A power tower reconstruction method based on ground-based LiDAR point cloud

Abstract

The application discloses a power tower reconstruction method based on ground LiDAR point cloud, comprising the following steps: step 1, the spatial and morphological characteristics of the tower ground LiDAR point cloud are comprehensively analyzed from the elevation, corner point, edge point, flat point and component form multi-angle; step 2, on the basis of the tower point cloud characteristic analysis of step 1, the nodes and edge lines on the tower main material point cloud are extracted; step 3, according to the results of step 1 and step 2, the tower component and rod point cloud are segmented and extracted; step 4, according to the segmented rod point cloud in step 3, the rod model and point cloud are registered, and the tower model reconstruction work is indirectly completed. On the basis of obtaining high-precision ground LiDAR point cloud data, based on the tower point cloud structure characteristics and point cloud processing technology, the model reconstruction method based on the tower point cloud segmentation is proposed, the details of the rod can be constructed, and through comparison with the modeling software, it is proved that the application has certain effectiveness.

Description

Technical Field

[0001] This invention relates to the field of power transmission line engineering technology, and in particular to a method for reconstructing power poles based on ground-based LiDAR point clouds. Background Technology

[0002] In the era of intelligent power grid inspection, towers are subject to various challenges. Construction and assembly errors result in towers that do not perfectly match the design drawings, and long-term exposure to the elements leads to deformation or tilting due to natural erosion and human damage. Furthermore, the diverse types of towers, the large number of parts, and the complexity of their local structures have become major factors hindering the improvement of 3D modeling for the entire transmission line. While measuring and detecting important parameters such as tower height, dimensions, and attachment point locations only requires a wireframe model, tasks like simulated installation, component collision verification, and automatic analysis of tower defects require high-precision and detailed surface or solid models.

[0003] Currently, point cloud data processing and information extraction methods for power equipment mainly focus on foundations, fittings, insulator strings, and conductors / ground wires. Research on point cloud processing methods for poles and towers, which are more diverse, have more parts, and exhibit more complex structures, is relatively lacking, and sufficient feature analysis and information extraction of the pole and tower point clouds themselves are not performed. Currently, 3D reconstruction methods based on pole and tower point clouds can be mainly divided into three categories: model-driven, data-driven, and hybrid-driven methods. In model-driven pole and tower 3D reconstruction, the main reliance is on matching, identifying, and stitching core component model libraries, without sufficient precision in matching and identifying component model libraries, resulting in high model library construction costs and insufficient flexibility. In data-driven and hybrid-driven pole and tower 3D reconstruction, algorithms focus on key point identification, feature line extraction, and pole and tower type identification, lacking a focus on pole and tower image or point cloud segmentation processing, resulting in mostly simplified single-line models in automatic or semi-automatic pole and tower model construction.

[0004] While CAD design drawings can create high-precision and detailed pole models, they cannot reveal the actual condition of poles in use. Pole models reconstructed from UAV imagery often suffer from numerous misalignments due to shadows. Airborne LiDAR, while acquiring point clouds with large spatial point spacing and blurred details of individual poles, is often used for constructing single-line pole models. Ground-based LiDAR, though less portable, is more effective at acquiring high-precision point cloud data for tall, small-footprint man-made structures like poles, facilitating detailed 3D reconstruction. Summary of the Invention

[0005] The technical problem to be solved by this invention is to provide a method for reconstructing power poles based on ground-based LiDAR point clouds, which can solve the problems of low precision in pole model reconstruction and high cost of building model library.

[0006] To address the aforementioned technical problems, this invention provides a method for reconstructing power poles based on ground-based LiDAR point clouds, comprising the following steps:

[0007] Step 1: Conduct a comprehensive analysis of the spatial and morphological characteristics of the LiDAR point cloud of the tower foundation from multiple perspectives, including elevation, corner points, edge points, flat points, and component morphology.

[0008] Step 2: Based on the feature analysis of the tower point cloud in Step 1, extract the nodes and edges on the point cloud of the main tower material;

[0009] Step 3: Based on the results of Step 1 and Step 2, segment and extract the point cloud of tower components and poles;

[0010] Step 4: Based on the segmented point cloud of the pole members in Step 3, register the pole member model and the point cloud to indirectly complete the reconstruction of the pole tower model.

[0011] Preferably, step 1 involves a comprehensive analysis of the spatial and morphological characteristics of the LiDAR point cloud of the tower foundation from multiple perspectives, including elevation, corner points, edge points, flat points, and component morphology. This analysis specifically includes the following steps:

[0012] Step 11: Use the point cloud slicing algorithm for analysis. That is, divide the tower point cloud into equal-interval longitudinal point cloud slices at a certain interval from bottom to top, count the number of points in each slice, and analyze the distribution on the longitudinal section at each elevation. Based on the point cloud slicing, use the Euclidean clustering algorithm on each slice to analyze the distribution on the horizontal plane at each elevation.

[0013] Step 12: Extract corner and edge points of the tower point cloud using the 3D Harris keypoint extraction algorithm and the fast 3D line segment detection algorithm. Combine the tower structure analysis to determine the distribution of corner and edge points. The flat points in the tower point cloud are basically distributed between two corner points. Analyze the distribution characteristics of flat points using cylinder connectivity analysis. The specific steps are: ① Use the k-nearest neighbor algorithm to fuse corner points near the same node. Calculate the average coordinates of multiple corner points at a connection point using the k-nearest neighbor algorithm to fuse them; ② Using the two corner points (x... i y i , z i ), (x j y j , z j1) Fit a straight line according to the two-point equation of the spatial line, where i and j are distinct positive integers from 1 to the number of corner points; 2) Sample 5 equally spaced spatial points on the spatial line; 3) Use the k-nearest neighbor algorithm again to count the number of spatial points within a certain distance of the point cloud data. If the number is greater than the threshold, it is determined that there is a flat point between the two corner points, and the spatial coordinate index of the two corner points is saved; 4) Repeat steps 2, 3, and 4 until there are no other combinations of two different corner points.

[0014] Step 13: Manually segment out the point clouds of some rods and components, and calculate their spatial morphological characteristics using the spatial characteristic formula of the covariance matrix, including linear index, surface index, spherical index, anisotropy, total variance and roughness.

[0015] Preferably, step 2, extracting nodes and edges from the point cloud of the main tower material, specifically includes the following steps:

[0016] Step 21: Point cloud extraction at connection points; the specific steps are as follows: ① Based on the characteristic of the regular distribution of tower components in different planes, perform cross-sectional processing in the vertical direction; ② Perform coarse segmentation based on the clustering characteristics of point clouds in different cross sections to extract point clouds of transverse diaphragms and point cloud data of only the connection points of diagonal and main members; ③ Based on the peak value of the number of points within the transverse diaphragm point cloud slice in the statistical graph, classify the two types of point cloud data of transverse diaphragms and only the connection points of diagonal and main members; ④ Classify tower point clouds and other transverse diaphragm point clouds based on the linear and planar characteristics of the transverse diaphragm point clouds; ⑤ Use the average height of the tower point clouds as the height threshold, those greater than the threshold are tower head connection point clouds, and those less than the threshold are tower body connection point clouds.

[0017] Step 22, Node Extraction; the specific steps are as follows: ① First, fit the point cloud at the connection point into two intersecting planes using the RANSAC plane fitting algorithm, a1x+b1y+c1z+d1=0 and a2x+b2y+c2z+d2=0; ② Then, calculate the spatial vector of the intersecting line ((b1c2-b2c1), (a2c1-a1c2), (a1b2-a2b1)) by the cross product of the direction vectors (a1, b1, c1) and (a2, b2, c2) in the two planes, and find any point on the intersecting line. Thus, the spatial straight line is determined. ③Then calculate the spatial points (x) in the point cloud. i y i , z i ) The distance from i = 1...n to the spatial line; ④ The spatial points that meet a certain threshold are the edge point cloud. The centroid of the edge, i.e. the key point, is obtained by calculating the average value of all points in the edge point cloud; ⑤ Repeat steps ① to ④ until the key points of all connection point clouds are extracted.

[0018] Step 23, Auxiliary line extraction; the specific steps are as follows: ① Cluster all the key points obtained above using the elevation z of the spatial points within the key points as the criterion, calculate the average elevation of the spatial points that are clustered into a cluster, and use the average elevation value as the new elevation of these spatial points; ② Then, cluster all the key points again using the plane coordinates (x, y) of these spatial points as the criterion, and use the RANSAC line fitting algorithm to fit the spatial lines of the spatial points that are clustered into a cluster.

[0019] Preferably, step 3, which involves registering the pole model and the point cloud to indirectly complete the reconstruction of the tower model, specifically includes the following steps:

[0020] Step 31: Extraction of component point cloud; the specific steps are as follows: ① Estimate the elevation value at the boundary between the tower head and the tower body by using the average elevation of the point cloud data of the cross diaphragm at this location; ② Calculate the distance from the spatial points in the tower point cloud to the auxiliary line, and count the spatial points that meet certain distance and elevation thresholds, which are the main component point clouds; ③ Perform RANSAC straight line segmentation on the main component point cloud to remove noise points from the point clouds of other members connected to the main material; ④ Segment the remaining point cloud using Euclidean clustering algorithm, and classify the crossarm and tower plate point clouds after removing the horizontal plane according to the roughness.

[0021] Step 32: Extraction of point clouds from tower plate members; the specific steps are as follows: ① Perform RANSAC plane fitting on the point clouds of tower plate members to obtain a spatial plane. ② Then project the spatial points (xi, yi, zi) in the point clouds of tower plate members onto this plane to obtain new spatial points (Xi, Yi, Zi), i = 1...n; ③ After the above processing, the diagonal member point cloud can be segmented from the new spatial points using the RANSAC line fitting algorithm; ④ Repeat steps ① to ③ until the diagonal member point cloud of all tower plate members is extracted; ⑤ Obtain the point cloud of the lacing material by using the point cloud difference algorithm on the point clouds of tower plate members and diagonal member point clouds; ⑥ Obtain the point cloud of a single lacing member by using Euclidean clustering on the point cloud of the lacing material; ⑦ Repeat steps ⑤ to ⑥ until the lacing material of all tower plates is segmented.

[0022] Step 33: Extraction of point clouds from transverse diaphragm members; the specific steps are as follows: ① Extract the point clouds of individual members on the transverse diaphragms other than the crossarms through RANSAC plane fitting, point cloud projection onto the plane, and RANSAC line fitting; ② Calculate the required spatial points for the 3D convex hull using the intersection formula based on the fitted tower head auxiliary line and the highest and lowest elevation surfaces z = -d of the tower point cloud; ③ Segment the point cloud of the middle part using the CropHull intersection algorithm of the 3D convex hull; ④ Process the transverse diaphragm point clouds segmented from the tower point cloud in the same way as the other transverse diaphragm point clouds in Step 1 to segment the point clouds of individual members.

[0023] Step 34: Point cloud segmentation and extraction of horizontal members of the crossarm; the specific steps are as follows: ① Obtain the point cloud of the horizontal crossarm by performing a differential operation on the point cloud of the tower and the cross diaphragm of the segmented middle part; ② Then perform voxel filtering, RANSAC plane fitting and point cloud projection onto the plane preprocessing operation on the point cloud of the horizontal crossarm; ③ Next, use the directed bounding box algorithm to extract the lengths of the first principal direction and the second principal direction of the point cloud of the horizontal crossarm, calculate the ratio of these two lengths, and count the point cloud parts of the members with large and small ratios respectively; ④ For the point cloud of the members with small ratios (obvious planarity), extract the point cloud of each member by RANSAC line fitting; for the point cloud of the members with large ratios, it is necessary to first extract the point cloud of the main member, then perform Euclidean clustering on the remaining point cloud to obtain the tower plate point cloud composed of two diagonal members, and then use the RANSAC line fitting algorithm to obtain the point cloud of a single diagonal member.

[0024] Preferably, step 4, which involves registering the pole model and the point cloud to indirectly complete the reconstruction of the tower model, specifically includes the following steps:

[0025] Step 41: Constructing the surface model of the angle steel member. Set the center of the angle steel edge at the origin of the coordinate system. The edge of the angle steel is on the x-axis, and the two legs of the angle steel are on the z and y axes respectively, and the x, y, and z axes are perpendicular to each other.

[0026] Step 42: Use the Oriented Bounding Box (OBB) algorithm to perform x-axis mapping of the rod point cloud. max y max z max x min y min z min The geometric parameters of the length, width, and height of the bar point cloud are extracted and calculated. BarLength (BL) represents the length of the bar, and Bar Width (BW) represents the width. When segmenting the point cloud of the angle steel double-limb bar, the limb width BW of the bar model is... The width of the rod point cloud is multiplied by a factor of 1. For cases where only a single rod point cloud is segmented, the limb width BW of the rod model is the width of the rod point cloud, and the length BL of the rod model is the length of the rod point cloud. Finally, the vertex coordinates of the corrected predefined model are translated and rotated to the positions of each rod point cloud in the global coordinate system using a rotation and translation matrix from the local coordinate system to the global coordinate system, thereby completing the coarse registration of the predefined rod model and the rod point cloud.

[0027] Step 43: ICP fine registration of the rods in the local coordinate system; the specific steps are as follows: ① Fit the corresponding set of straight lines by the endpoint coordinates of each model rod in the global coordinate system, calculate the distance from the spatial points in the tower point cloud to each straight line, count the spatial points that meet a certain threshold, and obtain the completed rod point cloud; ② Perform a traditional ICP algorithm on the rod point cloud and the model point cloud to complete the fine registration of the rod point cloud in the first principal direction; ③ Then, for the rod model point cloud that failed to register, rotate the model point cloud 180° in the local coordinate system with its first principal direction as the axis, and then perform a fine registration of ICP in the global coordinate system to complete the fine registration of the rod point cloud in the second principal direction; ④ Finally, save the translation and rotation matrices corresponding to the translation and rotation operations performed on each model point cloud during the fine registration process, and perform the same translation and rotation processing on the rod model to complete the fine registration of the rod model and the tower point cloud.

[0028] The beneficial effects of this invention are as follows: Based on the acquisition of high-precision ground-based LiDAR point cloud data, this invention studies a method for three-dimensional reconstruction of tower models based on the structural characteristics of tower point clouds and point cloud processing technology. Addressing existing theories on the structural characteristics of tower point clouds, this invention proposes a model reconstruction method based on tower point cloud segmentation. This method can construct detailed parts of the tower components. Comparison with modeling software demonstrates the effectiveness of this invention. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the method flow of the present invention.

[0030] Figure 2 This is a preprocessed point cloud image of the tower in this invention.

[0031] Figure 3(a) is a segmentation result diagram of the point cloud of the main material components of the tower body and tower head of the present invention.

[0032] Figure 3(b) is a diagram showing the segmentation result of the oblique point cloud of the tower body and tower head of the present invention.

[0033] Figure 3(c) is a segmentation result diagram of the point cloud of the lacing members of the tower body and tower head of the present invention.

[0034] Figure 4(a) shows the point cloud segmentation result of the crossbeam members excluding the horizontal plane in this invention.

[0035] Figure 4(b) shows the point cloud segmentation result of the inclined members excluding the horizontal plane in this invention.

[0036] Figure 5(a) shows the point cloud segmentation results of individual members of the tower and other transverse diaphragms in this invention.

[0037] Figure 5(b) shows the point cloud segmentation results of individual members of the tower and other transverse diaphragms in this invention.

[0038] Figure 5(c) shows the point cloud segmentation result of a single diagonal member of the tower and other transverse diaphragms in this invention.

[0039] Figure 6 This is a schematic diagram of the tower model reconstruction of the present invention. Detailed Implementation

[0040] like Figure 1 As shown, a method for reconstructing power poles based on ground-based LiDAR point clouds includes the following steps:

[0041] Step 1: Conduct a comprehensive analysis of the spatial and morphological characteristics of the LiDAR point cloud of the tower foundation from multiple perspectives, including elevation, corner points, edge points, flat points, and component morphology.

[0042] Step 2: Based on the feature analysis of the tower point cloud in Step 1, extract the nodes and edges on the point cloud of the main tower material;

[0043] Step 3: Based on the results of Step 1 and Step 2, segment and extract the point cloud of tower components and poles;

[0044] Step 4: Based on the segmented point cloud of the pole members in Step 3, register the pole member model and the point cloud to indirectly complete the reconstruction of the pole tower model.

[0045] Taking a power tower located in eastern Nanjing as an example, point cloud data of the tower was collected using a RIEGL VZ-400i terrestrial 3D laser scanner, and preprocessed with point cloud denoising and smoothing. The results are as follows: Figure 2 As shown. To solve the above technical problems, this invention provides a method for reconstructing a 3D model of a power pole based on LiDAR point clouds, including the following steps:

[0046] (1) First, the elevation distribution characteristics of the tower point cloud were analyzed using a point cloud slicing algorithm. The tower point cloud was sliced ​​longitudinally at equal intervals from bottom to top, and the number of points in each slice was counted. It was found that the number of points in the slices at the transverse diaphragms was at the peak of the entire tower slice, with the peak being more obvious at the tower. Based on the completed point cloud slicing, Euclidean clustering algorithm was applied to each slice. It was found that the point clouds at the transverse diaphragms in the slice basically clustered into one block, while the number of clusters at the connection between the members was basically consistent with the number of longitudinal main members.

[0047] Then, the corner and edge points of the tower point cloud were extracted using the 3D Harris key point extraction algorithm and the fast 3D line segment detection algorithm. The corner points are distributed near the joints of the members, and the corner points at approximately the same height are axially and centrally symmetrical, with a many-to-one relationship between the corner points and the connecting plates. From the distribution of edge points, the entire tower is basically composed of straight lines, and the edge lines of the angle steel tower members have a many-to-one relationship with the members.

[0048] The flat points in the tower point cloud are basically distributed between two corner points. This paper analyzes the distribution characteristics of flat points through the connectivity of cylinders. The specific steps are: ① Use the k-nearest neighbor algorithm to fuse corner points near the same node, that is, calculate the average value of the coordinates of multiple corner points at a connection point and then fuse them. ② Then, through the two corner points (x i y i , z i ), (x j y j , z j Fit a straight line using the two-point form of the equation of a straight line in space:

[0049]

[0050] Where i and j are distinct positive integers from 1 to the number of corner points. ③ Sample 5 equally spaced spatial points on the spatial straight line. ④ Use the k-nearest neighbor algorithm again to count the number of spatial points within a certain distance of the point cloud data. If it is greater than the threshold, it is determined that there is a flat point between the two corner points, and the spatial coordinate index of the two corner points is saved. ⑤ Repeat steps 2, 3, and 4 until there are no other combinations of two different corner points. From the distribution of flat points, the rod point cloud is distributed in a hollow shape, and there are distributions with the same direction and intersections between the individual rod point clouds.

[0051] Finally, the point clouds of some rods and components were manually segmented, and their spatial morphological characteristics were calculated using the following formulas for spatial characteristics of the covariance matrix, including linearity index, surface index, spherical index, anisotropy, total variance, and roughness. The main tower members exhibit a rough linear distribution with approximately longitudinal direction and small angles between them. The point clouds of the tower's diagonal members exhibit a smooth linear distribution with intersecting parallel structures. The point clouds of the tower sections have surface or linear characteristics, forming smooth planar or linear distributions. The point cloud of the tower platform has a smooth linear distribution, while the point clouds of other transverse diaphragms have a smooth surface distribution. The point clouds of the crossarms on the horizontal plane exhibit both rough surface and rough linear distributions.

[0052]

[0053] (2) To facilitate the acquisition of point clouds of the main tower components, this invention first extracts key points and auxiliary lines from the tower point cloud. First, longitudinal slices are made along the tower point cloud. Euclidean clustering is performed on the point clouds within each slice. The number of Euclidean clusters for the slices at the diaphragm, tower base, diagonal members only, and main component connections is less than or equal to the number of main tower components. The slices are then merged and Euclidean clusters are performed within each slice. The number of point clouds at the main component connections is much smaller than that at the diaphragm and tower base. The linear characteristics of the tower base are more pronounced than those at the diaphragm, and the point clouds at the connection between the tower head and tower body can be distinguished using the lower chord of the tower base as an elevation threshold.

[0054] Since the point cloud at the connection point can be approximated as two intersecting planes, this invention uses two RANSAC plane fittings of the point cloud at the connection point and the intersection of the fitted planes to obtain the edge point cloud at the connection point. The two fitted planes and the edge are shown in Equation 1 below.

[0055]

[0056] Where (a1, b1, c1) and (a2, b2, c2) are direction vectors in two planes, their cross product yields the spatial vector of the intersecting line ((b1c2-b2c1), (a2c1-a1c2), (a1b2-a2b1)), and any point on the intersecting line is... Points within a certain threshold distance from the edge line constitute the edge point cloud of the connection point cloud. The nodes at the connection between the tower head and tower body are obtained through the centroid of the edge point cloud. Nodes within the same elevation range are coplanar, and their projections onto the horizontal plane form four disjoint linear distributions based on the positions of the four main members. Auxiliary lines are obtained through straight line fitting.

[0057] (3) Based on a thorough analysis of the structural characteristics of the tower point cloud, the point cloud of individual tower components is segmented and extracted. The distance from the spatial points within the tower point cloud to the auxiliary lines is calculated, and spatial points that meet certain distance and elevation thresholds are statistically analyzed. Figure 3(a) shows the point cloud of the main structural member. After completing the segmentation of the point cloud of the transverse diaphragm, the tower platform, and the main structural member, the remaining point cloud is segmented using a Euclidean clustering algorithm. Based on the characteristic that the roughness of the point cloud of the crossarm excluding the horizontal plane is significantly greater than that of the point cloud of the tower plate component composed of inclined and lacing materials, the point clouds of the crossarm excluding the horizontal plane and the tower plate are extracted respectively.

[0058] The spatial distribution of the point cloud of the tower section composed of diagonal and lacing members tends to be linear and has certain planar distribution characteristics, while the point cloud of the diagonal members has very obvious linear characteristics and almost no planar characteristics. We can first use RANSAC plane fitting and then project it onto the plane to obtain new spatial points, thereby increasing the planar characteristics of the component point cloud and reducing the linear characteristics, making the distribution characteristics of the point cloud of the diagonal members more different. This facilitates the segmentation of the diagonal member point cloud as shown in Figure 3(b) using the RANSAC line fitting algorithm. The point cloud of the lacing members is only adjacent to the point cloud of the diagonal members, and the point clouds of the lacing members are in a disjoint state. Therefore, by removing the diagonal members from the other point clouds of the tower section, Euclidean clustering can be used to obtain the point cloud of a single lacing member as shown in Figure 3(c).

[0059] Besides the planar and linear characteristics of the point cloud of the crossarm members on the horizontal plane, the point cloud with the most obvious linear distribution can be segmented after a series of point cloud processing steps, including RANSAC plane fitting, point cloud projection onto the plane, and RANSAC line fitting, as shown in Figure 4(a). By removing the point cloud of the crossarm on the horizontal plane and the planar point cloud, the point cloud of the diagonal crossarm in the front plane is obtained through point cloud differentiation. Then, Euclidean clustering is performed to obtain the point cloud of a single crossarm lacing member. The point cloud of the diagonal crossarm members can be obtained by differentiating the point cloud of the crossarm main member and the planar point cloud. Finally, RANSAC line fitting can be used to segment the point cloud of a single diagonal crossarm member. The segmentation result of the diagonal member is shown in Figure 4(b).

[0060] The tower's point cloud exhibits a linear spatial distribution. The point cloud of the middle section of the tower is the point cloud of the transverse diaphragm, which can be segmented using the CropHull intersection algorithm with a 3D convex hull. The transverse diaphragm point cloud exhibits a planar spatial distribution. The point cloud of a single member, as shown in Figures 5(a) and (b), can be extracted through RANSAC plane fitting, point cloud projection onto the plane, and RANSAC line fitting. The crossarm's horizontal plane point cloud can be obtained by performing a difference operation on the transverse diaphragm point cloud of the tower and its segmented middle section. Then, through voxel filtering, RANSAC plane fitting, and point cloud preprocessing by projecting the point cloud onto the plane, and by using a directed bounding box (OBB) to extract the lengths in the first and second principal directions and calculate the ratio of these two lengths, the planarity of the member point cloud is more pronounced. The member point cloud can be directly extracted using RANSAC line fitting. For point clouds of components with large ratios, it is necessary to first extract the main member point cloud, then perform Euclidean clustering on the remaining point clouds to obtain the tower segment point cloud composed of two diagonal members, and then use the RANSAC linear fitting algorithm to obtain the point cloud of a single diagonal member. The final segmentation result is shown in Figure 5(c). The results of the member point cloud segmentation and extraction method proposed in this paper are statistically analyzed from five dimensions: total number of member point clouds, number of extracted member point clouds, recall rate of member point cloud extraction, precision and efficiency of member point cloud extraction. The formulas for calculating recall rate and precision are as follows:

[0061]

[0062]

[0063] Where RBE represents the recall rate of pole point cloud extraction, NCE represents the number of correctly extracted pole point clouds, NUE represents the number of unextracted pole point clouds, ABE represents the precision of pole point cloud extraction, and NEE represents the number of incorrectly extracted pole point clouds. The statistical results are shown in Table 1.

[0064] Table 1. Statistics of Point Cloud Segmentation Results for Rod Members

[0065]

[0066]

[0067] (4) Based on the segmented member point cloud in step (3), register the member model and the point cloud to indirectly complete the reconstruction of the tower model. The thickness of the angle steel member is not on the same order of magnitude as its length and width, and the thickness is only between 3 and 24 mm, which is negligible. The predefined angle steel member surface model only represents the length and width. The first, second, and third principal directions of the member point cloud are extracted by OBB to estimate the length and width of the two limbs of the angle steel member, and the vertex coordinates of the predefined member model are kept consistent with the endpoint coordinates under the local coordinate system of each member point cloud.

[0068] By using a rotation and translation matrix from the local coordinate system to the global coordinate system, the vertex coordinates of the corrected predefined model are translated and rotated to the positions of the point clouds of each member in the global coordinate system, thus completing the coarse registration of the predefined member model and its point clouds. Using the endpoint coordinates extracted from the local coordinate system via OBB and the determined leg widths of each member, point clouds are sampled uniformly at equal intervals, and then translated and rotated to the global coordinate system to obtain the point cloud data of each member model. The ICP algorithm is then used to perform fine registration of the member models resulting from the coarse registration. For member model point clouds that failed to register, the first principal direction is used as the axis to rotate the model point clouds 180° in the local coordinate system, and then fine registration is performed in the global coordinate system using ICP. The reconstruction results are as follows: Figure 6 As shown, the difference is only 0.044m compared to the corresponding point modeled manually.

Claims

1. A method for reconstructing power poles based on ground-based LiDAR point clouds, characterized in that, Includes the following steps: Step 1: Conduct a comprehensive analysis of the spatial and morphological characteristics of the LiDAR point cloud of the tower foundation from multiple perspectives, including elevation, corner points, edge points, flat points, and component morphology; specifically, this includes the following steps: Step 11: Use the point cloud slicing algorithm for analysis. That is, divide the tower point cloud into equal-interval longitudinal point cloud slices at a certain interval from bottom to top, count the number of points in each slice, and analyze the distribution on the longitudinal section at each elevation. Based on the point cloud slicing, use the Euclidean clustering algorithm on each slice to analyze the distribution on the horizontal plane at each elevation. Step 12: Extract the corner points and edge points of the tower point cloud, and analyze the distribution of corner points and edge points in combination with the tower structure. The flat points of the tower point cloud are basically distributed between two corner points. Analyze the distribution characteristics of flat points through the cylinder connectivity. The specific steps are as follows: ① Use the k-nearest neighbor algorithm to merge corner points near the same node. The k-nearest neighbor algorithm is used to calculate the average coordinates of multiple corner points at a connection point, thus merging them; ② Use the coordinates of two corner points (x... i y i , z i ), (x j y j , z j 1) Fit a straight line according to the two-point equation of the spatial line, where i and j are distinct positive integers from 1 to the number of corner points; 2) Sample 5 equally spaced spatial points on the spatial line; 3) Use the k-nearest neighbor algorithm again to count the number of spatial points in the point cloud data. If the number is greater than the threshold, it is determined that there is a flat point between the two corner points, and the spatial coordinate index of the two corner points is saved; 4) Repeat steps 2, 3, and 4 until there are no other combinations of two different corner points. Step 13: Manually segment out the point cloud of some rods and components, and calculate their spatial morphological characteristics using the spatial characteristic formula of the covariance matrix, including linear index, surface index, spherical index, anisotropy, total variance and roughness; Step 2: Based on the feature analysis of the tower point cloud in Step 1, extract the nodes and edges on the point cloud of the main tower material; Step 3: Based on the results of Step 1 and Step 2, segment and extract the point cloud of tower components and poles; Step 4: Based on the segmented point cloud of the pole members in Step 3, register the pole member model and the point cloud to indirectly complete the reconstruction of the pole tower model.

2. The method for reconstructing power poles based on ground-based LiDAR point clouds as described in claim 1, characterized in that, Step 2, extracting nodes and edges from the point cloud of the main tower material, specifically includes the following steps: Step 21: Point cloud extraction at connection points; the specific steps are as follows: ① Based on the characteristic of the regular distribution of tower components in different planes, perform cross-sectional processing in the vertical direction; ② Perform coarse segmentation based on the clustering characteristics of point clouds in different cross sections to extract point clouds of transverse diaphragms and point cloud data of only the connection points of diagonal and main members; ③ Based on the peak value of the number of points within the transverse diaphragm point cloud slice in the statistical graph, classify the two types of point cloud data of transverse diaphragms and only the connection points of diagonal and main members; ④ Classify tower point clouds and other transverse diaphragm point clouds based on the linear and planar characteristics of the transverse diaphragm point clouds; ⑤ Use the average height of the tower point clouds as the height threshold, those greater than the threshold are tower head connection point clouds, and those less than the threshold are tower body connection point clouds. Step 22: Node extraction; The specific steps are as follows: ① First, fit the point cloud at the connection point into two intersecting planes using the RANSAC plane fitting algorithm, a1x+b1y+c1z+d1=0 and a2x+b2y+c2z+d2=0; ② Then, calculate the spatial vector of the intersecting line ((b1c2-b2c1), (a2c1-a1c2), (a1b2-a2b1)) by the cross product of the direction vectors (a1, b1, c1) and (a2, b2, c2) in the two planes, and find any point on the intersecting line. Thus, the spatial straight line is determined. ③Then calculate the spatial points (x) in the point cloud. i y i , z i ) The distance from i = 1...n to the spatial line; ④ The spatial points that meet the conditions are the edge point cloud. The centroid of the edge, i.e. the key point, is obtained by calculating the average value of all points in the edge point cloud; ⑤ Repeat steps ① to ④ until the key points of all connection point clouds are extracted. Step 23: Extract auxiliary lines; The specific steps are as follows: ① Cluster all the key points obtained above using the elevation z of the spatial points within the key points as the criterion, calculate the average elevation of the spatial points that are clustered into a cluster, and use the average elevation value as the new elevation of these spatial points. ②Then, using the planar coordinates (x, y) of these spatial points as criteria, all key points are clustered, and the RANSAC line fitting algorithm is used to fit the spatial lines within the clusters.

3. The method for reconstructing power poles based on ground-based LiDAR point clouds as described in claim 1, characterized in that, Step 3, the segmentation and extraction of the point cloud of tower components and members, specifically includes the following steps: Step 31: Extraction of component point clouds; The specific steps are as follows: ① Estimate the elevation value at the boundary between the tower head and the tower body by using the average elevation of the point cloud data of the transverse diaphragm; ② Calculate the distance from the spatial points in the tower point cloud to the auxiliary lines, and count the spatial points that meet the conditions, which are the main material component point clouds; ③ Perform RANSAC straight line segmentation on the main material component point cloud to remove noise points from the point clouds of other members connected to the main material; ④ Segment the remaining point cloud using Euclidean clustering algorithm, and classify the point clouds of the crossarm and tower pieces after removing the horizontal plane according to the roughness. Step 32: Extraction of point cloud from tower components; The specific steps are as follows: ① Perform RANSAC plane fitting on the point cloud of the tower components to obtain a spatial plane; ② Project the spatial points (xi, yi, zi) in the point cloud of the tower components onto this plane to obtain new spatial points (Xi, Yi, Zi), i = 1...n; ③ After the above processing, use the RANSAC line fitting algorithm to segment the oblique bar point cloud from the new spatial points; ④ Repeat steps ① to ③ until all tower component point clouds have completed the extraction of oblique bar point clouds; ⑤ Use the point cloud difference algorithm to obtain the point cloud of the lacing material from the point cloud of the tower components and the oblique bar point cloud; ⑥ Use Euclidean clustering to obtain the point cloud of a single lacing bar member from the point cloud of the lacing material; ⑦ Repeat steps ⑤ to ⑥ until the lacing material of all tower components has been segmented. Step 33: Extraction of point cloud from transverse diaphragm members; The specific steps are as follows: ① Extract the point cloud of individual members on the transverse diaphragms other than the crossarm by processing RANSAC plane fitting, point cloud projection onto the plane, and RANSAC line fitting; ② Calculate the required spatial points for the 3D convex hull by using the intersection formula for the highest and lowest elevation surfaces z = -d of the fitted tower head auxiliary line and the tower point cloud; ③ Segment the point cloud of the middle part by using the CropHull intersection algorithm of the 3D convex hull; ④ Process the transverse diaphragm point cloud segmented from the tower point cloud in the same way as the other transverse diaphragm point clouds in step 1 to segment the point cloud of individual members. Step 34: Point cloud segmentation and extraction of horizontal members of the crossarm; The specific steps are as follows: ① Obtain the point cloud of the horizontal plane of the crossarm by performing a differential operation on the point cloud of the tower and the transverse diaphragm of its segmented middle section; ② Then perform voxel filtering, RANSAC plane fitting, and point cloud projection onto the plane preprocessing operation on the point cloud of the horizontal plane of the crossarm; ③ Next, use the directed bounding box algorithm to extract the lengths of the first principal direction and the second principal direction of the point cloud of the horizontal plane of the crossarm, calculate the ratio of these two lengths, and count the component point cloud parts with large and small ratios respectively; ④ For the component point cloud with small ratio, extract the point cloud of each member by RANSAC line fitting; for the component point cloud with large ratio, it is necessary to first extract the main member point cloud, then perform Euclidean clustering on the remaining point cloud to obtain the tower plate point cloud composed of two diagonal members, and then use the RANSAC line fitting algorithm to obtain the point cloud of a single diagonal member.

4. The method for reconstructing power poles based on ground-based LiDAR point clouds as described in claim 1, characterized in that, Step 4 involves registering the pole model and point cloud to indirectly complete the reconstruction of the tower model. This includes the following steps: Step 41: Constructing the surface model of the angle steel member. Set the center of the angle steel edge at the origin of the coordinate system. The edge of the angle steel is on the x-axis, and the two legs of the angle steel are on the z and y axes respectively, and the x, y, and z axes are perpendicular to each other. Step 42: Use the directed bounding box algorithm to obtain the x-axis of the rod point cloud. max y max z max x min y min z min The geometric parameters of the length, width, and height of the bar point cloud are extracted and calculated. BarLength (BL) represents the length of the bar, and Bar Width (BW) represents the width. When segmenting the point cloud of the angle steel double-limb bar, the limb width BW of the bar model is... The width of the rod point cloud is multiplied by a factor of 1. For cases where only a single rod point cloud is segmented, the limb width BW of the rod model is the width of the rod point cloud, and the length BL of the rod model is the length of the rod point cloud. Finally, the vertex coordinates of the corrected predefined model are translated and rotated to the positions of each rod point cloud in the global coordinate system using a rotation and translation matrix from the local coordinate system to the global coordinate system, thereby completing the coarse registration of the predefined rod model and the rod point cloud. Step 43: ICP fine registration of the rod under local coordinate system rotation; The specific steps are as follows: ① Fit the corresponding set of straight lines using the endpoint coordinates of each model member in the global coordinate system, calculate the distance from the spatial point in the tower point cloud to each straight line, count the spatial points that meet the conditions, and obtain the completed member point cloud; ② Perform a traditional ICP algorithm on the member point cloud and the model point cloud to complete the fine registration of the member point cloud in the first principal direction; ③ Then, for the member model point cloud that failed to register, rotate the model point cloud 180° in the local coordinate system with its first principal direction as the axis, and then perform a fine registration of ICP in the global coordinate system to complete the fine registration of the member point cloud in the second principal direction; ④ Finally, save the translation and rotation matrices corresponding to the translation and rotation operations performed on each model point cloud during the fine registration process, and perform the same translation and rotation processing on the member model to complete the fine registration of the member model and the tower point cloud.

Images