A point cloud registration method based on planar primitives in structured scenes

Abstract

The present invention belongs to the field of computer vision technology and relates to a point cloud registration method based on plane primitives in a structured point cloud scene. The method relies on a pre-processing module in the system to filter and downsample to output point cloud information, a plane extraction module to extract planes with obvious features, and a plane matching module to search for corresponding surfaces; a transformation matrix calculation module calculates the transformation matrix of the successfully matched plane features, and an optimal selection module selects the transformation matrix with the highest registration accuracy. The method filters, downsamples, and segments the source point cloud and the target point cloud to be registered to obtain planes, calculates plane feature descriptors and angle feature descriptors and performs geometric matching based on them, obtains a set of successfully matched planes and calculates the transformation matrix, selects the transformation matrix with the highest point cloud overlap to perform point cloud registration and quantization, thereby realizing point cloud registration in a structured scene. The method improves the accuracy of point cloud registration in structured scenes and reduces the running time of point cloud registration.

Description

Technical Field

[0001] The present invention belongs to the field of computer vision technology, and relates to a point cloud registration method based on plane primitives for structured scenes, and in particular to a point cloud registration method based on plane primitives in structured point cloud scenes. Background Art

[0002] With the rise of artificial intelligence and the rapid development of three-dimensional sensors, LiDAR (Light Detection and Ranging) has been used more frequently in various structured scenes as an advanced and efficient sensor. However, due to the limitations of LiDAR information collection and the occlusion of lasers, single-angle collection cannot obtain a complete data description of the scene or object. Therefore, when collecting point cloud data of a large scene or scanning the surface information of a single object, point cloud registration technology is required to completely restore the geometric information of the measured scene or object surface. The purpose of point cloud registration is to find the optimal transformation matrix to convert different point clouds into the same coordinate system.

[0003] Traditional point cloud registration methods, such as the Iterative Closest Point (ICP) algorithm and the Normal Distributions Transform (NDT) algorithm, experience significant degradation in registration performance in structured scenes. To address this registration challenge, the present invention leverages the abundant planar elements present in structured scenes. Point cloud registration based on planar primitives involves five steps: point cloud preprocessing, plane extraction, plane matching, transformation matrix calculation, and optimal selection. Plane extraction is a crucial prerequisite for achieving high-precision registration results. Commonly used plane extraction methods include Random Sample Consensus (RANSAC) and the Hough transform (HT). RANSAC and HT algorithms are model-based plane fitting methods with high segmentation accuracy and minimal noise interference. However, these two algorithms require precise initialization, and the quality of plane extraction significantly degrades when point cloud features are blurred. Plane matching is a key step in the registration process based on planar primitives, and its precision directly determines the accuracy of registration. Existing plane matching algorithms include using the area and centroid of corresponding planes for plane matching, and using the angle between the intersection of two planes as a feature descriptor. However, these plane matching methods only consider the correspondence between two planes simultaneously, while solving the transformation matrix requires at least three planes. Therefore, their matching results are extremely dependent on the quality of point cloud features. This application aims to address the problem of low point cloud registration accuracy in structured scenes. Summary of the Invention

[0004] The purpose of the present invention is to solve the problem of low accuracy of point cloud registration in structured point cloud scenes, and proposes a point cloud registration method based on plane primitives for structured scenes. The method optimizes the plane extraction and plane matching processes in plane registration based on a plane feature descriptor of distribution characteristics and an angle feature descriptor based on geometric characteristics, respectively. Specifically: in the plane extraction part, a new plane feature descriptor is proposed according to the divergence degree of the center of mass on the plane, the distribution and aggregation degree of points, and the number of points; in the plane matching part, an angle feature descriptor based on three plane center of mass and normal vector is proposed to realize point cloud registration in structured scenes; the method improves the accuracy of point cloud registration in structured scenes and reduces the running time of point cloud registration.

[0005] In order to achieve the above object, the present invention is implemented through the following technical solutions:

[0006] The point cloud registration method based on plane primitives for structured scenes relies on a point cloud registration system, including a preprocessing module, a plane extraction module, a plane matching module, a transformation matrix calculation module, and an optimal selection module;

[0007] The preprocessing module outputs point cloud information to the plane extraction module. The plane extraction module extracts the plane with obvious features and inputs it into the plane matching module. After the plane matching module searches for the corresponding surface, the transformation matrix calculation module calculates the transformation matrix of the matched plane features, and the optimal selection module selects the transformation matrix with the highest registration accuracy.

[0008] The working process of the point cloud registration system is as follows:

[0009] First, two frames of point clouds to be registered are input into the preprocessing module, which performs filtering and downsampling in sequence to obtain the processed point cloud information;

[0010] Second, the processed point cloud information output by the preprocessing module is input into the plane extraction module to extract the plane with obvious features;

[0011] Third, the extracted planes with obvious features are input into the plane matching module to search for corresponding surfaces, and a set of successfully matched planes is obtained;

[0012] Fourth, the transformation matrix calculation module calculates the transformation matrix of the plane features in the successfully matched plane set;

[0013] Fifth, the transformation matrix is ​​input into the optimal selection module to select the transformation matrix with the highest registration accuracy.

[0014] The point cloud registration method based on plane primitives for structured scenes includes filtering, downsampling, and segmenting the source point cloud and the target point cloud to be registered. After obtaining the plane, the following operations are performed:

[0015] S4, using the Euclidean distance from the points in the segmented plane to the plane centroid, the minimum nearest neighbor distance of each point, and the number of points to construct a plane feature descriptor and extract the plane with obvious features of the source point cloud and the target point cloud based on it;

[0016] S5. Select three planes from each of the source point cloud and the target point cloud with obvious features, calculate the geometric feature descriptors and angular feature descriptors of the two planes in sequence, and perform geometric matching based on the descriptors.

[0017] S6, traverse all three planes in the source point cloud and the target point cloud with obvious features, repeat S5 to obtain several groups of successfully matched planes, and form a set of successfully matched planes;

[0018] S7, calculating the transformation matrix of the plane features in the successfully matched plane set to form a transformation matrix set;

[0019] S8. Select the transformation matrix with the highest point cloud overlap for point cloud registration and quantification.

[0020] The filtering specifically includes: filtering the outliers in the source point cloud and the target point cloud to be registered to obtain two frames of filtered point cloud data.

[0021] The downsampling is to downsample the two frames of filtered point cloud data to obtain preprocessed point cloud data; the segmentation is specifically to segment the plane information of the preprocessed point cloud data using the region growing method and the RANSAC algorithm in sequence to obtain the segmented plane.

[0022] The filtering process described in S1 is filtering using a straight-through filter.

[0023] The downsampling downsamples the source point cloud and the target point cloud respectively through a voxelized grid.

[0024] The extraction of planes with obvious features from the source point cloud and the target point cloud in S4 is achieved by sorting the plane feature descriptors and selecting the l planes with the largest plane feature descriptors.

[0025] The value range of l is 8 to 12.

[0026] The geometric matching described in S5 is specifically as follows: if the angle difference between the geometric feature descriptor and the angular feature descriptor is less than 1 degree and the offset difference is less than 0.1 meter, the matching is successful; otherwise, the matching fails.

[0027] S7 calculates the transformation matrix of the plane features in the successfully matched plane set using the face-to-face SVD algorithm.

[0028] S8 includes the following sub-steps:

[0029] S81: Use voxelized grids to downsample the source point cloud and the target point cloud respectively to obtain downsampled source point cloud and target point cloud;

[0030] S82: Select a transformation matrix from the transformation matrix set obtained in S7, and transform the downsampled source point cloud into a new point cloud according to the transformation matrix;

[0031] S83: Calculate the minimum distance between each point in the new point cloud and the point in the sampled target point cloud. If the distance is less than or equal to 1 cm, the point cloud coincidence is increased by 1. Otherwise, the closest point is discarded to obtain the point cloud coincidence corresponding to the transformation matrix.

[0032] S84: Jump to S82 until the transformation matrices in the transformation matrix set of S7 are traversed and the transformation matrix with the largest point cloud overlap is selected as the optimal transformation matrix.

[0033] Beneficial effects

[0034] The point cloud registration method based on planar primitives for structured scenes proposed in this invention has the following advantages compared with the existing point cloud registration method based on planar primitives:

[0035] 1. The method can improve the noise resistance of plane extraction and reduce the required initial parameters, ensuring that the extracted plane has good feature information, effectively improving the accuracy of the registration algorithm, and effectively solving the problem that existing plane extraction algorithms require many initial parameters and have poor noise resistance;

[0036] 2. The method can improve the matching degree between the point clouds to be registered, reduce the running time of the algorithm, and effectively solve the problem that the existing plane matching algorithm only performs feature matching through the relative geometric relationship of the two planes, resulting in poor matching effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is an algorithm model diagram of a planar primitive-based registration algorithm for structured scenes used in Example 1 of the present invention;

[0038] Figure 2 This is a comparison chart of the cumulative errors of the planar primitive-based registration algorithm for structured scenes used in Example 1 on the Apartment dataset of ASL, where (2a) represents the cumulative angle error and (2b) represents the cumulative offset error.

[0039] Figure 3 The present invention is a planar primitive-based registration algorithm for structured scenes. The accumulated error is used on the ETH dataset of ASL in Example 2, where (3a) represents the accumulated angle error and (3b) represents the accumulated offset error.

[0040] Figure 4The present invention is a planar primitive-based registration algorithm for structured scenes. The accumulated error is used on the Stairs dataset of ASL in Example 3, where (3a) represents the accumulated angle error and (3b) represents the accumulated offset error.

[0041] Figure 5 This is a comparison chart of the improvement in algorithm operation efficiency achieved by the plane matching part based on the angle feature descriptors of three planes used in Example 4 of the present invention for a registration algorithm based on planar primitives for structured scenes. DETAILED DESCRIPTION

[0042] The present invention will be described below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to describe the present invention in more detail, but are not intended to limit the scope of application of the planar primitive-based registration algorithm for structured scenes of the present invention.

[0043] Example 1

[0044] With the rapid development of mapping, positioning and other fields, LiDAR is being used more frequently in structured scenes. Point cloud registration is an essential pre-process in LiDAR 3D point cloud processing technology and is the basis for large-scale 3D scene construction. In order to verify the performance of the present invention in structured scenes, the present invention selected the Apartment dataset from the Autonomous Systems Lab (ASL) data for testing. Figure 1 The basic process of the point cloud registration algorithm based on plane primitives is described, which includes the preprocessing module, plane extraction module, plane matching module, transformation matrix calculation module and optimal selection module. The specific steps are as follows:

[0045] The point cloud registration method based on planar primitives for structured scenes comprises the following steps:

[0046] S1. Use a straight-through filter to filter the outliers in the two frames of point clouds to be registered, and obtain two frames of filtered point cloud data;

[0047] The two frames of point clouds to be registered are respectively recorded as:

[0048] S={s i =(x i ,y i ,z i ),i∈[1,N]},T={t j =(x j ,y j ,z j ),j∈[1,M]};

[0049] Among them, s i , t jRepresent the points in the source and target point clouds respectively; x, y, z are the homogeneous coordinates in their respective spaces; N and M represent the number of points in the source and target point clouds respectively, and N≠M;

[0050] S2, down-sampling the two frames of filtered point cloud data using a voxelized grid to obtain pre-processed point cloud data;

[0051] The size of the voxelized grid is r vo ;

[0052] The pre-processed point cloud data are denoted as S' and T' respectively;

[0053] S3, using the region growing method and the RANSAC algorithm in sequence to extract the plane information of the pre-processed point cloud data to obtain the segmented plane;

[0054] S4, construct a plane feature descriptor Δ using the Euclidean distance from the plane midpoint to the plane centroid, the minimum neighbor distance of each point, and the number of points obtained in S3;

[0055] The plane extraction process selects a total of l planes as the output of the plane extraction module;

[0056] The calculation of Δ is as follows:

[0057]

[0058] in, m represents the number of points on the plane, x i represents the coordinates of a point on the plane, represents the center of mass of the plane; x closest The KD tree search method is used to find the nearest neighbor of each point on the plane;

[0059] A total of l planes are selected as the output of the plane extraction module, which are respectively denoted as s P={ s p i |i∈[1,l]} and t P={ t p j |j∈[1,l]}, where the plane is represented by

[0060] The input of S5, the plane matching module, is the l planes selected by S4. The matching feature descriptors are extracted in the two frame point clouds respectively. Specifically, the angle between the two plane normal vectors in the same frame point cloud is calculated and satisfies formula (2); two planes are taken from different frame point clouds respectively, the angle between the two plane normal vectors in each point cloud is calculated and satisfies formula (3); two planes are taken from different frame point clouds respectively, the Euclidean distance between the two plane centroids in each point cloud is calculated and satisfies formula (4).

[0061]

[0062]

[0063]

[0064] Among them, k, k'=1, 2, 3, and k≠k'. Its function is to prevent the three planes of the same frame point cloud from being parallel or overlapping. The function of is to ensure the accuracy of the corresponding surface angle difference. th The role of is to ensure the accuracy of the position of the corresponding surface relative to the center of mass.

[0065] S6, the plane extraction module proposes an angle feature descriptor based on the centroid and normal vector of the three planes. Specifically, it extracts the centroid of the three sets of planes in the source point cloud and the target point cloud respectively, and constructs a new plane based on this. Calculate the new plane s p ic 、 t p ic The sum of the angles between the normal vectors of the three original planes. Based on the threshold value, a feature descriptor that meets the requirements is obtained:

[0066] θ source =|arccos(- s n i1 · s n ic )+arccos(- s n i2 · s n ic )+arccos(- s n i3 · s n ic )|, (5)

[0067] θ target =|arccos(- t n j1 · t n jc )+arccos(- t n j2 · t n jc )+arccos(- t n j3 · t n jc ), (6)

[0068] θ source -θ target<θ thresold , (7)

[0069] Among them, θ thresold Represents the threshold value of the angle feature. When the feature descriptor of the extracted plane meets the above four conditions, it is used as the output of the plane matching step and recorded as

[0070] S7, the conversion matrix calculation module uses the face-to-face SVD algorithm, which specifically includes the following sub-steps:

[0071] S71: The corresponding planes are respectively denoted as in s n i and s n i Represents the unit normal vectors in the source point cloud and target point cloud planes respectively. The rotation matrix R is solved by minimizing the error function E. 3×3 :

[0072]

[0073] in, s n i and t n i The corresponding surfaces s p i 、 t p i The normal vector of .

[0074] S72: Define the correlation matrix by the normal vectors of the corresponding faces in the source point cloud and the target point cloud:

[0075]

[0076] Decompose H using the SVD algorithm and determine the rotation matrix R through the two unitary matrices 3×3 .

[0077] R=VU T . (10)

[0078] After determining the rotation matrix, calculate the offset matrix t 3×1 .

[0079] S73: Define the error function to solve the offset vector t 3×1 :

[0080]

[0081] t=(A T A) -1 A T b, (12)

[0082] Among them, A and b can be described by formula (13) and formula (14):

[0083]

[0084]

[0085] S74: The rotation matrix and the offset vector are obtained through sub-steps S72 and S73, and are combined to obtain the transformation matrix:

[0086]

[0087] Therefore, through this step calculation, the set of transformation matrices is

[0088] S8, the optimal selection module includes the following sub-steps:

[0089] S81: down-sample the filtered two-frame point cloud data using a voxelized grid to obtain point cloud data;

[0090] Where the size of the voxelized grid is r' vo ;

[0091] The pre-processed point cloud data are denoted as S” and T” respectively;

[0092] S82: Transformation matrix set from the output of S7 Choose a transformation matrix Put S" and follow Get a new point cloud S t ”.

[0093] S83: Calculate point cloud S" t And the number of points in point cloud T' that satisfy formula (16):

[0094]

[0095] Among them, x i represents a point in the target point cloud T', Represents point cloud S' t Zhong and x i For the points with the closest Euclidean distance, if the relationship between them satisfies formula (12), the point cloud overlap degree is increased by one;

[0096] In the simulation process of this example, the parameters are selected as follows: In the preprocessing part, the downsampling rate of the data set is r vo =0.05. In the plane matching part, d th ,θ thresold The values ​​are 1°, 1°, 0.05 meters and 1° respectively. In the optimal selection part, rv ' o =0.1=d cd ×10.

[0097] In this simulation, the registration was considered successful when the angular error of the point cloud registration was less than 2.5° and the offset error was less than 0.1m. To evaluate the performance of this registration algorithm, its rotational and offset errors were compared with those of the commonly used iterative closest point algorithm, the normal distribution variation algorithm (both implemented using the PCL library, hence referred to as PCL-ICP and PCL-NDT), the sample consensus initial alignment algorithm (SAC-IA), the fast global registration algorithm (FGR), and the coherent point drift (CPD) algorithm.

[0098] Apartment data is a flat containing five rooms. Figure 2 As shown, (2a) represents the cumulative angle error and (2b) represents the cumulative offset error. From Figures (2a) and (2b), it can be seen that 98% of the point clouds achieve successful registration with a rotation error of less than 2.5° and an offset error of less than 0.1m, which is better than other algorithms except FGR. At the same time, it can be found from Figures (2a) and (2b) that the present invention is better than the FGR algorithm in specific performance. The average angle error of the present invention is 0.55° and the average offset error is 0.023m, while the average angle error of the FGR algorithm is 0.94° and the average offset error is 0.028cm. In addition, since the overlap rate between the two frames of point clouds in the Apartment dataset is low and there is movement of local objects in the dataset (such as movement of furniture), the registration accuracy of the remaining four algorithms is less than 70%. In summary, the comprehensive performance of the present invention on the Apartment dataset is better than other algorithms.

[0099] Example 2

[0100] In order to verify the performance of the present invention in structured scenarios, the present invention selected the ETH dataset in the ASL data for testing and compared it with the performance of PCL-ICP, PCL-NDT, SAC-IA, and FGR algorithms. The comparison results are shown in the attached figure. Figure 3 As shown, (3a) represents the cumulative angle error and (3b) represents the cumulative offset error. The detailed process of this example test is the same as that shown in Example 1. During the simulation process of this example, the parameters are selected as follows: In the preprocessing part, the downsampling rate of the data set is r vo =0.05. In the plane matching part, d th ,θthresold The values ​​are 1°, 1°, 0.05 meters and 1° respectively. In the optimal selection part, r v ' o =0.1=d cd ×10. The test results are as follows Figure 3 As shown in Figure 2, when the angle error between point clouds is less than 2.5° and the offset error is less than 0.1m, it is considered a successful registration.

[0101] The ETH data is a structured scene containing a corridor. According to Figures (3a) and (3b), the percentage of point clouds with a rotation error of less than 2.5° and an offset error of less than 0.1m in this scene is 97%. Due to the presence of a large number of repeated columns and arch structures in the scene, the feature information is blurred and ambiguous. Therefore, the FGR algorithm has a poor registration effect on the ETH dataset, with an accuracy of 0%; while the registration accuracies of the traditional ICP and NDT algorithms are 91.4% and 82.8%, respectively. The CPD algorithm takes too long to run on the ETH dataset, so it is not compared in this dataset. In summary, the present algorithm is significantly superior to other algorithms in terms of registration accuracy on the ETH dataset.

[0102] Example 3

[0103] To verify the performance of the present invention in structured scenarios, the present invention selected the Stairs dataset in the ASL data for testing and compared the performance with PCL-ICP, PCL-NDT, SAC-IA, FGR, and CPD algorithms. The comparison results are shown in the attached figure. Figure 4 As shown, (4a) represents the cumulative angle error and (4b) represents the cumulative offset error. The detailed process of this example test is the same as that shown in Example 1. During the simulation process of this example, the parameters are selected as follows: In the preprocessing part, the downsampling rate of the data set is r vo =0.05. In the plane matching part, d th ,θ thresold The values ​​are 1°, 1°, 0.05 meters and 1° respectively. In the optimal selection part, r' vo =0.1=d cd ×10. The test results are as follows Figure 4 As shown in Figure 2, when the angle error between point clouds is less than 2.5° and the offset error is less than 0.1m, it is considered a successful registration.

[0104] The Stairs dataset contains both indoor and outdoor structured scenes and is characterized by a high sampling frequency, so the precision registration algorithm performs well on this dataset. As shown in Figures (4a) and (4b), the CPD and ICP algorithms achieve registration accuracies of 93% and 83%, respectively, on this dataset. Meanwhile, our method achieves 100% registration accuracy on this dataset. While the equally accurate CPD algorithm takes an average of over 800 seconds to run, our algorithm only takes 12.9 seconds.

[0105] Example 4

[0106] In order to make the technical advantages of the three-plane angular feature descriptor proposed in the plane matching part of the plane primitive-based registration algorithm for structured scenes more obvious, the following is combined with Example 4 and the attached Figure 5 , the plane extraction part of the present invention is described in more detail:

[0107] As shown in the simulation process in Example 1, the plane matching process is divided into two steps, S5 and S6. S5 solves the feature descriptor based on the centroid and normal vector relationship of two distinct feature planes; S6 constructs a new plane based on the centroid of three distinct feature planes and constructs a new angle feature descriptor based on the angle between this new plane and the normal vectors of the original three distinct feature planes. This descriptor improves the accuracy of matching planes between the point clouds to be registered and significantly reduces the algorithm's running time. To verify the beneficial effects of the proposed angle feature descriptor, a control experiment was conducted using the following method: Two sets of control simulation experiments were conducted in this example. One set of experimental steps was the same as in Example 1; the other set of experiments did not perform step S6 of Example 1, but remained the same. Tests were conducted on the Apartment, ETH, and Stairs datasets of ASL data.

[0108] This example uses the ASL dataset to analyze the role of the new angle feature descriptor, as shown in the attached Figure 5 As shown, Figure (5a) shows the running time required by the two comparison algorithms on the Apartment dataset; Figure (5b) shows the running time required by the two comparison algorithms on the ETH dataset; and Figure (5c) shows the running time required by the two comparison algorithms on the Stairs dataset. As shown in Figures (5a), (5b), and (5c), the average running time of the Apartment, ETH, and Stairs datasets is reduced by 4.22s, 7.40s, and 1.84s, respectively; and, specifically, time optimization is achieved to varying degrees for each data point. In summary, the feature descriptor proposed in the present invention achieves the purpose of improving the overall model operation efficiency, which is in line with the expectations of the present invention.

[0109] In summary, the above are only preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention.

Claims

1. A point cloud registration method based on planar primitives for structured scenes, characterized by: include: S1. Use a straight-through filter to filter the outliers in the two frames of point clouds to be registered, and obtain two frames of filtered point cloud data; S2, down-sampling the two frames of filtered point cloud data using a voxelized grid to obtain pre-processed point cloud data; S3, using the region growing method and the RANSAC algorithm in sequence to extract the plane information of the pre-processed point cloud data to obtain the segmented plane; S4, constructing a plane feature descriptor by multiplying the Euclidean distance from the point in the segmented plane to the plane centroid by the number of points and then dividing it by the minimum nearest neighbor distance of each point, and extracting the plane with obvious features of the source point cloud and the target point cloud based on it; S5, selecting three planes from the planes with obvious features in the source point cloud and the target point cloud respectively so that the angle between the normal vectors of the two planes in the same frame point cloud and the source point cloud and the target point cloud is greater than a set threshold and the Euclidean distance between the centroids of the two planes in each point cloud is less than a set threshold, and performing geometric matching based on the above; S6, first extracting the centroids of the three sets of planes based on the source point cloud and the target point cloud respectively and constructing a new plane based on the centroids, and then calculating the sum of the angles between the new plane and the normal vectors of the original three planes; The difference between the sum of the normal angles of the new plane calculated by the source point cloud and the target point cloud and the original three planes is selected, and three sets of planes from the source point cloud and the target point cloud that are less than the set threshold are selected to form a set of successfully matched planes. S7, calculating the transformation matrix of the plane features in the successfully matched plane set to form a transformation matrix set; S8. Select the transformation matrix with the highest point cloud overlap for point cloud registration and quantification.

2. The point cloud registration method according to claim 1, characterized in that: The filtering specifically includes: filtering the outliers in the source point cloud and the target point cloud to be registered to obtain two frames of filtered point cloud data.

3. The point cloud registration method according to claim 1 or 2, characterized in that: The downsampling is to downsample the two frames of filtered point cloud data to obtain preprocessed point cloud data; the segmentation is specifically to segment the plane information of the preprocessed point cloud data using the region growing method and the RANSAC algorithm in sequence to obtain the segmented plane.

4. The point cloud registration method according to claim 2, characterized in that: The filtering process is filtering using a straight-through filter.

5. The point cloud registration method according to claim 3, characterized in that: The downsampling downsamples the source point cloud and the target point cloud respectively through a voxelized grid.

6. The point cloud registration method according to claim 1, characterized in that: The extraction of planes with obvious features from the source point cloud and the target point cloud in S4 is achieved by sorting the plane feature descriptors and selecting the l planes with the largest plane feature descriptors.

7. The point cloud registration method according to claim 6, characterized in that: The value range of l is 8 to 12.

8. The point cloud registration method according to claim 1, wherein: The geometric matching described in S5 is specifically as follows: if the angle difference between the geometric feature descriptor and the angular feature descriptor is less than 1 degree and the offset difference is less than 0.1 meter, the matching is successful; otherwise, the matching fails.

9. The point cloud registration method according to claim 1, characterized in that: S7 calculates the transformation matrix of the plane features in the successfully matched plane set using the face-to-face SVD algorithm.

10. The point cloud registration method according to claim 1, characterized in that: S8 includes the following sub-steps: S81: Use voxelized grids to downsample the source point cloud and the target point cloud respectively to obtain downsampled source point cloud and target point cloud; S82: Select a transformation matrix from the transformation matrix set obtained in S7, and transform the downsampled source point cloud into a new point cloud according to the transformation matrix; S83: Calculate the minimum distance between each point in the new point cloud and the point in the sampled target point cloud. If the distance is less than or equal to 1 cm, the point cloud coincidence is increased by 1. Otherwise, the closest point is discarded to obtain the point cloud coincidence corresponding to the transformation matrix. S84: Jump to S82 until the transformation matrices in the transformation matrix set of S7 are traversed and the transformation matrix with the largest point cloud overlap is selected as the optimal transformation matrix.

Images