Power corridor laser point cloud registration method and system, storage medium and equipment

By voxel division and feature extraction of point cloud data in power corridor scenes, combined with the analysis of covariance matrix and change matrix, the problem of difficult point cloud registration efficiency and accuracy in the existing technology is solved, and more efficient and Accurate point cloud registration.

CN119027466APending Publication Date: 2024-11-26YUNNAN POWER GRID CO LTD ELECTRIC POWER RES INST

Patent Information

Application Number
CN202410987140.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-23
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing point cloud registration technology is difficult to take into account both accuracy and efficiency when dealing with power corridor scenarios, especially in cases where density distribution is uneven, occlusion or severe noise, and deep learning-based methods ignore the non-point cloud. The effect of overlapping areas on registration accuracy.

Method used

Through the grid, point cloud data is divided in voxel form, point cloud collections of different voxels are obtained, rotation matrix and number of point clouds in voxels are analyzed to obtain local density features, point-by-point features are spliced ​​to identify overlapping areas, and covariance is used to analyze the rotation matrix and the number of point clouds in the voxels to obtain local density features. The matrix and change matrix are initially and quadraticly registered to optimize the position of the point cloud.

Benefits of technology

It improves the efficiency and accuracy of point cloud registration, accurately identify overlapping areas, reduces the amount of data participating in registration, and improves the overall performance of point cloud registration in power corridor scenes.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119027466A_ABST
    Figure CN119027466A_ABST
Patent Text Reader

Abstract

The embodiment of the invention discloses an electric power corridor laser point cloud registration method and system, a storage medium and equipment. The method comprises the following steps: acquiring point cloud data of an electric power corridor scene; dividing the point cloud data in a voxel form through a grid to obtain a point cloud set of different voxels; acquiring density characteristics according to the rotation matrix of the point clouds in the point cloud set of the current voxel and the number of the point clouds; splicing the density feature and the point-by-point feature of the point cloud data to obtain region masks of a source point cloud and a target point cloud in the point cloud data; if the area mask is determined to be a preset mask value, the point cloud is the overlapping area point cloud; according to the covariance matrix of the source point cloud and the target point cloud, determining a main shaft of the two groups of point clouds, and performing initial registration on the source point cloud based on the main shaft; according to the corresponding relation between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, obtaining a transformation matrix, carrying out the secondary registration of the source point cloud based on the transformation matrix, obtaining an optimal source point cloud position, and improving the efficiency and precision of the point cloud registration work.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of laser point cloud technology, and in particular to a method, system, storage medium and device for laser point cloud registration of a power corridor. Background Art

[0002] Laser point cloud refers to a data set of spatial points obtained by scanning with a three-dimensional laser radar device. Drones equipped with laser radars scan the ground and convert the image data of line towers, ground wires, line channels and their surrounding environment into three-dimensional point cloud data through spatial three-dimensional calculation, so that the inspection line becomes an accurate three-dimensional model that can be measured and analyzed arbitrarily. With this model, power transmission and inspection personnel can timely and accurately detect abnormalities and hidden dangers in the transmission line, providing more stable and reliable power guarantee for people's production and life. Point cloud registration is the process of finding the mapping relationship between different point clouds under different perspectives, and using certain algorithms to convert different point clouds of the same target scene into the same coordinate system to form a more complete point cloud. The point cloud registration process is generally divided into two steps: coarse registration and fine registration. 3D point cloud registration is an important part of laser point cloud processing technology.

[0003] Existing point cloud registration technology solutions can be classified into two categories: one is the traditional point cloud registration algorithm, and the other is the point cloud registration algorithm based on deep learning. Point cloud registration algorithms based on traditional methods often cannot take into account both accuracy and efficiency. For point cloud registration algorithms based on deep learning, in point cloud registration based on feature descriptor learning, point-based methods ignore the local geometric information of point clouds, voxel-based methods are difficult to balance the problem between resolution and fine features, and graph-based methods are less efficient in large scene point clouds; point cloud registration methods based on correspondence solution perform poorly in cases of uneven density distribution, large changes in overlapping areas, occlusion or severe noise; point cloud registration methods based on posture regression mostly focus on the registration of completely overlapping small target point clouds, and use an iterative approach to achieve higher accuracy; methods based on scene flow estimation largely ignore the problem of motion consistency within each target, and are sensitive to occlusion and noise. In addition, point cloud registration algorithms based on deep learning often ignore the impact of non-overlapping areas of point clouds on registration accuracy, and do not pay attention to the inherent characteristics of objects in specific scenes during feature extraction. Summary of the invention

[0004] Based on this, it is necessary to propose a power corridor laser point cloud registration method, system, storage medium and equipment to address the above problems.

[0005] A method for laser point cloud registration of a power corridor, the method comprising:

[0006] Get point cloud data of the power corridor scene.

[0007] The point cloud data is divided into voxels through a grid to obtain point cloud sets of different voxels.

[0008] The density characteristics of the local neighborhood of the point cloud are obtained according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel.

[0009] The density features of the local neighborhood of the point cloud are spliced ​​with the point-by-point features of the point cloud data to obtain regional masks of the source point cloud and the target point cloud in the point cloud data.

[0010] It is determined that the region masks of the source point cloud and the target point cloud in the point cloud data are preset mask values, and the point cloud is an overlapping region point cloud.

[0011] According to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, the principal axes of the two groups of point clouds are determined, and the source point cloud is initially aligned based on the principal axes.

[0012] According to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, the change matrix of the source point cloud and the target point cloud is obtained, and the source point cloud is re-registered based on the transformation matrix to obtain the optimal source point cloud position.

[0013] The step of dividing the point cloud data into voxels by grid to obtain point cloud sets of different voxels specifically includes:

[0014] The point cloud data is preprocessed to filter out outliers and noise points in the point cloud data.

[0015] The maximum and minimum values ​​of each point cloud coordinate in the initialized point cloud data are obtained.

[0016] The grid size is determined according to the maximum and minimum values ​​of the point cloud coordinates.

[0017] The voxel coordinates to which each point cloud coordinate belongs are determined according to the point cloud coordinates in the point cloud data.

[0018] Each point cloud coordinate is assigned to a corresponding voxel using the voxel coordinates, and the number of point clouds in each voxel is counted and saved.

[0019] The point cloud sets of different voxels are determined according to the number of voxels in each direction in the three-dimensional coordinate system and the number of point clouds in each voxel.

[0020] The step of obtaining the density feature of the local neighborhood of the point cloud according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel specifically includes:

[0021] The point cloud feature vector of the point cloud in the point cloud set of the current voxel is multiplied by the feature vector matrix to normalize the point cloud data.

[0022] Extract the normalized high-dimensional feature vector of the point cloud data.

[0023] A rotation matrix corresponding to the high-dimensional feature vector is obtained, and the high-dimensional feature vector is aligned by multiplying the rotation matrix with the point cloud feature vector of the point cloud in the point cloud set of the current voxel.

[0024] The aligned high-dimensional feature vectors are aggregated through a maximum pooling operation to obtain the global features of the current point cloud set, and the global features are compressed to output the point cloud local information within the point cloud set of the current voxel.

[0025] According to the position of the grid where the point cloud set of the current voxel is located, the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in adjacent directions are obtained.

[0026] The number of point clouds of the current voxel and the number of point clouds of its adjacent voxels are concatenated in spatial order to obtain the feature vectors between the voxel grids.

[0027] The compressed local information of the point cloud is concatenated with the feature vectors between the voxel grids to obtain the density features of the local neighborhood of the point cloud.

[0028] The step of obtaining the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in adjacent directions according to the position of the grid where the point cloud set of the current voxel is located specifically includes:

[0029] It is determined that the point cloud set of the current voxel is in the edge grid, then for the missing adjacent directions, the number of point clouds in the point cloud set is marked by a preset marking value, and the number of point clouds in the point cloud sets of all voxels in the adjacent directions that are not missing is obtained.

[0030] If it is determined that the point cloud set of the current voxel is not in the edge grid, the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in adjacent directions are obtained.

[0031] The density feature of the local neighborhood of the point cloud is spliced ​​with the point-by-point feature of the point cloud data to obtain the regional mask of the source point cloud and the target point cloud in the point cloud data, specifically including:

[0032] The point-by-point features of the source point cloud and the target point cloud in the preprocessed point cloud data are respectively extracted.

[0033] Average pooling and maximum pooling operations are performed on the point-by-point features to obtain global features, and shared global features are extracted from the global features.

[0034] The density features of the local neighborhood of the source point cloud and the point-by-point features are concatenated with the shared global features of the target point cloud to obtain a region mask of the source point cloud.

[0035] The density features of the local neighborhood of the target point cloud and the point-by-point features are concatenated with the shared global features of the source point cloud to obtain a region mask of the target point cloud.

[0036] Wherein, after the density feature of the local neighborhood of the point cloud is spliced ​​with the point-by-point feature of the point cloud data to obtain the regional mask of the source point cloud and the target point cloud in the point cloud data, it further specifically includes:

[0037] A source point cloud loss function is defined according to the true mask and the region mask of the source point cloud.

[0038] A target point cloud loss function is defined according to the true mask and the region mask of the target point cloud.

[0039] The total loss function is determined by the source point cloud loss function and the target point cloud loss function.

[0040] The region mask is optimized according to a total loss function.

[0041] The step of determining the principal axes of the two groups of point clouds according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and performing initial registration on the source point cloud based on the principal axes specifically includes:

[0042] Determine the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud.

[0043] The eigenvalues ​​and eigenvectors of the covariance matrix are obtained by singular value decomposition, and the principal axes of the two groups of point clouds are determined.

[0044] An initial rigid body transformation matrix is ​​determined according to the principal axes of the two groups of point clouds, and the source point cloud is transformed by rotation to obtain a point cloud aligned with the principal axes.

[0045] The centers of the source point cloud and the target point cloud are determined respectively.

[0046] According to the point cloud centroid of the source point cloud and the point cloud centroid of the target point cloud, combined with the centers of the source point cloud and the target point cloud, the direction of the line connecting the point cloud centroid and the center is determined, and the initial alignment of the source point cloud is performed through the direction of the line connecting the point cloud centroid and the center.

[0047] The step of determining an initial rigid body transformation matrix according to the principal axes of the two groups of point clouds and obtaining a point cloud aligned with the principal axes by rotationally transforming the source point cloud specifically includes:

[0048] according to Determine the initial rigid body transformation matrix, where R0 and T0 are the rotation matrix and translation vector respectively. is the point cloud centroid of the source point cloud, is the point cloud centroid of the target point cloud, Up is the principal axis of the source point cloud, and Uq is the principal axis of the target point cloud. is the transposed matrix of the principal axis of the source point cloud, is the transposed matrix of the principal axes of the target point cloud.

[0049] according to

[0050] P′=P×R0+T0

[0051] The source point cloud is transformed by rotation to obtain a point cloud aligned with the principal axis, wherein P is the source point cloud.

[0052] The step of obtaining the change matrix of the source point cloud and the target point cloud according to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, performing secondary registration on the source point cloud based on the transformation matrix, and obtaining the optimal source point cloud position specifically includes:

[0053] The source point cloud is sampled to obtain a subset of the source point cloud.

[0054] The closest point to each point in the source point cloud subset is found in the target point cloud initially aligned with the source point cloud to form an initial correspondence between the two point clouds.

[0055] The transformation matrix of the two point clouds is calculated according to the initial corresponding relationship to minimize the registration error and obtain a transformed point cloud subset.

[0056] A new source point cloud position is obtained according to the transformed point cloud subset, and the iteration is terminated according to the comparison between the registration error and the convergence condition or the preset number of iterations to obtain the optimal position of the source point cloud.

[0057] The step of obtaining a new source point cloud position according to the transformed point cloud subset until the registration error reaches a convergence condition or a preset number of iterations, then the iteration is terminated to obtain the optimal position of the source point cloud, specifically including:

[0058] according to Determine the registration error, where n is the number of iterations, Qi is the i-th target point cloud, and P i is the i-th source point cloud.

[0059] If it is determined that the registration error is less than a convergence condition or reaches a preset number of iterations, the iteration is terminated and the optimal point cloud subset is obtained.

[0060] If it is determined that the registration error is greater than the convergence condition or the preset number of iterations is not reached, then the iteration of the transformed point cloud subset continues.

[0061] A power corridor laser point cloud registration system, characterized in that the system comprises:

[0062] Point cloud data acquisition module, used to acquire point cloud data of the power corridor;

[0063] A point cloud set acquisition module of different voxels, used for dividing the point cloud data into voxels through a grid to acquire point cloud sets of different voxels;

[0064] A density feature acquisition module is used to acquire the density feature of the local neighborhood of the point cloud according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel;

[0065] A region mask acquisition module, used for splicing the density features of the local neighborhood of the point cloud with the point-by-point features of the point cloud data to obtain the region masks of the source point cloud and the target point cloud in the point cloud data;

[0066] An overlapping area point cloud determination module is used to determine that if the area mask of the original point cloud and the target point cloud in the point cloud data is a preset mask value, then the point cloud is the overlapping area point cloud;

[0067] An initial registration module, used to determine the principal axes of the two groups of point clouds according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and perform initial registration on the source point cloud based on the principal axes;

[0068] The optimal source point cloud position acquisition module is used to obtain the change matrix of the source point cloud and the target point cloud according to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, and perform secondary registration on the source point cloud based on the transformation matrix to obtain the optimal source point cloud position.

[0069] A computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the processor executes the steps of the method described above.

[0070] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the above method.

[0071] The embodiments of the present invention have the following beneficial effects:

[0072] The present invention divides the point cloud data of the power corridor scene in the form of voxels through a grid, and divides the point cloud data into smaller, easy-to-manage voxels (i.e., small cubes in three-dimensional space). Each voxel contains a certain amount of point cloud data, and a point cloud set of different voxels is obtained. This division helps to reduce the amount of calculation for subsequent processing and improve processing efficiency. At the same time, it is also convenient to perform local analysis and feature extraction on the point cloud data. Further, by analyzing the rotation matrix and number of point clouds in the voxel, the density features of the local area of ​​the point cloud are obtained, and the density features of the local neighborhood of the point cloud are spliced ​​with the point-by-point features of the point cloud data to obtain the regional mask of the source point cloud and the target point cloud in the point cloud data. By comparing the regional mask with the preset mask value, the overlapping area between the source point cloud and the target point cloud can be accurately identified, and the interference of the changing objects and the non-overlapping area can be eliminated. The registration problem of the point cloud with partial area overlap is converted into the registration problem of the point cloud with almost completely overlapped areas, reducing the amount of data involved in the registration, and the point clouds in the overlapping area have a high degree of similarity, which is conducive to finding a good initial value for the subsequent point cloud registration work. Furthermore, based on the initial registration, by more carefully analyzing the correspondence between the source point cloud and the target point cloud in the overlapping area point cloud, a more accurate transformation matrix can be calculated. The source point cloud can be re-registered using the transformation matrix to obtain the optimal position of the source point cloud in the target point cloud coordinate system, thereby improving the efficiency and accuracy of the point cloud registration work. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0074] in:

[0075] Figure 1 A schematic diagram of a flow chart of an embodiment of a method for laser point cloud registration of a power corridor provided by the present invention;

[0076] Figure 2 A schematic structural diagram of an embodiment of a density embedding module provided by the present invention;

[0077] Figure 3 A schematic diagram of the structure of an overlapping area detection network structure according to an embodiment of the present invention;

[0078] Figure 4 A schematic diagram of a flow chart of another embodiment of a method for laser point cloud registration of a power corridor provided by the present invention;

[0079] Figure 5A schematic diagram of the PointNet layer structure provided by the present invention;

[0080] Figure 6 A schematic diagram of a flow chart of another embodiment of a method for laser point cloud registration of a power corridor provided by the present invention;

[0081] Figure 7 A schematic structural diagram of an embodiment of a power corridor laser point cloud registration system provided by the present invention;

[0082] Figure 8 A schematic diagram of the structure of an embodiment of a computer device provided by the present invention;

[0083] Fig. 9 A schematic structural diagram of an embodiment of a storage medium provided by the present invention. DETAILED DESCRIPTION

[0084] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0085] like Figure 1 As shown, Figure 1 A schematic diagram of a flow chart of an embodiment of a method for laser point cloud registration of a power corridor provided by the present invention. A method for laser point cloud registration of a power corridor, the method comprising:

[0086] S101: Acquire point cloud data of the power corridor scene.

[0087] Exemplarily, point cloud data of a power corridor scene from multiple different perspectives are acquired through a laser scanning device, and data preprocessing is performed on the point cloud data. The data preprocessing mainly uses filtering algorithms (such as statistical filtering, Gaussian filtering, etc.) to filter out outliers and noise points in the acquired point cloud data.

[0088] S102: Divide the point cloud data into voxels through a grid to obtain point cloud sets of different voxels.

[0089] For example, referring to Figure 2 , Figure 2The schematic diagram of the structure of an embodiment of the density embedding module provided by the present invention. The point cloud of the power scene has obvious density characteristics. Vegetation points and surface points account for the vast majority of the total number of point clouds and their density distribution is uneven, while the point clouds of power towers and power lines have a relatively uniform density distribution in the vertical or horizontal direction. Therefore, the present invention proposes a density embedding module, through which the density change information of points and fields is extracted. First, the input point cloud data is voxelized through the voxel layer, and the index relationship between each point cloud and the voxel is obtained. Specifically, the maximum and minimum values ​​of each point cloud coordinate in the initialized point cloud data are obtained; the grid size is determined according to the maximum and minimum values ​​of the point cloud coordinates; the voxel coordinates to which each point cloud coordinate belongs are determined according to the point cloud data; each point cloud coordinate is assigned to the corresponding voxel using the voxel coordinates, and the number of point clouds in each voxel is counted and saved; the point cloud set of different voxels is determined according to the number of voxels in each direction in the three-dimensional coordinate system and the number of point clouds in each voxel.

[0090] S103: Obtain density features of a local neighborhood of the point cloud according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel.

[0091] For example, see Figure 2 The density embedding module mainly includes a PointNet layer for extracting local features of points within voxels and a density feature extraction layer. The point cloud set divided in the form of voxels extracts local information of the point cloud within a single voxel through the PointNet layer. Specifically, the point cloud feature vector of the point cloud within the point cloud set of the current voxel is multiplied by the feature vector matrix to normalize the point cloud data, extract the high-dimensional feature vector of the normalized point cloud data, obtain the rotation matrix corresponding to the high-dimensional feature vector, align the high-dimensional feature vector by multiplying the rotation matrix with the point cloud feature vector of the point cloud within the point cloud set of the current voxel, aggregate the aligned high-dimensional feature vector through the maximum pooling operation, obtain the global features of the current point cloud set, compress the global features and output the local information of the point cloud within the point cloud set of the current voxel.

[0092] At the same time, the density change information between voxels is obtained through the density feature extraction layer. Specifically, according to the position of the grid where the point cloud set of the current voxel is located, the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in the adjacent direction are obtained, and the number of point clouds of the current voxel and the number of point clouds of its adjacent voxels are spliced ​​in spatial order to obtain the feature vector between the voxel grids.

[0093] Furthermore, the compressed local information of the point cloud is concatenated with the feature vector between the voxel grids and passed into the MLP for further feature extraction, thereby obtaining a 1024-dimensional feature vector of the density feature of the local neighborhood of the point cloud.

[0094] S104: The density features of the local neighborhood of the point cloud are concatenated with the point-by-point features of the point cloud data to obtain the regional masks of the source point cloud and the target point cloud in the point cloud data.

[0095] For example, referring to Figure 3 , Figure 3 A schematic diagram of the structure of an embodiment of the overlapping region detection network structure provided by the present invention. The overlapping region detection network includes a point feature extraction module, a density embedding module, a mask estimation module, and a global shared feature module. First, the source point cloud S and the target point cloud T are input into the overlapping region detection network, and the point feature extraction modules are used to extract the point-by-point features of the preprocessed point cloud data, respectively. The point-by-point features are averaged pooled (avg) and max pooled (max) are performed on the point-by-point features to obtain global features, and then the extracted global features are passed to the shared global feature module to obtain shared global features.

[0096] Furthermore, the density features of the local neighborhood of the source point cloud S extracted by the density embedding module are concatenated with the point-by-point features of the source point cloud S and the shared global features of the copied target point cloud T, and passed into the mask estimation module to learn the regional mask of S. The same operation is performed on the target point cloud T to obtain the regional mask of T.

[0097] Among them, the mask estimation module is composed of a group of multi-layer perceptrons (MLPs) of size (512, 256, 128, 1) connected in series. The sigmoid activation function is used in the last layer of MLP to make the final output probability range within [0,1].

[0098] It should be noted that the loss function of the overlapping region detection network consists of two parts, one for evaluating the loss between the output category of each point in the source point cloud and the true category, and the other for evaluating the loss between the output category of each point in the target point cloud and the true category. For each part, the loss during training is defined as the mean square error between the output mask and the true mask. The source point cloud loss function is defined based on the true mask and region mask of the source point cloud, as shown in the following formula:

[0099]

[0100] in, They represent the mask value (i.e., internal point probability) of each point in the source point cloud X and the target point cloud Y through the overlapping area detection network.

[0101] The target point cloud loss function is defined according to the true mask and region mask of the target point cloud, as shown in the following formula:

[0102]

[0103] in, Represent the true mask value of each point in the source point cloud and the target point cloud respectively.

[0104] The total loss function is determined by the source point cloud loss function and the target point cloud loss function, and the region mask is optimized according to the total loss function. The total loss function is shown in the following formula:

[0105] loss=loss x +loss y .

[0106] S105: Determine that the region masks of the original point cloud and the target point cloud in the point cloud data are preset mask values, and the point cloud is an overlapping region point cloud.

[0107] Exemplarily, the preset mask value is 1, and it is determined whether the mask corresponding to each point cloud is 1. All points with a mask of 1 in the two point clouds are points in the overlapping area, and then a filtering algorithm is used to remove isolated points.

[0108] S106: Determine the principal axes of the two groups of point clouds according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and perform initial registration on the source point cloud based on the principal axes.

[0109] For example, given two point clouds, the source point cloud P = {p i ∈R 3 |i=1,2…,N} and target point cloud Q={q j ∈R 3 |j=1,2…,M}, determine the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud. The covariance matrix is ​​shown as follows:

[0110]

[0111] in, is the point cloud centroid of the source point cloud, is the point cloud centroid of the target point cloud, N is the number of source point clouds, M is the number of target point clouds, and the point cloud centroid is used to decentralize the data.

[0112] Furthermore, the eigenvalues ​​and eigenvectors of the covariance matrix are obtained by singular value decomposition to determine the principal axes of the two groups of point clouds. The eigenvalues ​​and eigenvectors of the covariance matrix are shown in the following formula:

[0113]

[0114] Among them, Up is the main axis of the source point cloud, Uq is the main axis of the target point cloud, is the transposed matrix of the principal axis of the source point cloud, is the transposed matrix of the principal axis of the target point cloud, Dp is the covariance matrix of the source point cloud, and Dq is the covariance matrix of the target point cloud.

[0115] Furthermore, an initial rigid body transformation matrix is ​​determined according to the principal axes of the two groups of point clouds, and the source point cloud is transformed by rotation to obtain a point cloud aligned with the principal axis.

[0116] The initial rigid body transformation matrix is ​​shown as follows:

[0117]

[0118] Among them, R0, T0 are rotation matrices and translation vectors respectively, is the point cloud centroid of the source point cloud, is the point cloud centroid of the target point cloud, Up is the principal axis of the source point cloud, and Uq is the principal axis of the target point cloud. is the transposed matrix of the principal axis of the source point cloud, is the transposed matrix of the principal axis of the target point cloud;

[0119] The point cloud aligned with the principal axis is determined according to the following formula:

[0120] P′=P×R0+T0;

[0121] Among them, P is the source point cloud.

[0122] Furthermore, for two similar point clouds, although the main axis is aligned, there is a possibility that the main axis is reversed, so the opposite main axis situation needs to be considered. Specifically, first determine the center of the source point cloud and the target point cloud respectively, as shown in the following formula:

[0123]

[0124] Among them, L max , L min It is the maximum point coordinate and the minimum point coordinate in two point clouds (source point cloud and target point cloud).

[0125] Furthermore, according to the point cloud centroid of the source point cloud and the point cloud centroid of the target point cloud Combined with the centers of the source point cloud and the target point cloud, the direction of the line connecting the point cloud centroid and the center is determined, and the initial registration of the source point cloud is performed through the direction of the line connecting the point cloud centroid and the center.

[0126] S107: According to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, the change matrix of the source point cloud and the target point cloud is obtained, and the source point cloud is re-registered based on the transformation matrix to obtain the optimal source point cloud position.

[0127] For example, the source point cloud is sampled to obtain a subset P0 of the source point cloud, P0∈P, P0 represents a subset of the source point cloud P, and the nearest point min to each point in the source point cloud subset P0 is found in the target point cloud Q initially aligned with the source point cloud, that is, P0qi -p i P0=min, forming the initial correspondence between the two point clouds, where q i is the target point cloud, p i is the source point cloud, and the RANSAC algorithm is used to remove incorrect matching point pairs.

[0128] Furthermore, the transformation matrix of the two point clouds is calculated according to the initial corresponding relationship to minimize the registration error and obtain a transformed point cloud subset. The new source point cloud position is obtained according to the transformed point cloud subset. The iteration is terminated according to the comparison between the registration error and the convergence condition or the preset number of iterations to obtain the optimal source point cloud position.

[0129] It should be noted that in the secondary registration stage of the point cloud, if the amount of point cloud data is too large, resulting in slow registration, the two point clouds can be uniformly sampled in the same way to further reduce the amount of data.

[0130] From the above description, it can be seen that the present invention divides the point cloud data of the power corridor scene in the form of voxels through a grid, and divides the point cloud data into smaller, easy-to-manage voxels (i.e., small cubes in three-dimensional space). Each voxel contains a certain amount of point cloud data, and a point cloud set of different voxels is obtained. This division helps to reduce the amount of calculation for subsequent processing and improve processing efficiency. At the same time, it is also convenient to perform local analysis and feature extraction on the point cloud data. Furthermore, by analyzing the rotation matrix and number of point clouds in the voxel, the density features of the local area of ​​the point cloud are obtained, and the density features of the local neighborhood of the point cloud are spliced ​​with the point-by-point features of the point cloud data to obtain the regional masks of the source point cloud and the target point cloud in the point cloud data. By comparing the regional mask with the preset mask value, the overlapping area between the source point cloud and the target point cloud can be accurately identified, eliminating the interference of the changing objects and the non-overlapping area, and converting the registration problem of the partially overlapping point cloud into the registration problem of the almost completely overlapping point cloud, reducing the amount of data involved in the registration, and the point clouds in the overlapping area have a high degree of similarity, which is conducive to finding a good initial value for the subsequent point cloud registration work. Furthermore, on the basis of the initial registration, by more finely analyzing the correspondence between the source point cloud and the target point cloud in the overlapping area point cloud, a more accurate transformation matrix can be calculated, and the transformation matrix is ​​used to perform secondary registration on the source point cloud to obtain the optimal position of the source point cloud in the target point cloud coordinate system, thereby improving the efficiency and accuracy of the point cloud registration work.

[0131] like Figure 4 As shown, Figure 4 A schematic diagram of another embodiment of a method for laser point cloud registration of a power corridor provided by the present invention. A method for laser point cloud registration of a power corridor, the method comprising:

[0132] S201: Acquire point cloud data of the power corridor scene.

[0133] S202: Preprocess the point cloud data to filter out outliers and noise points in the point cloud data.

[0134] It should be noted that steps S201-S202 are Figure 1 This has been discussed in detail in the implementation scenario shown and will not be repeated here.

[0135] S203: Obtain the maximum and minimum values ​​of each point cloud coordinate in the initialized point cloud data.

[0136] For example, the overlapping area detection network first voxelizes the input point cloud data through the voxel layer, obtains the index relationship between each point cloud and the voxel, and voxelizes the point cloud data using a grid of size 5m×5m×5m. First, the maximum value P of the point cloud data coordinates is counted. max (x max ,y max ,z max ) and the minimum value P min (x min ,y min ,z min ).

[0137] S204: Determine the size of the voxel grid according to the maximum and minimum values ​​of the point cloud coordinates.

[0138] Exemplarily, the size of the voxel grid is then calculated. x ,grid y ,grid z ), the calculation formula is as follows:

[0139]

[0140] Among them, grid i is the voxel grid size.

[0141] S205: Determine the voxel coordinates to which each point cloud coordinate belongs according to the point cloud data.

[0142] For example, the voxel coordinates to which the point cloud Pi (xi, yi, zi) belongs in the point cloud data Determined according to the following formula:

[0143]

[0144] in, is rounded up, m is the three-dimensional coordinate axis, including the x-axis, y-axis, and z-axis, m min It is the minimum value of the point cloud coordinates of the current point cloud.

[0145] S206: using voxel coordinates to assign each point cloud coordinate to a corresponding voxel, and counting and saving the number of point clouds in each voxel.

[0146] For example, using the calculated voxel coordinates Index and add the point cloud Pi(xi,yi,zi) to the corresponding voxel. In the process of allocating point clouds, the number of point clouds in each voxel is counted and stored.

[0147] S207: Determine point cloud sets of different voxels according to the number of voxels in each direction in the three-dimensional coordinate system and the number of point clouds in each voxel.

[0148] Exemplarily, the representation of the point cloud set of different voxels is (B, L, W, H, T, 3), where T is the number of point clouds in each voxel, B, L, W are the dimensions of the voxel grid, that is, the number of voxels in the x, y, and z axis directions respectively, and H is.

[0149] S208: Normalize the point cloud data by multiplying the point cloud feature vector of the point cloud in the point cloud set of the current voxel with the feature vector matrix.

[0150] For example, referring to Figure 5 , Figure 5 The schematic diagram of the PointNet layer structure provided by the present invention. The point cloud set of different voxels is input into the PointNet layer in the density embedding layer. The PointNet layer first learns the feature vector rotation matrix through a T-Net network, and then multiplies the rotation matrix with the point cloud feature vector of the point cloud in the point cloud set of the current voxel input to realize the normalization of the point cloud data.

[0151] S209: extracting a high-dimensional feature vector of the normalized point cloud data.

[0152] Exemplarily, the normalized point cloud data is passed into an MLP (Multilayer Perceptron) to extract a high-dimensional feature vector.

[0153] S210: Obtain a rotation matrix corresponding to the high-dimensional feature vector, and align the high-dimensional feature vector by multiplying the rotation matrix with the point cloud feature vector of the point cloud in the point cloud set of the current voxel.

[0154] Exemplarily, the rotation matrix corresponding to the high-dimensional feature vector is learned through the second T-Net network, and the high-dimensional feature vector is aligned by multiplying the rotation matrix with the point cloud feature vector of the point cloud in the point cloud set of the current voxel.

[0155] S211: The aligned high-dimensional feature vectors are aggregated through a maximum pooling operation to obtain the global features of the current point cloud set, and the global features are compressed to output the local point cloud information within the point cloud set of the current voxel.

[0156] Exemplarily, the aligned high-dimensional feature vector is passed to the second MLP and the point cloud features are aggregated using maximum pooling (MaxPooling) to obtain the global features of the point cloud set, and finally passed to the third MLP to compress and refine the global features of the point cloud set and output the local information of the point cloud within the point cloud set.

[0157] S212: According to the position of the grid where the point cloud set of the current voxel is located, the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in adjacent directions are obtained.

[0158] Exemplarily, in the process of density feature extraction, since the point cloud contained in each voxel grid has been determined at the voxel layer, and the size of each voxel is the same, the point cloud density in the voxel is linearly related to the number of point clouds, so it is only necessary to obtain the relationship between the number of point clouds of each voxel grid and its six adjacent voxel grids. If it is determined that the point cloud set of the current voxel is in the edge grid and there are no adjacent grids in some directions, then for the missing adjacent directions, the number of point clouds in the point cloud set is recorded as -1, and the number of point clouds of the point cloud set of all voxels in the adjacent directions that are not missing is obtained; if it is determined that the point cloud set of the current voxel is not in the edge grid, the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud set of all voxels in the adjacent directions are obtained.

[0159] S213: Concatenate the number of point clouds of the current voxel and the number of point clouds of its adjacent voxels in spatial order to obtain feature vectors between voxel grids.

[0160] Exemplarily, the current voxel and its adjacent grid point cloud numbers are finally concatenated in the order of top, bottom, left, right, front, and back to obtain a feature vector between grids.

[0161] S214: Concatenate the compressed local information of the point cloud with the feature vectors between the voxel grids to obtain the density features of the local neighborhood of the point cloud.

[0162] For example, in combination with participation Figure 3 and Figure 5 , the output of the feature vector between the local information of the point cloud and the voxel grid is spliced ​​and passed into the MLP in the density embedding module for further feature extraction, so as to obtain a 1024-dimensional feature vector of the local density feature of the embedded point cloud.

[0163] S215: extracting point-by-point features of the source point cloud and the target point cloud in the preprocessed point cloud data respectively.

[0164] Exemplarily, it is composed of five spatial self-attention modules of size (64, 64, 64, 128, 192) connected in series. By connecting these five features of different depths, 512-dimensional point-by-point features with internal connections at different levels are obtained.

[0165] S216: Perform average pooling and maximum pooling operations on the point-by-point features to obtain global features, and extract shared global features from the global features.

[0166] Exemplarily, average pooling and maximum pooling operations are performed on point-by-point features to obtain global features, and the global features are input into a shared global feature extraction module. The shared global feature extraction module is composed of three channel mutual attention modules stacked together. By calculating the cross-attention maps of all channels, information interaction between two features is achieved at the same time, and shared global features are extracted.

[0167] S217: Concatenate the density features and point-by-point features of the local neighborhood of the source point cloud with the shared global features of the target point cloud to obtain a region mask of the source point cloud.

[0168] S218: Concatenate the density features and point-by-point features of the local neighborhood of the target point cloud with the shared global features of the source point cloud to obtain a region mask of the target point cloud.

[0169] Exemplarily, the density features of the local neighborhood of the source point cloud S extracted by the density embedding module are spliced ​​with the point-by-point features of the source point cloud S and the shared global features of the copied target point cloud T, and are passed into the mask estimation module to learn the regional mask of S. The same operation is performed on the target point cloud T to obtain the regional mask of T.

[0170] like Figure 6 As shown, Figure 6 A schematic diagram of a flow chart of another embodiment of a method for laser point cloud registration of a power corridor provided by the present invention. A method for laser point cloud registration of a power corridor, the method comprising:

[0171] S301: Determine the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud.

[0172] For example, given two point clouds, the source point cloud P = {p i ∈R 3 |i=1,2…,N} and target point cloud Q={q j ∈R 3 |j=1,2…,M}, determine the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud. The covariance matrix is ​​shown as follows:

[0173]

[0174] in, is the point cloud centroid of the source point cloud, is the point cloud centroid of the target point cloud, N is the number of source point clouds, M is the number of target point clouds, and the point cloud centroid is used to decentralize the data.

[0175] S302: Obtain the eigenvalues ​​and eigenvectors of the covariance matrix through singular value decomposition, and determine the principal axes of the two groups of point clouds.

[0176] Exemplarily, the eigenvalues ​​and eigenvectors of the covariance matrix are obtained by singular value decomposition, and the principal axes of the two groups of point clouds are determined. The eigenvalues ​​and eigenvectors of the covariance matrix are shown in the following formula:

[0177]

[0178] Among them, Up is the main axis of the source point cloud, Uq is the main axis of the target point cloud, is the transposed matrix of the principal axis of the source point cloud, is the transposed matrix of the principal axis of the target point cloud, Dp is the covariance matrix of the source point cloud, and Dq is the covariance matrix of the target point cloud.

[0179] S303: determining an initial rigid body transformation matrix according to the principal axes of the two groups of point clouds, and performing a rotation transformation on the source point cloud to obtain a point cloud aligned with the principal axis.

[0180] Exemplarily, the initial rigid body transformation matrix is ​​as follows:

[0181]

[0182] Among them, R0, T0 are rotation matrices and translation vectors respectively, is the point cloud centroid of the source point cloud, is the point cloud centroid of the target point cloud, Up is the principal axis of the source point cloud, Uq is the principal axis of the target point cloud, Vp is the transposed matrix of the principal axis of the source point cloud, and Vq is the transposed matrix of the principal axis of the target point cloud;

[0183] The point cloud aligned with the principal axis is determined according to the following formula:

[0184] P′=P×R0+T0;

[0185] Among them, P is the source point cloud.

[0186] S304: Determine the centers of the source point cloud and the target point cloud respectively.

[0187] For example, for two similar point clouds, although the main axis is aligned, there is a possibility that the main axis is reversed, so the opposite main axis situation needs to be considered. Specifically, first determine the center of the source point cloud and the target point cloud respectively, as shown in the following formula:

[0188]

[0190] Among them, L max , L min It is the maximum point coordinate and the minimum point coordinate in two point clouds (source point cloud and target point cloud).

[0191] S305: According to the point cloud centroid of the source point cloud and the point cloud centroid of the target point cloud, combined with the centers of the source point cloud and the target point cloud, the direction of the line connecting the point cloud centroid and the center is determined, and the initial alignment of the source point cloud is performed through the direction of the line connecting the point cloud centroid and the center.

[0192] S306: Sampling the source point cloud to obtain a subset of the source point cloud.

[0193] S307: searching the target point cloud initially aligned with the source point cloud for the closest point to each point in the source point cloud subset, so as to form an initial correspondence between the two point clouds.

[0194] For example, the source point cloud is sampled to obtain a subset P0 of the source point cloud, P0∈P, P0 represents a subset of the source point cloud P, and the nearest point min to each point in the source point cloud subset P0 is found in the target point cloud Q initially aligned with the source point cloud, that is, P0q i -p i P0=min, forming the initial correspondence between the two point clouds, where q i is the target point cloud, p i is the source point cloud, and the RANSAC algorithm is used to remove incorrect matching point pairs.

[0195] S308: Calculate the transformation matrix of the two point clouds according to the initial correspondence relationship to minimize the registration error and obtain a transformed point cloud subset.

[0196] S309: obtaining a new source point cloud position according to the transformed point cloud subset, terminating the iteration according to the comparison between the registration error and the convergence condition or the preset number of iterations, and obtaining the optimal position of the source point cloud.

[0197] Exemplarily, the transformation matrix of the two point clouds is calculated based on the initial correspondence to minimize the registration error and obtain a transformed point cloud subset. The new source point cloud position is obtained based on the transformed point cloud subset. The iteration is terminated based on the comparison between the registration error and the convergence condition or the preset number of iterations to obtain the optimal source point cloud position.

[0198] Specifically, the registration error is expressed as follows:

[0199]

[0200] Where n is the number of iterations, Qi is the i-th target point cloud, P i is the i-th source point cloud;

[0201] If the registration error is determined to be less than the convergence condition (pre-set threshold) or reaches the preset number of iterations, the iteration is terminated and the optimal point cloud subset is obtained; if the registration error is determined to be greater than the convergence condition or does not reach the preset number of iterations, the iteration of the transformed point cloud subset continues.

[0202] like Figure 7 As shown, Figure 7 A schematic diagram of the structure of an embodiment of a power corridor laser point cloud registration system provided by the present invention. A power corridor laser point cloud registration system, the system comprises:

[0203] The point cloud data acquisition module 11 is used to acquire the point cloud data of the power corridor.

[0204] The point cloud set acquisition module 12 of different voxels is used to divide the point cloud data into voxels through a grid to acquire point cloud sets of different voxels.

[0205] The density feature acquisition module 13 is used to acquire the density feature of the local neighborhood of the point cloud according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel.

[0206] The region mask acquisition module 14 is used to combine the density features of the local neighborhood of the point cloud with the point-by-point features of the point cloud data to obtain the region masks of the source point cloud and the target point cloud in the point cloud data.

[0207] The overlapping area point cloud determination module 15 is used to determine that if the area mask of the original point cloud and the target point cloud in the point cloud data is a preset mask value, then the point cloud is the overlapping area point cloud.

[0208] The initial registration module 16 is used to determine the principal axes of the two groups of point clouds according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and perform initial registration on the source point cloud based on the principal axes.

[0209] The optimal source point cloud position acquisition module 17 is used to obtain the change matrix of the source point cloud and the target point cloud according to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, and perform secondary registration on the source point cloud based on the transformation matrix to obtain the optimal source point cloud position.

[0210] Exemplarily, in the point cloud data acquisition module 11, the point cloud data of the power corridor scene is acquired, and in the point cloud set acquisition module 12 of different voxels, the point cloud data is preprocessed to filter out outliers and noise points in the point cloud data; the maximum and minimum values ​​of each point cloud coordinate in the initialized point cloud data are acquired; the grid size is determined according to the maximum and minimum values ​​of the point cloud coordinates; the voxel coordinates to which each point cloud coordinate belongs in the point cloud data are determined; each point cloud coordinate is assigned to a corresponding voxel using the voxel coordinates, and the number of point clouds in each voxel is counted and saved; and the point cloud sets of different voxels are determined according to the number of voxels in each direction in the three-dimensional coordinate system and the number of point clouds in each voxel. Furthermore, in the density feature acquisition module 13, the point cloud feature vector of the point cloud in the point cloud set of the current voxel is multiplied by the feature vector matrix to normalize the point cloud data; the high-dimensional feature vector of the normalized point cloud data is extracted; the rotation matrix corresponding to the high-dimensional feature vector is obtained, and the high-dimensional feature vector is aligned by multiplying the rotation matrix with the point cloud feature vector of the point cloud in the point cloud set of the current voxel; the aligned high-dimensional feature vector is aggregated through the maximum pooling operation to obtain the global features of the current point cloud set, and the global features are compressed to output the local information of the point cloud in the point cloud set of the current voxel; according to the position of the grid where the point cloud set of the current voxel is located, the number of point clouds in the point cloud set of the current voxel and the number of point clouds in the point cloud sets of all voxels in adjacent directions are obtained; the number of point clouds of the current voxel and the number of point clouds of its adjacent voxels are spliced ​​in spatial order to obtain the feature vector between the voxel grids; the compressed point cloud local information is spliced ​​with the feature vector between the voxel grids to obtain the density features of the local neighborhood of the point cloud. Further, in the regional mask acquisition module 14, the point-by-point features of the preprocessed point cloud data are extracted; the point-by-point features are average pooled and max pooled to obtain global features, and shared global features are extracted from the global features; the density features of the local neighborhood of the point cloud, the shared global features and the point-by-point features of the point cloud data are spliced ​​to obtain the regional masks of the source point cloud and the target point cloud in the point cloud data. In the overlapping area point cloud determination module 15, the principal axes of the two groups of point clouds are determined according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and the source point cloud is initially aligned based on the principal axis. Further, in the initial alignment module 16, the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud is determined; the eigenvalues ​​and eigenvectors of the covariance matrix are obtained by singular value decomposition, and the principal axes of the two groups of point clouds are determined; the initial rigid body transformation matrix is ​​determined according to the principal axes of the two groups of point clouds, and the source point cloud is transformed by rotation to obtain a point cloud aligned with the principal axis; the source point cloud is initially aligned by comparing the center of mass of the point cloud with the direction of the center line.Furthermore, in the optimal source point cloud position acquisition module 17, the source point cloud is sampled to obtain a subset of the source point cloud; the nearest point to each point in the source point cloud subset is found in the target point cloud that is initially aligned with the source point cloud to form an initial correspondence between the two point clouds; the transformation matrix of the two point clouds is calculated based on the initial correspondence to minimize the alignment error and obtain a transformed point cloud subset; the new source point cloud position is obtained based on the transformed point cloud subset, and the iteration is terminated based on the comparison of the alignment error with the convergence condition or the preset number of iterations to obtain the optimal position of the source point cloud.

[0211] like Figure 8 As shown, Figure 8 The computer device 20 includes a memory 21 and a processor 22. The memory 21 stores a computer program, and the processor 22 executes the computer program when working to implement the following. Figure 1 , Figure 4 , Figure 6 The method shown.

[0212] The specific technical details of the flexible power supply method based on interconnection of the distributed terminals implemented when the terminal 20 executes the computer program have been discussed in detail in the aforementioned method steps, so they will not be repeated here.

[0213] like Fig. 9 As shown, Fig. 9 The storage medium 30 stores at least one computer program 31, which is executed by the processor 22 to implement the following. Figure 1 , Figure 4 , Figure 6 In one embodiment, the storage medium 30 may be a storage chip, a hard disk, a mobile hard disk, a USB flash drive, an optical disk, or other readable and writable storage tools, or a server, etc.

[0214] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0215] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0216] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the present application. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the present application shall be subject to the attached claims.

Claims

1. A method for laser point cloud registration of power corridors, characterized in that: The method comprises: Obtain point cloud data of the power corridor scene; Dividing the point cloud data into voxel forms through a grid to obtain point cloud sets of different voxels; Obtain density features of the local neighborhood of the point cloud according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel; The density feature of the local neighborhood of the point cloud is spliced ​​with the point-by-point feature of the point cloud data to obtain a regional mask of a source point cloud and a target point cloud in the point cloud data; Determining that the region masks of the source point cloud and the target point cloud in the point cloud data are preset mask values, then the point cloud is an overlapping region point cloud; Determine the principal axes of the two groups of point clouds according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and perform initial registration on the source point cloud based on the principal axes; According to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, the change matrix of the source point cloud and the target point cloud is obtained, and the source point cloud is re-registered based on the transformation matrix to obtain the optimal source point cloud position.

2. The method for laser point cloud registration of a power corridor according to claim 1, characterized in that: The step of dividing the point cloud data into voxels by grid to obtain point cloud sets of different voxels specifically includes: Preprocessing the point cloud data to filter out outliers and noise points in the point cloud data; Obtaining the maximum and minimum values ​​of each point cloud coordinate in the initialized point cloud data; Determine the grid size according to the maximum and minimum values ​​of the point cloud coordinates; Determine the voxel coordinates to which each point cloud coordinate belongs according to the point cloud data; Using the voxel coordinates, each point cloud coordinate is assigned to a corresponding voxel, and the number of point clouds in each voxel is counted and saved; The point cloud sets of different voxels are determined according to the number of voxels in each direction in the three-dimensional coordinate system and the number of point clouds in each voxel.

3. A power corridor laser point cloud registration method according to claim 2, characterized in that: The step of obtaining the density feature of the local neighborhood of the point cloud according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel specifically includes: Normalizing the point cloud data by multiplying the point cloud feature vector of the point cloud in the point cloud set of the current voxel with the feature vector matrix; Extracting a high-dimensional feature vector of the normalized point cloud data; Obtaining a rotation matrix corresponding to the high-dimensional feature vector, and aligning the high-dimensional feature vector by multiplying the rotation matrix with the point cloud feature vector of the point cloud in the point cloud set of the current voxel; The aligned high-dimensional feature vectors are aggregated through a maximum pooling operation to obtain global features of the current point cloud set, and the global features are compressed to output local point cloud information within the point cloud set of the current voxel; According to the position of the grid where the point cloud set of the current voxel is located, the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in adjacent directions are obtained; The number of point clouds of the current voxel and the number of point clouds of its adjacent voxels are spliced ​​in spatial order to obtain the feature vector between the voxel grids; The compressed local information of the point cloud is concatenated with the feature vectors between the voxel grids to obtain the density features of the local neighborhood of the point cloud.

4. The method for laser point cloud registration of a power corridor according to claim 3 is characterized in that: The step of obtaining the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in adjacent directions according to the position of the grid where the point cloud set of the current voxel is located specifically includes: It is determined that the point cloud set of the current voxel is in the edge grid, then for the missing adjacent direction, the number of point clouds of the point cloud set is marked by a preset marking value, and the number of point clouds of the point cloud sets of all voxels in the adjacent direction that are not missing is obtained; If it is determined that the point cloud set of the current voxel is not in the edge grid, the number of point clouds of the point cloud set of the current voxel and the number of point clouds of the point cloud sets of all voxels in adjacent directions are obtained.

5. The method for laser point cloud registration of a power corridor according to claim 4 is characterized in that: The density feature of the local neighborhood of the point cloud is spliced ​​with the point-by-point feature of the point cloud data to obtain the regional mask of the source point cloud and the target point cloud in the point cloud data, specifically including: Extracting point-by-point features of the source point cloud and the target point cloud in the preprocessed point cloud data respectively; Performing average pooling and maximum pooling operations on the point-by-point features to obtain global features, and extracting shared global features from the global features; The density feature of the local neighborhood of the source point cloud and the point-by-point feature are concatenated with the shared global feature of the target point cloud to obtain a region mask of the source point cloud; The density features of the local neighborhood of the target point cloud and the point-by-point features are concatenated with the shared global features of the source point cloud to obtain a region mask of the target point cloud.

6. The speech synthesis method based on the multi-subband generation strategy according to any one of claims 1 to 5, characterized in that: After the density feature of the local neighborhood of the point cloud is spliced ​​with the point-by-point feature of the point cloud data to obtain the regional mask of the source point cloud and the target point cloud in the point cloud data, the method further specifically includes: Defining a source point cloud loss function according to a true mask and a region mask of the source point cloud; Defining a target point cloud loss function according to a true mask and a region mask of the target point cloud; Determine a total loss function by using the source point cloud loss function and the target point cloud loss function; The region mask is optimized according to a total loss function.

7. A power corridor laser point cloud registration method according to claim 6, characterized in that: Determining the principal axes of the two groups of point clouds according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and performing initial registration on the source point cloud based on the principal axes specifically includes: Determine the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud; Obtain the eigenvalues ​​and eigenvectors of the covariance matrix by singular value decomposition, and determine the principal axes of the two groups of point clouds; Determine an initial rigid body transformation matrix according to the principal axes of the two groups of point clouds, and obtain a point cloud aligned with the principal axes by rotationally transforming the source point cloud; Determine the centers of the source point cloud and the target point cloud respectively; According to the point cloud centroid of the source point cloud and the point cloud centroid of the target point cloud, combined with the centers of the source point cloud and the target point cloud, the direction of the line connecting the point cloud centroid and the center is determined, and the initial alignment of the source point cloud is performed through the direction of the line connecting the point cloud centroid and the center.

8. The method for laser point cloud registration of a power corridor according to claim 7, characterized in that: The initial rigid body transformation matrix is ​​determined according to the principal axes of the two groups of point clouds, and the source point cloud is transformed by rotation to obtain a point cloud aligned with the principal axis, specifically comprising: according to Determine the initial rigid body transformation matrix, where R0 and T0 are the rotation matrix and translation vector respectively. is the point cloud centroid of the source point cloud, is the point cloud centroid of the target point cloud, Up is the principal axis of the source point cloud, and Uq is the principal axis of the target point cloud. is the transposed matrix of the principal axis of the source point cloud, is the transposed matrix of the principal axis of the target point cloud; according to P′=P×R0+T0 The source point cloud is transformed by rotation to obtain a point cloud aligned with the principal axis, wherein P is the source point cloud.

9. The method for laser point cloud registration of a power corridor according to claim 7, characterized in that: The step of obtaining a change matrix of the source point cloud and the target point cloud according to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, and performing secondary registration on the source point cloud based on the transformation matrix to obtain the optimal source point cloud position specifically includes: Sampling the source point cloud to obtain a subset of the source point cloud; Finding the closest point to each point in the source point cloud subset in the target point cloud initially aligned with the source point cloud to form an initial correspondence between the two point clouds; Calculating a transformation matrix of the two point clouds according to the initial correspondence relationship to minimize the registration error and obtain a transformed point cloud subset; A new source point cloud position is obtained according to the transformed point cloud subset, and the iteration is terminated according to the comparison between the registration error and the convergence condition or the preset number of iterations to obtain the optimal position of the source point cloud.

10. A power corridor laser point cloud registration method according to claim 9, characterized in that: The step of obtaining a new source point cloud position according to the transformed point cloud subset until the registration error reaches a convergence condition or a preset number of iterations, then the iteration is terminated, and the optimal position of the source point cloud is obtained, specifically including: according to Determine the registration error, where n is the number of iterations, Q i is the i-th target point cloud, P i is the i-th source point cloud; If it is determined that the registration error is less than a convergence condition or reaches a preset number of iterations, the iteration is terminated and an optimal point cloud subset is obtained; If it is determined that the registration error is greater than the convergence condition or the preset number of iterations is not reached, then the iteration of the transformed point cloud subset continues.

11. A power corridor laser point cloud registration system, characterized in that: The system comprises: Point cloud data acquisition module, used to acquire point cloud data of the power corridor; A point cloud set acquisition module of different voxels, used for dividing the point cloud data into voxels through a grid to acquire point cloud sets of different voxels; A density feature acquisition module is used to acquire the density feature of the local neighborhood of the point cloud according to the rotation matrix and the number of point clouds in the point cloud set of the current voxel; A region mask acquisition module, used for splicing the density features of the local neighborhood of the point cloud with the point-by-point features of the point cloud data to obtain the region masks of the source point cloud and the target point cloud in the point cloud data; An overlapping area point cloud determination module is used to determine that if the area mask of the original point cloud and the target point cloud in the point cloud data is a preset mask value, then the point cloud is the overlapping area point cloud; An initial registration module, used to determine the principal axes of the two groups of point clouds according to the covariance matrix of the source point cloud and the target point cloud in the overlapping area point cloud, and perform initial registration on the source point cloud based on the principal axes; The optimal source point cloud position acquisition module is used to obtain the change matrix of the source point cloud and the target point cloud according to the correspondence between the source point cloud in the overlapping area point cloud and the target point cloud in the initial registration, and perform secondary registration on the source point cloud based on the transformation matrix to obtain the optimal source point cloud position.

12. A computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 10.

13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 10.

Citation Information

Patent Citations

  • Single-line point cloud and multi-line point cloud registration method and device

    CN112446907A

  • Cross-country automatic trolley long-distance real-time positioning system and method based on laser radar

    CN115372989A

  • Laser point cloud iterative registration method and device and storage medium

    CN115760939A

  • Registration method for inclined dense matching point cloud data and laser radar point cloud data

    CN116245925A

  • Three-dimensional morphology method for large-scale cuboid in moving state

    CN116503490A

Cited By

  • Partial-to-overall non-rigid point cloud registration method and system

    CN119205865A

  • Point cloud acquisition equipment calibration method and equipment, storage medium and program product

    CN119494882A

  • Multi-modal interaction method and system for virtual images of smart exhibition hall

    CN119648928A

  • A multimodal interactive method and system for virtual images in a smart exhibition hall

    CN119648928B

  • Self-adaptive point cloud registration method based on GM robust kernel function

    CN120355765A