extracting the central skeleton of the point cloud.

4. The registration method according to claim 2 or 3, characterized in that: The preprocessing includes: extracting edge feature points, specifically: first fitting a least square plane based on k nearest neighbors and the target point, projecting the target point and its neighborhood points onto the plane, and finally constructing an edge point detection operator using the idea of ​​weighted equivalent force and the density parameters of the points, and extracting edge feature points by setting a threshold.

5. The registration method according to claim 4, characterized in that: Based on the PCA initial registration algorithm, the initial registration transformation matrix R0 and T0 of the original point cloud model are obtained, and the original point cloud data P source0 After the initial transformation, we get P' source0 ; Finally, P' source0 The ICP algorithm is used to iteratively solve the target0 The fine registration transformation matrix R and T are used to obtain the optimal transformation between the model and the target point cloud model. R0 is the rotation matrix of the initial registration transformation, T0 is the translation vector of the initial registration transformation, R is the rotation matrix of the fine registration transformation, and T is the translation vector of the fine registration transformation.

Images