Large-scale point cloud registration method, system and device based on power grid transmission line and storage medium

Abstract

The invention discloses a large-scale point cloud registration method, system and device based on a power grid power transmission line and a storage medium, and the method comprises the steps: collecting point cloud data of the power transmission line under different visual angles, and forming a source point cloud set and a target point cloud set; down-sampling point cloud data in the source point cloud set and the target point cloud set, dynamically adjusting the sampling granularity of different density regions, and obtaining representative point cloud data from the source point cloud set and the target point cloud set; extracting point cloud feature points, accelerating neighborhood search through a KD-Tree structure, and adaptively adjusting a neighborhood search radius; constructing a geometric topology model; performing coarse registration based on an improved RANSAC algorithm, selecting non-collinear feature points, and optimizing a transformation model; an improved ICP algorithm is adopted to carry out fine registration, and normal vector features are utilized to introduce weighting factors to optimize registration precision. According to the invention, large-scale power transmission line point cloud data can be efficiently and accurately processed, and the method is especially suitable for power inspection, three-dimensional modeling and maintenance scenes.

Description

Technical Field

[0001] The present invention relates to computer vision and three-dimensional point cloud processing technology, and in particular to a large-scale point cloud registration method, system, equipment and storage medium based on power grid transmission lines. Background Art

[0002] With the development of remote sensing technologies such as drones and LiDAR, 3D point clouds have been widely used in the fields of inspection, maintenance, and modeling of power transmission lines. Point cloud data can accurately describe the 3D structure of equipment such as towers and conductors in transmission lines. However, since the point cloud data of transmission lines is usually large in scale and complex in structure, traditional point cloud registration methods have many shortcomings in this application.

[0003] The ICP (Iterative Closest Point) algorithm is a classic point cloud registration method, but it is easy to fall into the local optimal solution when processing large-scale point clouds, and the computational efficiency is low. Especially under the complex geometric structure of the transmission line (such as the linear structure of the conductor and the vertical features of the tower), the ICP algorithm cannot fully utilize these global features, resulting in a decrease in registration accuracy and robustness. In addition, existing feature matching methods, such as octree downsampling, ISS feature point extraction, directional histogram, etc., although they perform well in some scenarios, they still face the following technical bottlenecks in complex scenarios of power grid transmission lines:

[0004] (1) Low computational efficiency of large-scale point clouds: The point cloud data of power transmission lines has a wide range and uneven density. Traditional methods require huge computational effort when processing such large-scale data, which makes it difficult to meet the requirements of practical applications.

[0005] (2) Insufficient local features: Existing feature extraction algorithms usually rely on local geometric information, which makes it difficult to effectively utilize the global structural features in the transmission line and cannot ensure the accurate matching of key geometric structures.

[0006] (3) Poor registration robustness: In complex scenes, noise points, missing points, etc. can easily affect the existing registration methods, resulting in registration failure or reduced accuracy. Summary of the invention

[0007] Purpose of the invention: The purpose of the present invention is to provide a large-scale point cloud registration method, system, device and storage medium based on power grid transmission lines, which can solve the problems of low computational efficiency, insufficient registration accuracy and poor robustness of existing point cloud registration methods when dealing with complex transmission line scenes.

[0008] Technical solution: A large-scale point cloud registration method based on power grid transmission lines of the present invention comprises:

[0009] Collect point cloud data of the transmission line from different perspectives to form a source point cloud set P and a target point cloud set Q;

[0010] Adaptively downsampling the point cloud data of the source point cloud set P and the target point cloud set Q respectively, and dynamically adjusting the intervals of the selected points in different density areas, obtaining representative point cloud data from the source point cloud set P to generate a first point cloud data set, and obtaining representative point cloud data from the target point cloud set Q to generate a second point cloud data set;

[0011] The improved ISS algorithm is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively. In the process of extracting point cloud feature points, the KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure.

[0012] Based on the feature point set, the directional histogram feature description model is constructed in combination with the geometric topology information to generate a geometric topology model; the geometric topology model captures the spatial distribution direction of the feature points in the feature point set within the neighborhood, reflects the local structural characteristics of the point cloud, and provides stable description features for subsequent registration;

[0013] The improved RANSAC algorithm is used to roughly align the feature point sets of the source point cloud set and the target point cloud set, and the non-collinear feature points that meet the geometric relationship are selected to optimize the rigid body transformation matrix; by minimizing the matching error, the initial alignment matrix from the source point cloud set to the target point cloud set is obtained;

[0014] Based on the rough registration results, the improved ICP algorithm is used for fine registration; the weighting factor is introduced by using the normal vector feature and combined with KD-Tree accelerated search, and the rigid body transformation matrix is ​​gradually optimized to improve the registration accuracy, and finally the precise registration results of the source point cloud set and the target point cloud set are output.

[0015] Further, adaptive octree downsampling is performed on the point cloud data of the source point cloud set and the target point cloud set respectively, and the intervals of the selected points in different density areas are dynamically adjusted, representative point cloud data are obtained from the source point cloud set to generate a first point cloud data set, and representative point cloud data are obtained from the target point cloud set to generate a second point cloud data set, including:

[0016] Initialize the octree: Determine the three-dimensional space range occupied by the point cloud data in the source point cloud set P and the target point cloud set Q of the transmission line, which is defined as [x min ,x max ,y min ,y max ,z min ,z max ];

[0017] Divide octree nodes: For the root node of the octree, use the following formula to recursively divide it into eight child nodes, taking the x direction as an example, and the y and z directions are similar:

[0018]

[0019] Among them, i = 0, 1, 2, 3, 4, 5, 6, 7, representing the index of the eight child nodes in the x direction;

[0020] Determine adaptive subdivision: Calculate the point cloud density ρ in each child node, the formula is as follows:

[0021]

[0022] Where N represents the number of points in the node; V represents the spatial volume of the node. For a rectangular node, its volume is:

[0023] V=(x max -x min )(y max -y min )(z max -z min )

[0024] Set the density threshold of the point cloud data in the transmission tower area and the density threshold of the point cloud data in the open area. When ρ>ρ threshold When , the node is further subdivided; wherein, the density threshold of the point cloud data in the transmission tower area is greater than the density threshold of the point cloud data in the open area;

[0025] Downsampling to select representative points: After completing the construction of the octree, select representative points from each node;

[0026] Through the above steps, a part of representative points can be selected from the original large-scale point cloud data to complete the adaptive octree downsampling.

[0027] Furthermore, after the construction of the octree is completed, representative points are selected from each node, including:

[0028] Using the node centroid method, for a point set {P i (x i ,y i ,z i )}, center of gravity P center The calculation formula is as follows:

[0029]

[0030] Where N represents the number of nodes inside the node.

[0031] Furthermore, the improved ISS algorithm is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively. In the process of extracting point cloud feature points, the KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure, including:

[0032] The local neighborhood is determined using the KD-Tree structure. For each point P = (x, y, z) in the point cloud data, its local neighborhood point cloud set N(P) within the radius r is determined.

[0033] For each point P in the local neighborhood point cloud set N(P) q , establish a local coordinate system, and set the search radius, where the neighborhood search radius of the point cloud near the transmission tower is set to be larger than the neighborhood search radius of the point cloud in the smooth area of ​​the transmission line; find all points within the search radius and calculate the weight of each point. The formula is:

[0034]

[0035] In the formula, ω i Represents point p q The weight is inversely proportional to the number of points in its local neighborhood; r represents the search radius; p i Represents points in the neighborhood. The sparser the points in the sampling, the higher the weight;

[0036] Calculate the three-dimensional weighted covariance matrix for each point using the formula:

[0037]

[0038] Solve the eigenvalues ​​of each point based on the three-dimensional weighted covariance matrix of each point The three eigenvalues ​​are arranged from large to small. The relationship between them reflects the main features of the data points and their relative importance. Two thresholds ε1 and ε2 are set for the point cloud with complex structure near the transmission tower and the point cloud in the smooth area of ​​the transmission line. If the query point p q As an ISS feature point, it must meet

[0039]

[0040] Furthermore, based on the feature point set and combined with the geometric topological information, the directional histogram feature description is performed to construct a geometric topological model that reflects the local structural characteristics of the point cloud, including:

[0041] For each feature point P and its local neighborhood point cloud set N(P), calculate the direction vector of the neighborhood point relative to the feature point P;

[0042] Let the neighborhood point be Q i∈N(P), direction vector The calculation method is That is, the direction vector is obtained by subtracting the coordinates of the feature point from the coordinates of the neighborhood point, thereby obtaining a set of vectors reflecting the spatial distribution direction of the neighborhood points relative to the feature points;

[0043] Divide the vector set according to the direction space, count the number of direction vectors of the neighborhood points falling in each interval, and construct a histogram;

[0044] Set the distance threshold d threshold and the angle threshold θ threshold , if the distance between two neighboring points is less than d threshold And the angle between them and the feature point is less than θ threshold , it is considered that there is a connectivity relationship between the two points, and the connected point pairs in the neighborhood are identified, and then a geometric topological model is established to quantify the local topological information of the point cloud into feature descriptors.

[0045] Furthermore, the improved RANSAC algorithm is used to roughly align the feature point sets of the source point cloud set and the target point cloud set, select non-collinear feature points that meet the geometric relationship, and optimize the rigid body transformation matrix; by minimizing the matching error, the initial alignment matrix from the source point cloud set to the target point cloud set is obtained, including:

[0046] Randomly select 4 corresponding data points from the source point cloud set P and the target point cloud set Q, denoted as p1, p2, p3, p4 and q1, q2, q3, q4 respectively;

[0047] The rigid body transformation matrix H is solved by the least squares method, so that the source point cloud is transformed by the rigid body rotation and translation to obtain a new point cloud set P`=R c P+t c , where R c is the rotation matrix, t c is the translation vector;

[0048] Calculate the distance projection error between the new source point cloud set PP' and the corresponding points of the target point cloud set Q under the rigid body transformation matrix H of the source point cloud set P;

[0049] If the distance projection error of the point pair is less than the set threshold error d t (d t =d 2 (p` i ,Hp i ), where d is the distance between corresponding points, then add the point to the sample internal point set I and calculate the matching score S i , that is, the ratio of the internal points to the total number of point clouds;

[0050] Set the number of loops N and the matching score threshold S min, if the matching score of the point cloud is less than S min , then jump out of the loop, otherwise repeat the above steps until the number of iterations reaches N;

[0051] The number of sample inliers under different rigid body transformation matrices is counted, and the rigid body transformation matrix with the largest number of inliers is taken as the optimal rigid body transformation matrix. The rigid body transformation matrix is ​​optimized to obtain the source point cloud after coarse registration.

[0052] Furthermore, based on the rough registration results, the improved ICP algorithm is used for fine registration. The weighting factor is introduced by using the normal vector feature and combined with KD-Tree accelerated search, and the rigid body transformation matrix is ​​gradually optimized to improve the registration accuracy. Finally, the precise registration results of the source point cloud set and the target point cloud set are output, including:

[0053] Input source point cloud set P and target point cloud set Q to be registered;

[0054] Establish KD-Tree indexes for the source point cloud set P and the target point cloud set Q respectively; taking the source point cloud set as an example, for each point p, search the neighborhood point set N(P) and construct the covariance matrix C.

[0055]

[0056] Among them, q is a point in the neighborhood, and the eigenvector corresponding to the minimum eigenvalue of the covariance matrix is ​​the normal vector of point p;

[0057] Using the normal vector features of the source point cloud set and the target point cloud set, a weighted value is assigned to each pair of matching points in the registration process. According to the normal vector information, a weighting factor ω is defined to optimize the matching. Assume and is the normal vector of the corresponding point pair, and the weight ω can be expressed as:

[0058]

[0059] Where σ represents the smoothing parameter;

[0060] Construct a descriptor based on each point and its normal information to enhance the uniqueness of each feature point;

[0061] Traverse all points p in the source point cloud set P i , use KD-Tree to search for each p in the target point cloud set Q i The nearest neighbor point q i , forming a matching point pair (p i ,q i )’s set s;

[0062] Calculate the average distance d between matching points in the matching point set s m With standard deviation σ, if a point is d siThe distance satisfies the 3σ criterion, d si <d m +3σ, then the matching point pair is retained, otherwise it is removed;

[0063] According to d si Sort the matching point set s from small to large and select the top 70% matching point pairs (p i ,q i ) coordinates, and solve the rotation matrix R and translation vector t by weighted least squares method;

[0064] The rigid transformation matrix [RT] is obtained by using the SVD method, satisfying q i ≈R·p i +t;

[0065] Calculate the registration error evaluation function E k (RT):

[0066]

[0067] Among them, ω i represents the weighting factor; n represents the number of valid matching point pairs; k represents the kth conversion. k Difference D k =E k -E k-1 If the set threshold is met, it is converged, otherwise it continues to iterate.

[0068] Based on the same inventive concept, a large-scale point cloud registration system based on power grid transmission lines of the present invention comprises:

[0069] The data acquisition module is used to collect point cloud data of the transmission line from different viewing angles to form a source point cloud set P and a target point cloud set Q;

[0070] A downsampling module is used to perform adaptive octree downsampling on the point cloud data of the source point cloud set P and the target point cloud set Q, respectively, and dynamically adjust the interval of the selected points in different density areas, obtain representative point cloud data from the source point cloud set P to generate a first point cloud data set, and obtain representative point cloud data from the target point cloud set Q to generate a second point cloud data set;

[0071] A feature point extraction module is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively by using an improved ISS algorithm. In the process of extracting point cloud feature points, a KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure;

[0072] The model building module is used to build a directional histogram feature description model based on the feature point set and combined with geometric topological information to generate a geometric topological model; the geometric topological model captures the spatial distribution direction of the feature points in the feature point set within the neighborhood, reflects the local structural characteristics of the point cloud, and provides stable description features for subsequent registration;

[0073] The coarse registration module is used to coarsely register the feature point sets of the source point cloud set and the target point cloud set using the improved RANSAC algorithm, select non-collinear feature points that meet the geometric relationship, and optimize the rigid body transformation matrix; by minimizing the matching error, the initial alignment matrix of the source point cloud set to the target point cloud set is obtained;

[0074] The fine registration module is used to perform fine registration based on the rough registration results using the improved ICP algorithm. It uses the normal vector feature to introduce a weighting factor and combines it with the KD-Tree accelerated search to gradually optimize the rigid body transformation matrix to improve the registration accuracy. Finally, it outputs the precise registration results of the source point cloud set and the target point cloud set.

[0075] Based on the same inventive concept, a large-scale point cloud registration device based on power grid transmission lines of the present invention includes a processor and a memory, wherein computer instructions are stored in the memory, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the electronic device implements the steps of the above-mentioned large-scale point cloud registration method based on power grid transmission lines.

[0076] Based on the same inventive concept, a computer-readable storage medium of the present invention stores a computer program thereon, and when the program is executed by a processor, the steps of the above-mentioned large-scale point cloud registration method based on power grid transmission lines are implemented.

[0077] Beneficial effects: Compared with the prior art, the significant technical effects of the present invention are:

[0078] Adaptive octree downsampling strategy uses adaptive octree method to perform intelligent dynamic downsampling of point cloud data in different density areas. It solves the problem that traditional fixed downsampling is difficult to adapt to the complex structure and uneven density of transmission lines. It achieves efficient data compression while retaining key structural information and improving data processing efficiency.

[0079] The feature point extraction based on the ISS algorithm is improved. The neighborhood search is accelerated through KD-Tree. The neighborhood search radius is adaptively adjusted according to the local geometric characteristics to extract highly representative ISS feature points. This solves the problem that the traditional feature extraction algorithm is inefficient on large, uneven density point cloud data and difficult to capture key geometric structures. It achieves fast and stable feature point extraction and enhances the representativeness and distribution uniformity of the registration points.

[0080] The directional histogram is combined with the geometric topological model to form a geometric topological model. The directional histogram captures the spatial distribution of the feature point neighborhood and combines the geometric topological information of the feature points to form a geometric topological model. This solves the problem that the existing description methods have limited geometric information expression for complex scene point clouds and are difficult to guarantee accuracy. It improves the integrity and reliability of the point cloud feature description and helps to improve the registration accuracy in complex scenes.

[0081] Improve RANSAC rough registration and non-collinear feature point selection. Rough registration is performed through the improved RANSAC method, and non-collinear feature points are selected to improve matching robustness. The problem that traditional rough registration algorithms are difficult to avoid collinearity during random sampling, resulting in unstable registration, is solved. The accuracy and efficiency of rough registration are improved, providing reliable initial alignment for subsequent fine registration.

[0082] Improved ICP precision registration with normal vector weighting combines the normal vector weighting ICP algorithm for precision registration and introduces weighting factors to enhance feature description. This solves the problem that the traditional ICP algorithm is sensitive to noise and outliers and is prone to falling into local optimality. It enhances the robustness and accuracy of point cloud registration and is suitable for fine registration of large-scale data. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] Figure 1 It is a flow chart of a large-scale point cloud registration method based on power grid transmission lines disclosed in an embodiment of the present invention;

[0084] Figure 2 It is a structural schematic diagram of a large-scale point cloud registration system based on power grid transmission lines disclosed in an embodiment of the present invention;

[0085] Figure 3 It is a structural schematic diagram of a large-scale point cloud registration system based on power grid transmission lines disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0086] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. It will be understood by those skilled in the art that the purposes and advantages that can be achieved with the present invention are not limited to the specific description of the above beneficial effects, and the above and other purposes that can be achieved by the present invention will be more clearly understood based on the following detailed description.

[0087] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed in the present invention can be implemented in hardware, software or a combination of the two. Whether it is implemented in hardware or software depends on the specific application and design of the technical solution. Professional and technical personnel can use different methods to implement the functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention.

[0088] Reference to "embodiments" in the present invention means that a particular feature, structure, or characteristic described in conjunction with the embodiments may be included in at least one embodiment of the present invention. The appearance of the phrase in various places in the specification does not necessarily refer to the same embodiment, nor is it an independent or alternative embodiment that is mutually exclusive with other embodiments. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0089] Example 1

[0090] See also Figure 1 , Figure 1 1 is a flow chart of a large-scale point cloud registration method based on power grid transmission lines disclosed in an embodiment of the present invention. Figure 1 The large-scale point cloud registration method described is applied to power systems, such as three-dimensional modeling and inspection analysis of power transmission lines, which is not limited in the embodiments of the present invention. Figure 1 As shown, the large-scale point cloud registration method based on power grid transmission lines may include the following operations:

[0091] S1. Data collection: Collect point cloud data of the transmission line from different perspectives to form a source point cloud set P and a target point cloud set Q.

[0092] Obtain two sets of transmission line point cloud sets to be processed from different perspectives collected by drone inspection, namely the source point cloud set P and the target point cloud set Q.

[0093] S2, adaptive octree downsampling: adaptive octree downsampling is performed on the point cloud data of the source point cloud set P and the target point cloud set Q respectively, and the interval of the selected points in different density areas (i.e., sampling granularity) is dynamically adjusted to obtain representative point cloud data from the source point cloud set P to generate the first point cloud data set, and obtain representative point cloud data from the target point cloud set Q to generate the second point cloud data set. The specific steps are as follows:

[0094] S2.1. Initialize the octree: After obtaining the point cloud data to be processed, we first need to determine the three-dimensional space range occupied by the point cloud data in the source point cloud set P and the target point cloud set Q of the transmission line, which is defined as [x min ,x max ,y min ,y max ,z min ,z max ].

[0095] In this embodiment, the geographic coordinate range is set according to the actual data collection. For example, if the transmission line is located in an area with an east-west extension length of 1000 meters, a north-south width of 200 meters, a height range of 50 meters from the ground to the top of the transmission tower, and the geographic coordinate origin is located in the lower left corner of the area, then x can be set. min =0, x max =1000,y min =0,y max =200, z min =0, z max =50.

[0096] S2.2. Divide the octree nodes: For the root node of the octree, use the following formula to recursively divide it into eight child nodes (taking the x direction as an example, the y and z directions are similar):

[0097]

[0098] Among them, i = 0, 1, 2, 3, 4, 5, 6, 7, representing the index of the eight child nodes in the x direction.

[0099] S2.3, determine adaptive subdivision: calculate the point cloud density ρ in each child node, the formula is as follows:

[0100]

[0101] Where N represents the number of points in the node; V represents the spatial volume of the node. For a rectangular node, its volume is:

[0102] V=(x max -x min )(y max -y min )(z max -z min )

[0103] Considering the characteristics of the point cloud data of the transmission line, the point cloud data is dense near the transmission tower, and the point cloud in the open area is sparse. Set the density threshold of the point cloud data in the transmission tower area and the density threshold of the point cloud data in the open area; among them, the density threshold of the point cloud data in the transmission tower area is greater than the density threshold of the point cloud data in the open area;

[0104] In this embodiment, for the transmission tower area, the density threshold ρ of the point cloud data is set threshold =100 points / m3, the density threshold of point cloud data in the open area is set to ρ threshol =20 points / cubic meter. When ρ>ρ threshold , the node is further subdivided.

[0105] S2.4, downsampling and selecting representative points: After the octree is constructed, a representative point is selected from each node. In this embodiment, the node centroid method is used.

[0106] {P i (x i ,y i ,z i )}, center of gravity P center The calculation formula is as follows:

[0107]

[0108] Where N represents the number of nodes inside the node.

[0109] Through the above steps, a part of representative points can be selected from the original large-scale point cloud data to complete the adaptive octree downsampling.

[0110] In the present invention, when downsampling the point cloud, the point cloud is first divided into multiple octree structures, and the three-dimensional space is recursively subdivided to generate multiple sub-cubes. Considering the complexity and density difference of the point cloud data of the power grid transmission line, an adaptive octree downsampling algorithm is adopted. High-density areas (such as transmission towers) maintain a higher resolution, while low-density areas (such as open areas) reduce sampling points to ensure a balance between local refinement and global simplification. This method overcomes the problem that a fixed sampling scale cannot adapt to different point cloud densities.

[0111] After downsampling, the centroid of the point cloud within each octree node is selected as the representative point to generate a representative point cloud dataset, ensuring that important geometric features are preserved in the subsequent registration process.

[0112] S3, ISS feature point extraction: The improved ISS algorithm is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively. In the process of extracting point cloud feature points, the KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure.

[0113] After octree downsampling, the amount of point cloud data is still large, so the ISS algorithm needs to be used to extract feature points. The ISS algorithm calculates the eigenvalues ​​of the covariance matrix of each point field and selects feature points through the linear relationship between the eigenvalues. In order to improve the efficiency of extracting feature points of large-scale point cloud data, the ISS algorithm is improved, combined with the KD-Tree structure for accelerated processing, and the feature point extraction parameters are adaptively adjusted according to the local geometric characteristics of the point cloud. The specific steps of step S3 are as follows:

[0114] S3.1. Use the KD-Tree structure to determine the local neighborhood. For each point P = (x, y, z) in the point cloud data, determine its local neighborhood point cloud set N(P) within the radius r.

[0115] In this embodiment, starting from the root node, the distance d from the point represented by the root node to P is calculated using the Euclidean distance formula Where (x i ,y i ,z i ) is the point represented by the root node.

[0116] If d≤r, then add the point represented by the root node to the local neighborhood point cloud set N(P); if d≥r, and the root node is divided according to the x dimension, if x <x i , then search in the left subtree, if x>x i , then search in the right subtree until there is only one point in the subtree or the preset stopping condition is met, such as the subtree depth reaches 10 layers.

[0117] S3.2. After obtaining the local neighborhood point cloud set N(P), for each point P in the local neighborhood point cloud set N(P) q , establish a local coordinate system, and set the search radius. Among them, the neighborhood search radius of the point cloud near the transmission tower is set to be larger than the neighborhood search radius of the point cloud in the smooth area of ​​the transmission line; in this embodiment, for the point cloud near the transmission tower, due to the complex structure, the neighborhood search radius of the point cloud near the transmission tower is set to r = 0.8 meters, and in the area where the transmission line is relatively smooth, the neighborhood search radius of the point cloud in the smooth area of ​​the transmission line is set to r = 0.3 meters. Find all points within the search radius and calculate the weight of each point. The formula is:

[0118]

[0119] In the formula, ω i Represents point p q The weight is inversely proportional to the number of points in its local neighborhood; r represents the search radius; p i Represents points in the neighborhood. The sparser the points in the sampling, the higher the weight.

[0120] S3.3. Calculate the three-dimensional weighted covariance matrix of each point. The formula is:

[0121]

[0122] S3.4. Solve the eigenvalues ​​of each point based on the three-dimensional weighted covariance matrix of each point The three eigenvalues ​​are arranged from large to small. The relationship between them reflects the main features of the data points and their relative importance. Two thresholds ε1 and ε2 are set. For the point cloud with complex structure near the transmission tower, ε1 = 0.3 and ε2 = 0.1 are set. For the area with smooth transmission line, ε1 = 0.6 and ε2 = 0.2 are set. If the query point p q As an ISS feature point, it must meet

[0123]

[0124] In the present invention, after adaptive octree downsampling, a point cloud data set with reasonable data volume and intact structure is obtained. Based on this data set, the ISS algorithm is used to extract feature points. In order to improve the efficiency of the algorithm, the present invention combines the KD-Tree structure to accelerate the neighborhood search and adopts an adaptive neighborhood radius adjustment strategy. In complex areas such as transmission towers, a larger neighborhood radius is used to capture details; in relatively smooth areas such as conductors, a smaller radius is used. Such an optimized ISS algorithm can adapt to the complex geometric structures of different areas in the transmission line point cloud, and significantly improve the speed and accuracy of feature extraction.

[0125] S4, Direction Histogram (PFH) feature description: Based on the feature point set, the direction histogram feature description model is constructed in combination with the geometric topology information to generate a geometric topology model; the geometric topology model captures the spatial distribution direction of the feature points in the feature point set within the neighborhood, reflects the local structural characteristics of the point cloud, and provides stable description features for subsequent registration. Specifically, it includes the following:

[0126] S4.1. After the local neighborhood points are determined in the previous step, for each feature point P and its local neighborhood point cloud set N(P), the direction vector of the neighborhood point relative to the feature point P is calculated.

[0127] S4.2, let the neighborhood point be Q i ∈N(P), direction vector The calculation method is That is, the direction vector is obtained by subtracting the coordinates of the feature point from the coordinates of the neighborhood point. In this way, a set of vectors reflecting the spatial distribution direction of the neighborhood points relative to the feature points is obtained.

[0128] S4.3, dividing the vector set according to the direction space; in this embodiment, considering the structural characteristics of the power transmission line and the data accuracy requirements, the 360° direction space is divided into 24 intervals, each interval is 15°. Then, the number of direction vectors of the neighborhood points falling in each interval is counted to construct a histogram.

[0129] S4.4, in order to further combine the geometric topological information and fully reflect the geometric structure in the complex scene, a distance threshold d is set. threshold= 0.2 m and angle threshold θ threshol =30°. If the distance between two neighboring points is less than d threshold And the angle between them and the feature point is less than θ threshold , then it is considered that there is a connected relationship between the two points; then the connected point pairs in the neighborhood are identified, and then the geometric topological model is established. Finally, the local topological information of the point cloud is quantified as a feature descriptor.

[0130] For example, for a neighborhood point Q and a feature point P, the distance d and the angle θ between them are calculated, and these distance and angle information are encoded. The distance and angle are respectively mapped to a fixed numerical interval and combined into a new numerical value as the geometric topological descriptor of the neighborhood point.

[0131] In the present invention, the feature description of the directional histogram PFH mainly includes three steps: selection of neighborhood points, calculation of directional vectors, and histogram quantization. The local geometric form is reflected by capturing the spatial distribution direction of points in the neighborhood. However, the spatial distribution of points described by the directional histogram alone is not enough to fully reflect the geometric structure in complex scenes, especially in complex areas such as power grid transmission line tower connection nodes and wire intersections. For this reason, the present invention combines geometric topological information and establishes a geometric topological model in the neighborhood by analyzing the connection relationship between feature points and neighborhood points. The model reflects the linear and branching characteristics of the local structure, which is particularly evident at the wire connection points and tower nodes. For example, in the transmission line point cloud, for the connection points of the wires, there is an obvious connectivity relationship between the points in its neighborhood, and this connectivity can be determined by analyzing information such as the distance and angle between the points.

[0132] Then the local topological information of the point cloud is quantified into feature descriptors as a supplement to the directional histogram. For each point in the neighborhood, the connectivity between it and the feature point is calculated, including distance, angle distribution, etc., and this information is encoded as a geometric topological descriptor. It is used to reflect the local topological structure in the neighborhood, such as straight lines, curves, or branches, so as to further enrich the description of feature points.

[0133] S5. Improved RANSAC rough registration: Use the improved RANSAC algorithm to roughly register the feature point sets of the source point cloud set and the target point cloud set, select non-collinear feature points that meet the geometric relationship, and optimize the rigid body transformation matrix. By minimizing the matching error, the initial alignment matrix from the source point cloud set to the target point cloud set is obtained. The specific steps are as follows:

[0134] S5.1. Randomly select 4 corresponding data points from the source point cloud set P and the target point cloud set Q, denoted as p1, p2, p3, p4 and q1, q2, q3, q4 respectively;

[0135] S5.2. Solve the rigid body transformation matrix H by the least squares method, so that the source point cloud is transformed by the rigid body rotation and translation to obtain a new point cloud set P`=R c P+t c , where R c is the rotation matrix, t c is the translation vector;

[0136] S5.3, calculating the distance projection error between the new source point cloud set PP' of the source point cloud set P under the rigid body transformation matrix H and the corresponding points of the target point cloud set Q;

[0137] S5.4, if the distance projection error of the point pair is less than the set threshold error d t (d t =d 2 (p` i ,Hp i ), where d is the distance between corresponding points, then add the point to the sample internal point set I and calculate the matching score S i , that is, the ratio of the internal points to the total number of point clouds;

[0138] S5.5. Set the number of cycles N and the matching score threshold S min , if the matching score of the point cloud is less than S min , then jump out of the loop, otherwise repeat the above steps until the number of iterations reaches N;

[0139] S5.6. Count the number of sample internal points under different rigid body transformation matrices, take the rigid body transformation matrix with the largest number of internal points as the best rigid body transformation matrix, optimize the rigid body transformation matrix, and obtain the source point cloud after coarse alignment.

[0140] The point cloud data after the previous feature description has met the requirements of RANSAC rough registration for point cloud data. The present invention adopts a RANSAC-based point cloud rough registration method, which randomly samples the source point cloud, finds the corresponding transformation model, and tests each random transformation model, and continuously cycles the process until the optimal transformation model is selected as the final result. In order to improve the robustness and operating efficiency of the algorithm, the present invention has made improvements in the following aspects:

[0141] In the selection of corresponding points, 4 corresponding point pairs are selected, and it is ensured that these 4 points are non-collinear. This corresponding point selection method can more comprehensively reflect the transformation relationship between point clouds, and compared with the traditional selection method, it improves the accuracy and stability of registration. Then the KD-Tree method is used to search for the nearest neighbor point. By dividing the point cloud data according to certain rules, the point near the target point can be quickly located, which greatly speeds up the search speed, thereby improving the operating efficiency of the entire coarse registration algorithm.

[0142] S6. Improved ICP fine registration: Based on the rough registration results, the improved ICP algorithm is used for fine registration; the weighting factor is introduced by using the normal vector feature and combined with KD-Tree accelerated search, and the rigid body transformation matrix is ​​gradually optimized to improve the registration accuracy, and finally the precise registration results of the source point cloud set and the target point cloud set are output. The specific steps are as follows:

[0143] S6.1. Input source point cloud set P and target point cloud set Q to be registered;

[0144] S6.2. Establish KD-Tree indexes for the source point cloud set P and the target point cloud set Q respectively. For each point p (taking the source point cloud set as an example), search the neighborhood point set N(P). Then, construct the covariance matrix C:

[0145]

[0146] Among them, q is a point in the neighborhood, and the eigenvector corresponding to the minimum eigenvalue of the covariance matrix is ​​the normal vector of point p;

[0147] S6.3. Using the normal vector features of the source point cloud and the target point cloud, a weighted value is assigned to each pair of matching points in the registration process (to control the influence of each pair of points during registration and increase the weight of the normal vector close point pairs). Based on the normal vector information, a weighting factor ω is defined to optimize the matching. Assume and is the normal vector of the corresponding point pair, and the weight ω can be expressed as:

[0148]

[0149] Here, σ represents the smoothing parameter.

[0150] S6.4. Construct a descriptor based on each point and its normal information to enhance the uniqueness of each feature point.

[0151] In order to enhance the robustness of point pair matching, we first calculate the covariance matrix M of a certain point with the help of the point cloud normal. i , let n be the number of neighboring points of a point, and the components of the normal vector of the point in the x, y, and z directions are n respectively x ,n y ,n z , the covariance matrix M i The calculation formula is:

[0152]

[0153] Calculate the weighted function f at this point i :

[0154] f i =10detM i -0.2(traceMi ) 2

[0155] Using KD-Tree for neighborhood search, the search condition for the nearest point is changed to the characteristic distance equation:

[0156]

[0157] where c i is the descriptor of the point cloud, and the feature distance filtering of the matching point pairs can also be performed based on this formula.

[0158] S6.5. Traverse all points p in the source point cloud set P i , use KD-Tree to search for each p in the target point cloud set Q i The nearest neighbor point q i , forming a matching point pair (p i ,q i )’s set s;

[0159] S6.6. Calculate the average distance d between matching points in the matching point set s m With standard deviation σ, if a point is d si The distance satisfies the 3σ criterion, d si <d m +3σ, then the matching point pair is retained, otherwise it is removed;

[0160] S6.7, according to d si Sort the matching point set s from small to large and select the top 70% matching point pairs (p i ,q i ) coordinates, and solve the rotation matrix R and translation vector t by weighted least squares method;

[0161] S6.8, use the SVD method to solve the rigid transformation matrix [RT], satisfying q i ≈R·p i +t;

[0162] S6.9, calculate the registration error evaluation function E k (RT):

[0163]

[0164] Among them, ω i represents the weighting factor; n represents the number of valid matching point pairs; k represents the kth conversion. k Difference D k =E k -E k-1 If the set threshold is met, it is converged, otherwise it continues to iterate.

[0165] On the basis of coarse registration, an improved ICP fine registration method is used to further improve the registration accuracy. By introducing a weighting factor based on point cloud normal features and a new feature descriptor, the robustness of point pair matching is enhanced, and the neighborhood search and registration error evaluation mechanism are optimized to achieve more accurate point cloud registration.

[0166] First, use the point cloud normal to calculate the covariance matrix M of a certain point i , and then calculate the weighted function f at a certain point i , the weighted values ​​of all points are calculated, and these weighted values ​​will be used to adjust the importance of different points in the subsequent fine registration process. Then, according to the distance between the normal vector and the point, the neighborhood normal angle invariance characteristic of the point cloud is calculated, and the feature descriptor of the point is formed with the coordinate value of the point. This method makes up for the defect of using only coordinate features to describe the point cloud, and improves the speed and accuracy of feature calculation and construction. At the same time, combined with the weighting factor f i All points are weighted synchronously, which increases the difference in feature description between feature-rich points and feature-poor points, and further improves the iterative convergence speed.

[0167] The technical solution of the present invention, for large-scale and complex transmission line point cloud data, first reduces data redundancy and retains key geometric features through adaptive octree downsampling. Then, the ISS feature point extraction algorithm is combined with KD-Tree acceleration to dynamically adjust the neighborhood radius and extract stable geometric feature points. In the feature description stage, the directional histogram is combined with geometric topological information to enhance the description ability of the feature points, ensuring good discrimination in complex structures (such as poles, wire connection points, etc.). The rough alignment step uses the improved RANSAC method to achieve rapid alignment of feature points, and combines KD-Tree to optimize search efficiency. In the fine alignment process, the improved ICP algorithm is used to further improve the registration accuracy and robustness of the point cloud through weighted error calculation, dynamic step size control and KD-Tree accelerated search. This method can efficiently and accurately process large-scale transmission line point cloud data, and is particularly suitable for power inspection, three-dimensional modeling and maintenance scenarios.

[0168] The present invention adopts adaptive octree downsampling, dynamic ISS feature point extraction, directional histogram (PFH) feature description combined with geometric topology information, improved RANSAC coarse registration and improved ICP fine registration technology to form an efficient and accurate point cloud registration scheme, which is particularly suitable for 3D modeling and inspection analysis of transmission lines.

[0169] Example 2

[0170] See also Figure 2 , Figure 2The present invention discloses a large-scale point cloud registration system based on power grid transmission lines, which can realize three-dimensional modeling and inspection analysis of power transmission lines, and specifically includes:

[0171] The data acquisition module is used to collect point cloud data of the transmission line from different viewing angles to form a source point cloud set P and a target point cloud set Q;

[0172] A downsampling module is used to perform adaptive octree downsampling on the point cloud data of the source point cloud set P and the target point cloud set Q, respectively, and dynamically adjust the interval of the selected points in different density areas, obtain representative point cloud data from the source point cloud set P to generate a first point cloud data set, and obtain representative point cloud data from the target point cloud set Q to generate a second point cloud data set;

[0173] A feature point extraction module is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively by using an improved ISS algorithm. In the process of extracting point cloud feature points, a KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure;

[0174] The model building module is used to build a directional histogram feature description model based on the feature point set and combined with geometric topological information to generate a geometric topological model; the geometric topological model captures the spatial distribution direction of the feature points in the feature point set within the neighborhood, reflects the local structural characteristics of the point cloud, and provides stable description features for subsequent registration;

[0175] The coarse registration module is used to coarsely register the feature point sets of the source point cloud set and the target point cloud set using the improved RANSAC algorithm, select non-collinear feature points that meet the geometric relationship, and optimize the rigid body transformation matrix; by minimizing the matching error, the initial alignment matrix of the source point cloud set to the target point cloud set is obtained;

[0176] The fine registration module is used to perform fine registration based on the rough registration results using the improved ICP algorithm. It uses the normal vector feature to introduce a weighting factor and combines it with the KD-Tree accelerated search to gradually optimize the rigid body transformation matrix to improve the registration accuracy. Finally, it outputs the precise registration results of the source point cloud set and the target point cloud set.

[0177] In an optional embodiment, a large-scale point cloud registration method based on a power grid transmission line includes: a) collecting point cloud data of the transmission line from different perspectives to form a source point cloud set P and a target point cloud set Q; b) using an adaptive octree to downsample the point cloud data and dynamically adjust the sampling granularity of different density areas; c) using an improved ISS algorithm to extract point cloud feature points, accelerating the neighborhood search through KD-Tree, and adaptively adjusting the neighborhood radius; d) combining geometric topological information to perform directional histogram (PFH) feature description and constructing a geometric topological model of the local structure of the point cloud; e) performing coarse registration based on an improved RANSAC algorithm, selecting 4 non-collinear feature points, and optimizing the transformation model; f) using an improved ICP algorithm for fine registration, and using normal vector features to introduce a weighting factor to optimize the registration accuracy.

[0178] Example 3

[0179] See also Figure 3 , Figure 3 is a schematic diagram of the structure of a large-scale point cloud registration system based on power grid transmission lines disclosed in an embodiment of the present invention. Figure 3 The described device can be applied to power systems, such as for three-dimensional modeling and inspection analysis of transmission lines, and the embodiments of the present invention are not limited thereto.

[0180] like Figure 3 As shown, the device may include a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the electronic device implements the steps of the method described in the above embodiment and can achieve the technical effect consistent with the above method.

[0181] The memory may include computer system readable media in the form of volatile memory, such as random access memory (RAM) and / or cache memory. The device may further include other removable / non-removable, volatile / non-volatile computer system storage media. By way of example only, the memory may be used to read and write non-removable, non-volatile magnetic media (commonly referred to as a "hard drive"). A program / utility having a set (at least one) of program modules may be stored in, for example, the memory, such program modules including but not limited to an operating system, one or more application programs, other program modules, and program data, each of which or some combination may include the implementation of a network environment. The program modules typically perform the functions and / or methods in the embodiments described herein.

[0182] The processor executes various functional applications and data processing by running the program stored in the memory, such as implementing the method provided in the first embodiment of the present invention.

[0183] Example 4

[0184] Embodiment 4 of the present invention further provides a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the method described in the above embodiment are implemented and the technical effect consistent with the above method can be achieved.

[0185] The computer storage medium of the embodiment of the present invention can adopt any combination of one or more computer-readable media. The computer-readable medium can be a computer-readable signal medium or a computer-readable storage medium. The computer-readable storage medium can be, for example, but not limited to, a system, device or device of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination of the above. More specific examples (non-exhaustive list) of computer-readable storage media include: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In this document, a computer-readable storage medium can be any tangible medium containing or storing a program, which can be used by an instruction execution system, a device or a device or used in combination with it.

[0186] Computer-readable signal media may include data signals propagated in baseband or as part of a carrier wave, which carry computer-readable program code. Such propagated data signals may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than a computer-readable storage medium, which may send, propagate, or transmit a program for use by or in conjunction with an instruction execution system, apparatus, or device.

[0187] The program code embodied on the computer readable medium may be transmitted using any appropriate medium, including but not limited to wireless, wireline, optical fiber cable, RF, etc., or any suitable combination of the foregoing.

[0188] Computer program code for performing the operation of the present invention may be written in one or more programming languages ​​or combinations thereof, including object-oriented programming languages ​​such as Java, Smalltalk, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer through any type of network including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0189] Of course, the storage medium containing computer executable instructions provided by an embodiment of the present invention, whose computer executable instructions are not limited to the above method operations, can also execute related operations in the method provided by any embodiment of the present invention.

[0190] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A large-scale point cloud registration method based on power grid transmission lines, characterized in that: include: Collect point cloud data of the transmission line from different perspectives to form a source point cloud set P and a target point cloud set Q; Adaptively downsampling the point cloud data of the source point cloud set P and the target point cloud set Q respectively, and dynamically adjusting the intervals of the selected points in different density areas, obtaining representative point cloud data from the source point cloud set P to generate a first point cloud data set, and obtaining representative point cloud data from the target point cloud set Q to generate a second point cloud data set; The improved ISS algorithm is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively. In the process of extracting point cloud feature points, the KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure. Based on the feature point set, the directional histogram feature description model is constructed in combination with the geometric topology information to generate a geometric topology model; the geometric topology model captures the spatial distribution direction of the feature points in the feature point set within the neighborhood, reflects the local structural characteristics of the point cloud, and provides stable description features for subsequent registration; The improved RANSAC algorithm is used to roughly align the feature point sets of the source point cloud set and the target point cloud set, and the non-collinear feature points that meet the geometric relationship are selected to optimize the rigid body transformation matrix; by minimizing the matching error, the initial alignment matrix from the source point cloud set to the target point cloud set is obtained; Based on the rough registration results, the improved ICP algorithm is used for fine registration; use The weighting factor is introduced into the normal vector feature and combined with KD-Tree to accelerate the search, and the rigid body transformation matrix is ​​gradually optimized to improve the registration accuracy. Finally, the accurate registration results of the source point cloud set and the target point cloud set are output.

2. The large-scale point cloud registration method based on power grid transmission lines according to claim 1 is characterized in that: Adaptively downsampling the point cloud data of the source point cloud set and the target point cloud set by octree, and dynamically adjusting the intervals of the selected points in different density areas, acquiring representative point cloud data from the source point cloud set to generate a first point cloud data set, and acquiring representative point cloud data from the target point cloud set to generate a second point cloud data set, including: Initialize the octree: Determine the three-dimensional space range occupied by the point cloud data in the source point cloud set P and the target point cloud set Q of the transmission line, which is defined as [x min ,x max ,y min ,y max ,z min ,z max ]; Divide octree nodes: For the root node of the octree, use the following formula to recursively divide it into eight child nodes, taking the x direction as an example, and the y and z directions are similar: Among them, i = 0, 1, 2, 3, 4, 5, 6, 7, representing the index of the eight child nodes in the x direction; Determine adaptive subdivision: Calculate the point cloud density ρ in each child node, the formula is as follows: Where N represents the number of points in the node; V represents the spatial volume of the node. For a rectangular node, its volume is: V=(x max -x min )(y max -y min )(z max -z min ) Set the density threshold of the point cloud data in the transmission tower area and the density threshold of the point cloud data in the open area. When ρ>ρ threshold When , the node is further subdivided; wherein, the density threshold of the point cloud data in the transmission tower area is greater than the density threshold of the point cloud data in the open area; Downsampling to select representative points: After completing the construction of the octree, select representative points from each node; Through the above steps, a part of representative points can be selected from the original large-scale point cloud data to complete the adaptive octree downsampling.

3. The large-scale point cloud registration method based on power grid transmission lines according to claim 2 is characterized in that: After the octree is constructed, select representative points from each node, including: Using the node centroid method, for a point set {P i (x i ,y i ,z i )}, center of gravity P center The calculation formula is as follows: Where N represents the number of nodes inside the node.

4. The large-scale point cloud registration method based on power grid transmission lines according to claim 1 is characterized in that: The improved ISS algorithm is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively. In the process of extracting point cloud feature points, the KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure, including: The local neighborhood is determined using the KD-Tree structure. For each point P = (x, y, z) in the point cloud data, its local neighborhood point cloud set N(P) within the radius r is determined. For each point P in the local neighborhood point cloud set N(P) q , establish a local coordinate system, and set the search radius, where the neighborhood search radius of the point cloud near the transmission tower is set to be larger than the neighborhood search radius of the point cloud in the smooth area of ​​the transmission line; find all points within the search radius and calculate the weight of each point. The formula is: In the formula, ω i Represents point p q The weight is inversely proportional to the number of points in its local neighborhood; r represents the search radius; p i Represents points in the neighborhood. The sparser the points in the sampling, the higher the weight; Calculate the three-dimensional weighted covariance matrix for each point using the formula: Solve the eigenvalues ​​of each point based on the three-dimensional weighted covariance matrix of each point The three eigenvalues ​​are arranged from large to small. The relationship between them reflects the main features of the data points and their relative importance. Two thresholds ε1 and ε2 are set for the point cloud with complex structure near the transmission tower and the point cloud in the smooth area of ​​the transmission line. If the query point p q As an ISS feature point, it must meet 5. The large-scale point cloud registration method based on power grid transmission lines according to claim 1 is characterized in that: Based on the feature point set and combined with geometric topological information, the directional histogram feature description is performed to construct a geometric topological model that reflects the local structural characteristics of the point cloud, including: For each feature point P and its local neighborhood point cloud set N(P), calculate the direction vector of the neighborhood point relative to the feature point P; Let the neighborhood point be Q i ∈N(P), direction vector The calculation method is That is, the direction vector is obtained by subtracting the coordinates of the feature point from the coordinates of the neighborhood point, thereby obtaining a set of vectors reflecting the spatial distribution direction of the neighborhood points relative to the feature points; Divide the vector set according to the direction space, count the number of direction vectors of the neighborhood points falling in each interval, and construct a histogram; Set the distance threshold d threshold and the angle threshold θ threshold , if the distance between two neighboring points is less than d threshold And the angle between them and the feature point is less than θ threshold , it is considered that there is a connectivity relationship between the two points, and the connected point pairs in the neighborhood are identified, and then a geometric topological model is established to quantify the local topological information of the point cloud into feature descriptors.

6. The large-scale point cloud registration method based on power grid transmission lines according to claim 1 is characterized in that: The improved RANSAC algorithm is used to roughly align the feature point sets of the source point cloud set and the target point cloud set, and the non-collinear feature points that meet the geometric relationship are selected to optimize the rigid body transformation matrix. By minimizing the matching error, the initial alignment matrix from the source point cloud to the target point cloud is obtained, including: Randomly select 4 corresponding data points from the source point cloud set P and the target point cloud set Q, denoted as p1, p2, p3, p4 and q1, q2, q3, q4 respectively; The rigid body transformation matrix H is solved by the least squares method, so that the source point cloud is transformed by the rigid body rotation and translation to obtain a new point cloud set P`=R c P+t c , where R c is the rotation matrix, t c is the translation vector; Calculate the distance projection error between the new source point cloud set PP' and the corresponding points of the target point cloud set Q under the rigid body transformation matrix H of the source point cloud set P; If the distance projection error of the point pair is less than the set threshold error d t (d t =d 2 (p` i ,Hp i ), where d is the distance between corresponding points, then add the point to the sample internal point set I and calculate the matching score S i , that is, the ratio of the internal points to the total number of point clouds; Set the number of loops N and the matching score threshold S min , if the matching score of the point cloud is less than S min , then jump out of the loop, otherwise repeat the above steps until the number of iterations reaches N; The number of sample inliers under different rigid body transformation matrices is counted, and the rigid body transformation matrix with the largest number of inliers is taken as the optimal rigid body transformation matrix. The rigid body transformation matrix is ​​optimized to obtain the source point cloud after coarse registration.

7. The large-scale point cloud registration method based on power grid transmission lines according to claim 1 is characterized in that: Based on the rough registration results, the improved ICP algorithm is used for fine registration. The weighting factor is introduced by using the normal vector feature and combined with KD-Tree accelerated search, and the rigid body transformation matrix is ​​gradually optimized to improve the registration accuracy. Finally, the precise registration results of the source point cloud set and the target point cloud set are output, including: Input source point cloud set P and target point cloud set Q to be registered; Establish KD-Tree indexes for the source point cloud set P and the target point cloud set Q respectively; taking the source point cloud set as an example, for each point p, search the neighborhood point set N(P) and construct the covariance matrix C: Among them, q is a point in the neighborhood, and the eigenvector corresponding to the minimum eigenvalue of the covariance matrix is ​​the normal vector of point p; Using the normal vector features of the source point cloud set and the target point cloud set, a weighted value is assigned to each pair of matching points in the registration process. According to the normal vector information, a weighting factor ω is defined to optimize the matching. Assume and is the normal vector of the corresponding point pair, and the weight ω can be expressed as: Where σ represents the smoothing parameter; Construct a descriptor based on each point and its normal information to enhance the uniqueness of each feature point; Traverse all points p in the source point cloud set P i , use KD-Tree to search for each p in the target point cloud set Q i The nearest neighbor point q i , forming a matching point pair (p i ,q i )’s set s; Calculate the average distance d between matching points in the matching point set s m With standard deviation σ, if a point is d si The distance satisfies the 3σ criterion, d si <d m +3σ, then the matching point pair is retained, otherwise it is removed; According to d si Sort the matching point set s from small to large and select the top 70% matching point pairs (p i ,q i ) coordinates, and solve the rotation matrix R and translation vector t by weighted least squares method; The rigid transformation matrix [RT] is obtained by using the SVD method, satisfying q i ≈R·p i +t; Calculate the registration error evaluation function E k (RT): Among them, ω i represents the weighting factor; n represents the number of valid matching point pairs; k represents the kth conversion. k Difference D k =E k -E k-1 If the set threshold is met, it is converged, otherwise it continues to iterate.

8. A large-scale point cloud registration system based on power grid transmission lines, characterized in that: include: The data acquisition module is used to collect point cloud data of the transmission line from different viewing angles to form a source point cloud set P and a target point cloud set Q; A downsampling module is used to perform adaptive octree downsampling on the point cloud data of the source point cloud set P and the target point cloud set Q, respectively, and dynamically adjust the interval of the selected points in different density areas, obtain representative point cloud data from the source point cloud set P to generate a first point cloud data set, and obtain representative point cloud data from the target point cloud set Q to generate a second point cloud data set; A feature point extraction module is used to extract point cloud feature points from the first point cloud data set and the second point cloud data set respectively by using an improved ISS algorithm. In the process of extracting point cloud feature points, a KD-Tree structure is used to accelerate the neighborhood search, and the neighborhood search radius is adaptively adjusted according to the local geometric characteristics of the point cloud, so as to obtain a feature point set that adapts to the complex scene structure; The model building module is used to build a directional histogram feature description model based on the feature point set and combined with geometric topological information to generate a geometric topological model; the geometric topological model captures the spatial distribution direction of the feature points in the feature point set within the neighborhood, reflects the local structural characteristics of the point cloud, and provides stable description features for subsequent registration; The coarse registration module is used to coarsely register the feature point sets of the source point cloud set and the target point cloud set using the improved RANSAC algorithm, select non-collinear feature points that meet the geometric relationship, and optimize the rigid body transformation matrix; by minimizing the matching error, the initial alignment matrix of the source point cloud set to the target point cloud set is obtained; The fine registration module is used to perform fine registration based on the rough registration result using the improved ICP algorithm; The weighting factor is introduced by using the normal vector feature and combined with KD-Tree to accelerate the search, and the rigid body transformation matrix is ​​gradually optimized to improve the registration accuracy. Finally, the accurate registration results of the source point cloud set and the target point cloud set are output.

9. A large-scale point cloud registration device based on power grid transmission lines, characterized in that: The electronic device comprises a processor and a memory, wherein the memory stores computer instructions, and the processor is used to execute the computer instructions stored in the memory. When the computer instructions are executed by the processor, the electronic device implements the steps of the large-scale point cloud registration method based on power grid transmission lines as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of the large-scale point cloud registration method based on power grid transmission lines as claimed in any one of claims 1 to 7.

Images