Point cloud registration method based on geometric information

Abstract

The invention belongs to the technical field of point cloud registration, and particularly relates to a point cloud registration method based on geometric information. Obtaining source point cloud data and target point cloud data; step 2; extracting geometric features of the source point cloud and the target point cloud, wherein the geometric features comprise normal vectors, curvatures and shape contexts of points; a third step; performing feature matching based on the extracted geometric features to obtain a matching point pair set; step 4; calculating a transformation matrix between the point clouds by using the matching point pair set; (5) carrying out drying; and optimizing the transformation matrix to realize accurate registration of the point cloud. According to the method, geometric information is used for registration, and compared with a traditional distance-based registration method, the dependence on the initial position is low, and noise and partially overlapped point clouds can be better processed.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud registration, and specifically relates to a point cloud registration method based on geometric information. Background Art

[0002] Point cloud registration is an upstream task in 3D vision tasks. It aims to solve a transformation matrix to align point clouds from two different viewpoints in the same scene to the same coordinate system, and is widely used in fields such as 3D reconstruction, autonomous driving, simultaneous localization and mapping, and robot pose estimation.

[0003] Most current point cloud registration methods are based on correspondence-based point cloud registration, which mainly includes the following four processes. First, features are extracted from the two input frames of point clouds respectively. Next, correspondences are selected in the feature space. Then, outliers are filtered. Finally, a transformation matrix is solved using a relatively robust pose estimator based on the correspondences. Recently, correspondence-based point cloud registration methods generally use a coarse-to-fine registration method. They first find correspondences at the superpoint level and then perform fine registration at the point level.

[0004] Point cloud registration methods can be classified into single-modal point cloud registration methods and multi-modal point cloud registration methods according to the modality.

[0005] Single-modal Point Cloud Registration Method

[0006] Single-modal point cloud registration methods refer to those that only use the point cloud modality for registration. Currently, most methods are single-modal point cloud registration methods. The success of point cloud registration depends on the extraction of point cloud geometric features. Due to the disorder of point clouds, it is difficult to extract features. Initially, a shared multi-layer perceptron network was used separately for each point for feature extraction. Later, a hierarchical network structure was used to make the extracted features adapt to different point cloud densities. Inspired by image convolutional neural networks, some scholars proposed a point cloud feature extraction method based on variable kernel point convolution, which can adapt to objects of different shapes and pays more attention to the geometric structure characteristics of point clouds. In the point cloud registration task, people also pay more attention to the geometric consistency between two frames of point clouds. After using kernel point convolution for downsampling to extract superpoints, some people enhance the features by injecting the distance and angle consistency information between superpoints, which significantly improves the robustness of the algorithm. The performance of current single-modal point cloud registration methods has achieved good results. However, since these methods only use the single modality of point clouds, the extracted features only have geometric information and lack texture information, which makes them vulnerable to repeated geometric regions and weak geometric regions in low overlap rate scenarios, resulting in final registration failure. In addition, when current methods perform geometric consistency embedding, they often consider the relationship between each superpoint, mixing in ambiguous geometric information, which reduces the distinguishability of point cloud features and is not conducive to the correct selection of corresponding relationships.

[0007] Multi-modal point cloud registration methods

[0008] In recent years, with the rise of deep learning, through the use of technologies such as convolutional neural networks for feature extraction of images, significant breakthroughs have been made in tasks such as image classification and object detection. Therefore, in the past two years, some multi-modal point cloud registration algorithms based on images and point clouds have emerged in the point cloud registration task. Initially, some people tried an implicit multi-modal feature fusion method. It first encodes and extracts features from the point cloud and its corresponding image separately, and then uses the attention mechanism to globally fuse the features of the point cloud and the image to achieve the purpose of enhancing the point cloud features. Finally, a decoder is used to obtain point cloud features with both texture information and structural information. This method provides a reference for the fusion of multi-modal information. However, due to the use of an implicit image feature fusion method, the distinguishability of the point cloud features is reduced, resulting in its inability to achieve good performance in low overlap rate point cloud registration tasks. Later, some people proposed using only one image in the training stage, which can simultaneously cover the partially overlapping regions of the two input frames of point clouds. This method is beneficial for finding corresponding relationships and ensures that the correct transformation matrix can be solved later.

[0009] In addition, some people use the currently mature image matching technology to first extract the corresponding pixels from the image, then project them onto the point cloud using the internal and external parameter matrices, initialize the features of the point cloud as the pixel features within the two-dimensional overlapping area, and then use the conventional kernel convolution method to extract the features of the point cloud.

[0010] Traditional point cloud registration methods, such as the Iterative Closest Point (ICP) algorithm, although can achieve point cloud registration to a certain extent, have problems such as being sensitive to the initial position, high computational complexity, and being prone to falling into local optima. Therefore, a point cloud registration method based on geometric information is improved and designed. Summary of the Invention

[0011] In view of the above deficiencies in the prior art, the present invention provides a point cloud registration method based on geometric information to solve the problems in the above background technology.

[0012] To solve the above technical problems, the present invention adopts the following technical solutions:

[0013] The point cloud registration method based on geometric information includes the following steps:

[0014] The first step: Obtain the source point cloud data and the target point cloud data;

[0015] The second step: Extract the geometric features of the source point cloud and the target point cloud, and the geometric features include the normal vector, curvature, and shape context of the points;

[0016] The third step: Perform feature matching based on the extracted geometric features to obtain a set of matching point pairs;

[0017] The fourth step: Calculate the transformation matrix between the point clouds using the set of matching point pairs;

[0018] The fifth step: Optimize the transformation matrix to achieve accurate registration of the point clouds.

[0019] Furthermore, the extraction steps of the geometric features in the second step include the calculation of the normal vector, the calculation of the curvature, and the calculation of the shape context.

[0020] Furthermore, the calculation of the normal vector includes the following steps;

[0021] Neighborhood point selection; For point p i , select its neighborhood point set N=(p i );

[0022] Covariance matrix calculation; Calculate the covariance matrix C of the neighborhood points. Let Pi=(x i ,y i ,z i ), and the neighborhood point Pj=(x j ,yj , z j ), the elements of the covariance matrix C are calculated as follows:

[0023]

[0024] The covariance matrix C can be expressed as: Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C to obtain eigenvalues λ 1 , λ 2 , λ 3 and the corresponding eigenvectors and satisfy λ 1 ≥ λ 2 ≥ λ 3 , and the eigenvector corresponding to the smallest eigenvalue λ 3 is the normal vector of the point Pi. That is, the normal vector of the point Pi.

[0025] Furthermore, the calculation of the curvature includes the following steps;

[0026] Neighborhood point selection: For the point p i , select its neighborhood point set N = (p i );

[0027] Least squares plane fitting;

[0028] Let the neighborhood point set N(P i ) = {p j = (x j , y j , z j ) | j = 1, 2,..., k}, and the plane equation can be expressed as z = Ax + By + C (assuming the plane normal vector is (a, b, c) = (A, B, -1), then the error function is: By minimizing this error function, A, B, and C can be solved. Take the partial derivatives of E(A, B, C) with respect to A, B, and C respectively and set them to zero to obtain a linear equation system: where the summation is performed on the neighborhood points. Solving this linear equation system can obtain the coefficients of the plane equation;

[0029] Curvature calculation;

[0030] After obtaining the plane equation z = Ax + By + C, its curvature K can be calculated by the following formula:

[0031] Furthermore, the calculation of the shape context includes the following steps;

[0032] Log-polar coordinate division: For the input point set or the point set on the shape contour P = {p 1 , p2 ,..., p n}, for each point P i ∈ P, a logarithmic polar coordinate system is established with it as the center, and the angular range [0, 2π] of the polar coordinates is divided into M regions, and the angular interval of each region is The distance range is equally divided into N regions in logarithmic form, and the boundary of the K-th region is

[0033] Construct a feature histogram; for point P i for each neighborhood point p j belonging to N = (p i ), the neighborhood points can be obtained by K-nearest neighbor or fixed-radius search, calculate its polar coordinates relative to P i According to and a j map k to the corresponding logarithmic polar coordinate region, for each region k, count the number of neighborhood points falling into this region, and form an M×N-dimensional histogram H i (K); this histogram is the shape context feature of point p i ; by traversing each point in the point set p, calculate the shape context feature for each point, and obtain the shape context matrix H = {h 1 , h 2 ,..., h n}, where h i represents the shape context feature of point p i .

[0034] Furthermore, the feature matching in the third step includes shape matching based on shape context, and the shape matching based on shape context includes the following steps;

[0035] Feature extraction, for the two shapes to be matched, calculate their shape context feature matrices and

[0036] Feature matching; for each shape context feature find the most similar shape context feature in H 2 Establish a set of matching point pairs M = {(p i1 , p j1 ), (p i2 , p j2 ),...}, where p ik and p jk are point pairs with similar shape context features;

[0037] ​​Matching optimization: According to the set M of matching point pairs, use the transformation model to align the shapes, and use the least squares method or other optimization algorithms to solve the transformation matrix T to minimize the error of the matching point pairs after transformation, that is: Apply the transformation matrix T to all points of the shape to complete the matching of the shape; Local optimization: Take the distance between point pairs and the consistency of normal vectors as the optimization objective function, where ω is a weight coefficient used to balance the importance of position and normal vector consistency.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The present invention uses geometric information for registration. Compared with traditional distance-based registration methods, it has low dependence on the initial position and can better handle noisy and partially overlapping point clouds.

[0040] 2. Considering multiple geometric features, it improves the accuracy and robustness of matching and can more accurately describe the local and global geometric structures of point clouds.

[0041] 3. Adopting a staged optimization process not only ensures computational efficiency but also achieves high registration accuracy, providing a good foundation for subsequent point cloud processing and analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is a flowchart of the point cloud registration method based on geometric information of the present invention;

[0043] Figure 2 It is a schematic diagram of the feature points of the straight line extracted from the three-dimensional point cloud;

[0044] Figure 3 It is a schematic diagram of the straight line fitting result;

[0045] Figure 4 Schematic diagram of the result of three-dimensional point cloud registration; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] In order to enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be further described below with reference to the drawings and embodiments.

[0047] Among them, the drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as a limitation to this patent; In order to better illustrate the embodiments of the present invention, some components in the drawings will be omitted, enlarged or reduced, and do not represent the dimensions of actual products; For those skilled in the art, it is understandable that some well-known structures and their descriptions in the drawings may be omitted.

[0048] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if terms such as "upper", "lower", "left", "right", "inner", "outer", etc. are used to indicate the orientation or positional relationship, it is based on the orientation or positional relationship shown in the accompanying drawings. This is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and cannot be construed as a limitation of this patent. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0049] In the description of the present invention, unless otherwise clearly specified and limited, if terms such as "connection" are used to indicate the connection relationship between components, this term should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or integrated; it can be a mechanical connection or an electrical connection; it can be directly connected or indirectly connected through an intermediate medium, and it can be the communication inside two components or the interaction relationship between two components. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0050] Traditional point cloud registration methods, such as the Iterative Closest Point (ICP) algorithm, although can achieve point cloud registration to a certain extent, have problems such as being sensitive to the initial position, high computational complexity, and being easily trapped in local optima. The present invention proposes a point cloud registration method based on geometric information. This method utilizes geometric features in the point cloud data, including but not limited to information such as the normal vector, curvature, and shape context of points. By extracting and matching these geometric features, and then calculating the transformation matrix between the point clouds, the source point cloud is registered to the coordinate system of the target point cloud.

[0051] The following is a specific embodiment of the present invention:

[0052] Data preparation: Obtain the source point cloud data and the target point cloud data, which can be data from a laser scanner, a depth camera, or other 3D sensors.

[0053] Normal vector calculation

[0054] Geometric feature extraction: For the input source point cloud Ps and target point cloud pt, extract their geometric features respectively. For each point pi ∈ ps and qj ∈ pt, calculate their normal vectors ni and nj, and the normal vectors can be solved by performing principal component analysis (PCA) on the covariance matrix of the neighborhood points.

[0055] For a point pi in a point cloud P = {p1, p2,..., pn}, we usually estimate the normal vector of this point by considering its neighboring points. The following are the steps to calculate the normal vector using PCA (Principal Component Analysis):

[0056] Neighboring point selection:

[0057] For the point pi, select its neighboring point set N = (pi). Neighboring points can be selected in different ways, such as a neighborhood with a fixed radius (e.g., using the K-nearest neighbor algorithm or fixed-radius search). For example, using the K-nearest neighbor algorithm, find the K points closest to pi as neighboring points, where the choice of k is usually determined according to the density and noise level of the point cloud.

[0058] Covariance matrix calculation:

[0059] Calculate the covariance matrix C of the neighboring points. Let Pi = (xi, yi, zi) and the neighboring point Pj = (xj, yj, zj), then the elements of the covariance matrix C are calculated as follows:

[0060]

[0061] The covariance matrix C can be expressed as:

[0062]

[0063] Eigenvalue decomposition:

[0064] Perform eigenvalue decomposition on the covariance matrix C to obtain the eigenvalues λ 1 , λ 2 , λ 3 and the corresponding eigenvectors and satisfy λ 1 ≥ λ 2 ≥ λ 3 .

[0065] The eigenvector corresponding to the smallest eigenvalue λ 3 is the normal vector of the point Pi. Because this eigenvector corresponds to the direction with the smallest data variance, that is, the direction that best describes the local plane of the point cloud at this point, which is the normal vector. That is, the normal vector of the point Pi. Because this eigenvector corresponds to the direction with the smallest data variance, that is, the direction that best describes the local plane of the point cloud at this point, which is the normal vector.

[0066] Calculate the curvature information ki and kj of each point. The method of local surface fitting can be used, such as least squares plane fitting, and the curvature is calculated according to the residuals of the fitting plane.

[0067] Curvature calculation;

[0068] Curvature calculation based on local surface fitting:

[0069] For a point Pi in the point cloud P = {p1, p2,..., pn}, we can use the method of local least squares plane fitting to calculate its curvature. The following are the specific steps:

[0070] Neighborhood point selection:

[0071] For the point pi, select its neighborhood point set N = (pi). Neighborhood points can be selected in different ways, such as a neighborhood with a fixed radius (e.g., using the K-nearest neighbor algorithm or fixed-radius search).

[0072] Least squares plane fitting:

[0073] Assume the plane equation is; ax + by + cz + d = 0. We want to find a set of coefficients a, b, c, d such that the sum of the squares of the distances from the plane to the neighborhood points is minimized. The problem can be expressed as a least squares problem.

[0074] Let the neighborhood point set N(P i ) = {p j = (x j , y j , z j )|j = 1, 2,...k}. The plane equation can be expressed as z = Ax + By + C (assuming the plane normal vector is (a, b, c) = (A, B, -1), then the error function is: By minimizing this error function, A, B, and C can be solved. Take the partial derivatives of E(A, B, C) with respect to A, B, and C respectively and set them to zero to obtain a system of linear equations: where the summation is over the neighborhood points.

[0075] Solving this system of linear equations can obtain the coefficients of the plane equation.

[0076] Curvature calculation:

[0077] After obtaining the plane equation z = Ax + By + C, its curvature K can be calculated by the following formula:

[0078]

[0079] Calculation method of shape context

[0080] Log-polar coordinate division

[0081] For the input point set or the point set on the shape contour P = {p1, p2,..., pn}, for each point Pi ∈ P, establish a log-polar coordinate system with it as the center. Divide the angular range [0, 2π] of the polar coordinates into M regions, and the angular interval of each region is Divide the distance range Be equally divided into N regions in logarithmic form, where the boundary of the Kth region is This logarithmic polar coordinate division method can describe shapes at different scales and enhance the robustness to shape scaling.

[0082] Feature histogram construction:

[0083] For each neighborhood point pj of point Pi belonging to N=(pi), the neighborhood points can be obtained through K-nearest neighbor or fixed-radius search, and calculate its polar coordinates relative to Pi According to and θi j Map k to the corresponding logarithmic polar coordinate region.

[0084] For each region k, count the number of neighborhood points falling into this region to form an M×N-dimensional histogram H i (K). This histogram is the shape context feature of point pi.

[0085] By traversing each point in the point set p, calculate the shape context feature for each point to obtain the shape context matrix H={h 1 ,h 2 ,...,h n} where hi represents the shape context feature of point pi.

[0086] Similarity measurement of shape context

[0087] To measure the similarity between two shape context features, the chi-square distance (x 2 distance) is used as the measurement. For two shape context features hi and hj, their chi-square distance is calculated as follows:

[0088] where h i (k) and h j (k) respectively represent the statistical values of h i and h j in the kth region. This distance measurement can effectively quantify the difference between two shape contexts. The smaller the value, the more similar they are, and it has robustness to rotation, scaling, and partial deformation.

[0089] Shape matching method based on shape context

[0090] Feature extraction

[0091] For the two shapes to be matched, calculate their shape context feature matrices respectively and

[0092] Feature matching

[0093] For each shape context feature In H 2 Find the most similar shape context feature You can use the nearest neighbor search or the two-way matching strategy, with the chi-square distance as the similarity measure. The two-way matching strategy can improve the reliability of the matching and avoid false matching.

[0094] Establish a matching point pair set M = {(p i1 ,p j1 ),(p i2 ,p j2 ),...}, where p ik and p jk are point pairs with similar shape context features.

[0095] Matching Optimization

[0096] According to the set of matching point pairs M, the shapes are aligned using a transformation model (such as rigid transformation, affine transformation, etc.). The least squares method or other optimization algorithms can be used to solve the transformation matrix T to minimize the error of the matching point pairs after transformation, that is: The transformation matrix T can be applied to all points of the shape to complete the shape matching.

[0097] The transformation matrix is ​​further optimized using a local optimization algorithm, such as nonlinear least squares, to improve the registration accuracy. The distance between point pairs and the consistency of normal vectors are used as the optimization objective function, for example: Where ω is a weight coefficient that balances the importance of position and normal consistency.

[0098] Point cloud registration method based on straight line features

[0099] In the field of point cloud processing, point cloud registration is a key technology to unify point cloud data in different coordinate systems into the same coordinate system. This paper will introduce in detail a method of point cloud registration using 3D straight line features. The method mainly includes the following steps: detecting straight line feature points in 3D point clouds, fitting 3D straight lines, matching 3D straight lines, and using the matched straight lines for point cloud registration.

[0100] 1. Fitting of three-dimensional straight line

[0101] In three-dimensional space, there are many ways to fit a straight line. The common representation methods are as follows:

[0102] 1. General formula: A straight line can be regarded as the intersection of two planes, which is obtained by solving two plane equations together, that is, Ax+By+Cz+D=0.

[0103] 2. Point-direction form (standard equation): Let (m, n, p) be the direction vector of the line equation, and (x_0, y_0, z_0) be a point on the line. Then the equation is m(x - x_0) + n(y - y_0) + p(z - z_0) = 0.

[0104] 3. Parametric equations: x = x_0 + mt, y = y_0 + nt, z = z_0 + pt, where (x_0, y_0, z_0) is a point on the line and t is a parameter.

[0105] In practical applications, the least squares method is often used to fit the line equation. Its core is to minimize the sum of the squares of the residuals between the discrete points (actual points) and the points on the fitted line (ideal points). Suppose there is a set of three-dimensional point data, and the parameters of the fitted line are solved by constructing a matrix and solving the inconsistent equations.

[0106] Steps for three-dimensional line fitting

[0107] 1. Detect line feature points: In three-dimensional point cloud data, appropriate algorithms (such as curvature-based methods, RANSAC algorithm, etc.) are used to detect line feature points. These algorithms can identify the points that may be on the line according to the local geometric features of the point cloud.

[0108] 2. Fit the three-dimensional line: Using the detected line feature points, according to the above line representation method and the least squares method, calculate the parameters of the three-dimensional line. For example, if the parametric equations are used to represent the line, (x_0, y_0, z_0) and the direction vector (m, n, p) are determined by the least squares method.

[0109] II. Matching of three-dimensional lines

[0110] After obtaining the three-dimensional lines fitted from two sets of point clouds respectively, these lines need to be matched. The purpose of matching is to find the corresponding lines in the two sets of point clouds for subsequent point cloud registration. The method of line matching can be based on the direction vector and position information of the lines.

[0111] 1. Direction vector matching: Calculate the angle between the direction vectors of two lines. The smaller the angle, the closer the directions of the two lines. Set a threshold. When the angle is less than the threshold, the two lines are considered to have a matching direction.

[0112] 2. Position matching: Consider the position relationship between the line in the source point cloud and the line in the target point cloud. If the distance and the position angle relationship are within a certain range, the two lines are considered to have a matching position. By comprehensively considering the results of direction vector and position matching, the corresponding line pairs in the two sets of point clouds are determined.

[0113] III. Point cloud registration using paired lines

[0114] Point cloud registration essentially aims to find the transformation relationship between two coordinate systems, which can be described by a 3×3 rotation matrix R and a three-dimensional translation vector t. That is, registration is to solve for (R, t).

[0115] Steps of Point Cloud Registration Based on Paired Lines

[0116] 1. Calculate the rotation matrix: For each pair of matching lines, use relevant algorithms to calculate the rotation matrix based on information such as the direction vectors of the lines. For example, through the method of singular value decomposition (SVD), let the covariance matrix of two sets of corresponding point sets be H, perform SVD decomposition on H as H = UΣV^T, and the rotation matrix R = VU^T.

[0117] 2. Calculate the translation vector: After obtaining the rotation matrix, calculate the translation vector through the transformation relationship of the point cloud centroids. Let the centroids of the two point clouds be c_1 and c_2 respectively, then the translation vector t = c_2 - Rc_1.

[0118] 3. Point cloud transformation: Apply the calculated rotation matrix R and translation vector t to one of the point clouds to achieve point cloud registration. That is, for each point p_i = (x_i, y_i, z_i) in the point cloud P, the point p_i' after registration transformation is p_i' = Rp_i + t.

[0119] Through the above point cloud registration method based on three-dimensional line features, the point cloud data in different coordinate systems can be effectively unified into the same coordinate system, providing a basis for subsequent point cloud analysis, modeling, etc.

[0120] The above are only embodiments of the present invention. The circuits, electronic components, and modules involved are all prior arts and can be fully implemented by those skilled in the art without further elaboration. The content protected by this application does not involve improvements to software and methods. Common knowledge such as specific structures and characteristics in the solution is not described in detail here. Those of ordinary skill in the art know all the common general technical knowledge in the technical field to which the invention belongs before the application date or priority date, can know all the prior arts in this field, and have the ability to apply conventional experimental means before this date. Those of ordinary skill in the art can, under the inspiration given by this application, combine their own abilities to complete and implement this solution. Some typical well-known structures or well-known methods should not be an obstacle for those of ordinary skill in the art to implement this application. It should be noted that for those skilled in the art, without departing from the structure of the present invention, several deformations and improvements can still be made, which should also be regarded as the protection scope of the present invention, and these will not affect the implementation effect of the present invention and the practicality of the patent.

Claims

1. The point cloud registration method based on geometric information is characterized by The following steps are involved: first step; Obtain source point cloud data and target point cloud data; Step 2: Extract geometric features of the source point cloud and the target point cloud, wherein the geometric features include the normal vector, curvature and shape context of the point and the straight line feature; The third step is to perform feature matching based on the extracted geometric features to obtain a set of matching point pairs; Step 4: Calculate the transformation matrix between point clouds using the matching point pair set; Step 5: Optimize the transformation matrix to achieve accurate alignment of the point cloud.

2. The point cloud registration method based on geometric information according to claim 1, characterized in that: The geometric feature extraction steps in the second step include the calculation of normal vectors, curvature, and shape context.

3. The point cloud registration method based on geometric information according to claim 2, characterized in that: The calculation of the normal vector comprises the following steps: Neighborhood point selection; for point p i , select its neighborhood point set N = (p i ); Covariance matrix calculation: Calculate the covariance matrix C of the neighborhood points, set Pi = (x i ,y i ,z i ), the neighboring point Pj=(x j ,y j ,z j ), then the elements of the covariance matrix C are calculated as follows: The covariance matrix C can be expressed as: Eigenvalue decomposition: Perform eigenvalue decomposition on the covariance matrix C to obtain eigenvalues ​​λ1, λ2, λ3 and corresponding eigenvectors And satisfying λ1≥λ2≥λ3, the eigenvector corresponding to the minimum eigenvalue λ3 This is the normal vector of point Pi.

4. The point cloud registration method based on geometric information according to claim 2, characterized in that: The calculation of the curvature comprises the following steps: Neighborhood point selection: For point p i , select its neighborhood point set N = (p i ); Least squares plane fitting; Let the neighborhood point set N(P i )={p j =(x j ,y j ,z j )|j=1,2,...k}, the plane equation can be expressed as z=Ax+By+C (assuming the plane normal vector is (a,b,c)=(A,B,-1), then the error function is: By minimizing this error function, we can solve A, B, and C. We can take the partial derivatives of E(A, B, C) with respect to A, B, and C respectively and set them to zero, and obtain a system of linear equations: The summation is performed on the neighborhood points. Solving this system of linear equations can obtain the coefficients of the plane equation; Curvature calculation; After obtaining the plane equation z = Ax + By + C, its curvature K can be calculated by the following formula:

5. The point cloud registration method based on geometric information according to claim 2, characterized in that: The calculation of the shape context comprises the following steps: Logarithmic polar coordinate partitioning; for the input point set or point set on the shape contour P = {p1, p2, ..., p n }, for each point P i ∈P, a logarithmic polar coordinate system is established with it as the center, and the angle range of the polar coordinates [0,2π] is divided into M regions, and the angle interval of each region is Distance Range Logarithmic The form is divided into N regions, where the boundary of the Kth region is Construct feature histogram; For point P i Each neighborhood point p j Belong to N=(p i ), the neighboring points can be obtained by K nearest neighbors or fixed radius search to calculate their relative to P i Polar coordinates according to and θi j Map k to the corresponding logarithmic polar coordinate area. For each area k, count the number of neighborhood points that fall into the area to form an M×N dimensional histogram H i (K); the histogram is point p i By traversing each point in the point set p and calculating its shape context feature for each point, the shape context matrix H = {h1,h2,...,h n } where h i Represents point p i Shape context features.

6. The point cloud registration method based on geometric information according to claim 1, characterized in that: The feature matching in the third step includes shape matching based on shape context, and the shape matching based on shape context includes the following steps: Feature extraction: for the two shapes to be matched, their shape context feature matrices are calculated respectively. and Feature matching; for each shape context feature Find the most similar shape context feature in H2 Establish a matching point pair set M = {(p i1 ,p j1 ),(p i2 ,p j2 ),...}, where p ik and p jk are point pairs with similar shape context features; Matching optimization, based on the set of matching point pairs M, uses the transformation model to align the shapes, and uses the least squares method or other optimization algorithms to solve the transformation matrix T to minimize the error of the matching point pairs after transformation, that is: Apply the transformation matrix T to all points of the shape to complete the shape matching; Local optimization, taking the distance between point pairs and the consistency of normal vectors as the optimization objective function, Where ω is a weight coefficient that balances the importance of position and normal consistency.

Images