An analysis method and medium based on point cloud semantic segmentation and interpolation reconstruction of ground

Abstract

The application relates to the technical field of electric digital data processing, in particular to an analysis method and medium for ground based on point cloud semantic segmentation and interpolation reconstruction. The method comprises the following steps: acquiring three-dimensional point cloud data of a power transmission line; establishing a semantic segmentation model to perform semantic segmentation processing on the three-dimensional point cloud data, so as to obtain point cloud data of four categories of semantic labels; performing noise reduction processing on the point cloud data of the four categories of semantic labels respectively, so as to obtain corresponding point cloud data subsets; acquiring three-dimensional conductor trajectory data according to the conductor point cloud data subset; acquiring a tree barrier risk point set according to the vegetation point cloud data subset; adopting a multi-source collaborative reconstruction method to establish a digital ground model of the power transmission line; calculating the distance of each conductor point from the ground, and combining the tree barrier risk point set to output a safety evaluation report of the power transmission line. The method effectively solves the problems of three-dimensional modeling and safety evaluation of the power grid in a complex scene, and significantly improves the precision and reliability of automatic inspection.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to an analysis method and medium for ground reconstruction based on point cloud semantic segmentation and interpolation. Background Technology

[0002] With the widespread application of drones equipped with LiDAR technology, intelligent 3D point cloud inspection of transmission lines has become an important means of automated operation and maintenance in power systems. High-precision point cloud data allows for detailed modeling of key objects such as conductors, vegetation, towers, and the ground, enabling the automatic calculation of core safety indicators such as tree obstruction risk identification, conductor-to-ground distance assessment, and sag status monitoring. However, the actual point cloud data collected often faces challenges such as conductor obstruction and breakage, blurred vegetation boundaries, and sparse or missing ground points. Especially in complex scenarios such as multi-circuit, mountainous, and forested areas, traditional methods struggle to guarantee reconstruction accuracy and analytical reliability.

[0003] Currently, mainstream technologies typically employ a two-stage process: semantic segmentation + single data source interpolation. Specifically, existing processes first use general point cloud segmentation networks such as PointNet++ or RandLA-Net to classify the original LiDAR point cloud into different categories; then, local curve fitting is used to fill in the gaps in the segmented guide points, while cloth simulation filtering or inverse distance weighted interpolation is used to generate digital ground models for ground points; finally, tree barrier distances and ground heights are calculated based on the reconstructed guide points and the ground. However, existing technologies have the following problems: Ambiguous semantic segmentation boundaries: Existing models rely solely on category labels for supervision, lacking explicit modeling of geometric boundaries, leading to severe misclassification at the boundary between conductors and vegetation, directly affecting the accuracy of tree barrier distance calculation; Topological ambiguity in conductor reconstruction: The "nearest neighbor matching between slices" strategy only considers local spatial distances and ignores the global line topology, easily causing cross-loop incorrect connections in complex scenarios such as multi-loop paralleling, branching, and crossing; Ground reconstruction relies on single LiDAR data: In sparsely populated areas such as canyons and dense forests, traditional methods often exhibit "cloth penetration" or excessive smoothing, resulting in severe distortion of ground distance calculations. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides an analysis method and medium for ground reconstruction based on point cloud semantic segmentation and interpolation. This method acquires 3D point cloud data of transmission lines; establishes a semantic segmentation model to perform semantic segmentation processing on the 3D point cloud data, obtaining point cloud data with four semantic labels; performs noise reduction processing on the point cloud data of each of the four semantic labels to obtain corresponding point cloud data subsets; obtains 3D conductor trajectory data based on the conductor point cloud data subset; obtains a tree obstacle risk point set based on the vegetation point cloud data subset; establishes a digital ground model of the transmission line using a multi-source collaborative reconstruction method; calculates the ground distance of each conductor point; and outputs a safety assessment report of the transmission line based on the tree obstacle risk point set. This effectively solves the challenges of 3D modeling and safety assessment of power grids in complex scenarios, significantly improving the accuracy and reliability of automated inspections.

[0005] This invention employs the following technical solution: an analysis method for reconstructing the ground based on point cloud semantic segmentation and interpolation, the method comprising:

[0006] Acquire 3D point cloud data of power transmission lines;

[0007] A semantic segmentation model is established based on the improved PointTransformer network structure, and the semantic segmentation model is used to perform semantic segmentation processing on the three-dimensional point cloud data to obtain point cloud data with four categories of semantic labels: wires, vegetation, towers and ground.

[0008] Noise reduction processing was performed on the point cloud data of the four categories of semantic labels to obtain subsets of conductor point cloud data, vegetation point cloud data, tower point cloud data and ground point cloud data;

[0009] Vertical slicing and hierarchical clustering operations are performed on a subset of the traverse point cloud data to obtain the traverse topology map;

[0010] A global path search is performed based on the described traverse topology to obtain three-dimensional traverse trajectory data;

[0011] Obtain the minimum Euclidean distance from each vegetation point cloud in the vegetation point cloud data subset to the three-dimensional traverse trajectory data, and compare the minimum Euclidean distance corresponding to each vegetation point cloud with a set threshold to obtain the vegetation point clouds whose minimum Euclidean distance is less than the set threshold, thus obtaining the tree obstacle risk point set.

[0012] Based on the conductor points extracted from the ground point cloud data subset and the 3D conductor trajectory data, a digital ground model of the transmission line is established using a multi-source collaborative reconstruction method.

[0013] Based on the digital ground model using the three-dimensional conductor trajectory data, the distance to the ground of each conductor point is calculated, and a safety assessment report for the transmission line is output by combining the tree obstacle risk point set.

[0014] Furthermore, a semantic segmentation model is established based on the improved PointTransformer network structure, including:

[0015] A four-level encoder is used to perform hierarchical downsampling and feature extraction processing on the input 3D point cloud data;

[0016] The feature propagation decoder is connected to the fourth-level encoder and upsamples the features output by the fourth-level encoder through difference and skip connections.

[0017] The semantic classification head is connected to the output of the feature propagation decoder and is used to output the probability that each 3D point cloud data belongs to one of the four semantic categories.

[0018] The boundary distance prediction branch, connected in parallel with the semantic classification head to the output of the feature propagation decoder, is used to output the predicted boundary distance value for each 3D point cloud data.

[0019] Furthermore, when building a semantic segmentation model based on the improved PointTransformer network structure, boundary distance mapping constraints are also introduced;

[0020] The boundary distance mapping constraints include: distance mapping definition, network predicted distance graph, and joint loss function;

[0021] The distance mapping is specifically defined as follows: the shortest Euclidean distance between the 3D point cloud data of each category semantic label and the 3D point cloud data of different categories is used as the true value of the boundary distance of each 3D point cloud data.

[0022] The network prediction distance map specifically outputs the predicted boundary distance value of each 3D point cloud data through the boundary distance prediction branch;

[0023] The joint loss function is specifically constructed using standard cross-entropy loss and distance loss.

[0024] Furthermore, noise reduction processing is performed on the point cloud data of the four categories of semantic labels, specifically as follows:

[0025] For point cloud data with semantic labels for conductor categories, a combined filtering method that combines statistical outlier removal and radius outlier removal is used for noise reduction.

[0026] Point cloud data of vegetation category semantic labels are filtered by setting an elevation threshold based on their normalized elevation values ​​relative to the ground.

[0027] For point cloud data with semantic labels for pole / tower categories, adaptive filtering is performed based on the point density distribution in the vertical direction.

[0028] Furthermore, vertical slicing and hierarchical clustering operations are performed on a subset of the conductor point cloud data to obtain the conductor topology map, specifically:

[0029] Principal component analysis is performed on the subset of the conductor point cloud data to obtain the global direction of the transmission line;

[0030] Based on the global orientation of the transmission line, a longitudinal slicing operation with a set thickness is performed to obtain multiple longitudinal slices perpendicular to the global orientation.

[0031] The point clouds of the conductors in each vertical slice are clustered using a density-based clustering algorithm to obtain multiple clusters of conductor point clouds;

[0032] Based on the set conditions, multiple conductor point cloud clusters are filtered to obtain the valid point cloud clusters in each longitudinal slice;

[0033] Obtain valid point cloud clusters from all vertical slices as nodes, and construct a wire topology by connecting pairs of nodes in adjacent vertical slices.

[0034] Furthermore, the weights of the connecting edges between any two nodes in the conductor topology graph are calculated according to the following formula:

[0035] ;

[0036] in, For nodes and nodes The weight of the connecting edges between them It is a three-dimensional Euclidean distance. The angle between the principal direction vectors of two valid point cloud clusters. This is an indicator of split pattern consistency; when split patterns are consistent, then... =1, otherwise 0; These are preset weighting coefficients.

[0037] Furthermore, a global path search is performed based on the aforementioned traverse topology map to obtain three-dimensional traverse trajectory data, including:

[0038] The topology graph of the conductor is input into a graph neural network to perform a global optimal path search and output a preliminary conductor path.

[0039] For broken sections in the preliminary traverse path where point cloud is missing, interpolation is performed based on the point clouds corresponding to their adjacent nodes to generate corresponding virtual traverse point clouds.

[0040] Based on the preliminary traverse path and the virtual traverse point cloud, smoothing is performed in spatial order to obtain three-dimensional traverse trajectory data.

[0041] Furthermore, based on a subset of ground point cloud data combined with conductor points extracted from 3D conductor trajectory data, a multi-source collaborative reconstruction method is used to establish a digital ground model of the transmission line, including:

[0042] Acquire auxiliary terrain data for the power transmission line; the auxiliary terrain data includes: digital surface model and flight trajectory elevation data measured by airborne satellite;

[0043] The corridor area of ​​the transmission line is divided into an interpolated grid with adaptive resolution;

[0044] Obtain the three-dimensional coordinates of the traverse points and, in conjunction with the auxiliary terrain data as geometric constraints, establish an optimization objective function;

[0045] Solve the optimization objective function for each grid node in the interpolation grid to obtain the elevation data of each grid node;

[0046] A digital ground model of the transmission line is generated based on the interpolation data of each grid node.

[0047] The present invention further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements an analysis method for ground reconstruction based on point cloud semantic segmentation and interpolation as described above.

[0048] The beneficial effects of this invention are as follows: The semantic segmentation model established in this invention introduces boundary distance mapping constraints, which can significantly improve the classification accuracy and boundary clarity of the boundary area between the conductor and vegetation, providing a reliable foundation for subsequent analysis; By implementing targeted noise reduction on various point clouds, isolated noise, ground-hugging vegetation, and abnormal point clusters are effectively eliminated, enhancing data quality; By constructing a conductor topology map and combining it with graph neural networks for global path search and intelligent interpolation of broken segments, the problem of conductor misconnection and break repair in complex scenarios such as multi-loop and crossover is completely solved, achieving highly complete 3D reconstruction of conductors; By integrating multi-source data such as ground point clouds, conductor point constraints, and satellite trajectory elevation, and adopting an adaptive interpolation strategy, a high-fidelity digital ground model can still be reconstructed in sparse ground point areas such as canyons and dense forests; Finally, based on the accurately reconstructed conductor trajectory and ground model, high-precision automated calculation of tree obstacle distance, ground distance, and sag parameters is achieved, significantly improving the accuracy, completeness, and reliability of 3D modeling of transmission lines, providing an efficient and robust solution for intelligent power grid inspection and safety assessment in complex environments. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This is a schematic diagram of an analysis method for ground reconstruction based on point cloud semantic segmentation and interpolation, according to an embodiment of the present invention. Detailed Implementation

[0051] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0052] A schematic diagram of the analysis method for ground reconstruction based on point cloud semantic segmentation and interpolation according to an embodiment of the present invention is shown below. Figure 1 As shown, it includes:

[0053] Acquire 3D point cloud data of power transmission lines;

[0054] This invention aims to perform high-precision semantic segmentation on the raw point cloud data of power transmission lines collected by UAV lidar, thereby accurately distinguishing four key objects: conductors, vegetation, towers, and the ground. In particular, it improves the segmentation boundary accuracy of the area where conductors and vegetation meet, providing a reliable foundation for subsequent tree obstacle analysis and conductor reconstruction. Therefore, in this embodiment of the invention, the acquisition of three-dimensional point cloud data of power transmission lines requires the acquisition of high-density, high-precision three-dimensional lidar point cloud data covering the entire line corridor. The acquisition equipment can be a UAV platform equipped with lidar equipment, which also integrates a high-precision GNSS receiver and IMU (Inertial Measurement Unit), which can provide real-time centimeter-level absolute positioning and attitude data for georeferencing and accuracy assurance of the point cloud.

[0055] A semantic segmentation model is established based on the improved PointTransformer network structure, and the semantic segmentation model is used to perform semantic segmentation processing on the three-dimensional point cloud data to obtain point cloud data with four categories of semantic labels: wires, vegetation, towers and ground.

[0056] In this embodiment of the invention, the improved PointTransformer network structure uses the PointTransformer-V2 network model as the backbone network of the semantic segmentation model. This network model employs a vector attention mechanism to capture the geometric relationships between point clouds and retains multi-scale contextual information through hierarchical downsampling and upsampling. The input is the point cloud coordinates and the laser intensity I, and the output is the category probability distribution of each point cloud. Specifically, the structure of this semantic segmentation model includes:

[0057] The four-level encoder is used to perform hierarchical downsampling and feature extraction processing on the input 3D point cloud data. It aggregates the local features of the 3D point cloud data level by level, with downsampling rates of 1, 1 / 4, 1 / 16 and 1 / 64 respectively.

[0058] The feature propagation decoder, connected to the four-level encoder, upsamples the features output by the four-level encoder through interpolation and skip connections.

[0059] The semantic classification head, connected to the output of the feature propagation decoder, is used to output the probability of each 3D point cloud data belonging to one of the four semantic labels, i.e., the probability distribution of each point cloud belonging to the four categories of wire, vegetation, tower and ground.

[0060] The boundary distance prediction branch, connected in parallel with the semantic classification head to the output of the feature propagation decoder, consists of a multilayer perceptron and is used to output the predicted boundary distance value for each 3D point cloud data.

[0061] To train this semantic segmentation model, this embodiment of the invention further introduces boundary distance mapping constraints as a supervision signal. The boundary distance mapping constraints specifically include the distance mapping definition, the network-predicted distance graph, and the joint loss function, wherein:

[0062] The distance mapping is specifically defined as follows: the shortest Euclidean distance between the 3D point cloud data of each category semantic label and the 3D point cloud data of different categories is used as the ground truth of the boundary distance for each 3D point cloud data. The calculation expression is:

[0063] ;

[0064] In the formula, Point cloud The true value of the boundary distance. For point clouds True semantic tags, For point clouds True semantic tags, Representation and point cloud 3D point clouds with different semantic categories It is an L2 norm.

[0065] The network prediction distance graph specifically outputs the predicted boundary distance value for each 3D point cloud data through the boundary distance prediction branch. ;

[0066] The joint loss function is specifically constructed using standard cross-entropy loss and distance loss, and is expressed as follows:

[0067] ;

[0068] In the formula, For the joint loss function, Standard cross-entropy loss is used for class prediction; The distance loss is denoted by λ, which is a balance coefficient. A commonly preferred value is any value between 0.3 and 0.5. The distance loss is obtained through the distance mapping definition and the network's predicted distance map, and its expression is:

[0069] ;

[0070] In the formula, For distance loss, Point cloud The true value of the boundary distance. Point cloud The predicted boundary distance value, This represents the number of categories in the point cloud.

[0071] By establishing the boundary distance mapping constraint, the semantic segmentation model can pay attention to the geometric boundary information of the point cloud during training, enabling higher classification confidence at easily confused boundary points such as wire edge points.

[0072] In one specific embodiment of the present invention, the data used to train the semantic segmentation model is UAV LiDAR point cloud data from power transmission lines with different terrains and voltage levels in multiple provinces. The data is meticulously annotated with four semantic labels by professionals, covering approximately 120 kilometers of lines and containing approximately 240 million points. To improve the model's robustness, online data augmentation methods are employed during training, including random rotation around the Z-axis, random translation, random laser intensity jitter, and simulated conductor occlusion. The optimizer selected for the model is AdamW, and the initial learning rate is set to... The cosine annealing strategy was used for adjustment; during training, training batches were generated by a combination of random sampling and farthest point sampling, and the number of point clouds input in each batch was set to 8192 points.

[0073] After training, in order to further optimize the semantic segmentation results, the initial labels are further post-processed. Specifically, for the boundary region between the guide line and vegetation, if the proportion of guide line point clouds in a certain radius neighborhood of a point cloud classified as vegetation exceeds 60%, the point cloud is reclassified as a guide line to repair the eroded guide line edges. Finally, a conditional random field is used to optimize the segmentation results to improve the spatial smoothness and consistency of the labels.

[0074] Noise reduction processing was performed on the point cloud data of the four categories of semantic labels to obtain subsets of conductor point cloud data, vegetation point cloud data, tower point cloud data and ground point cloud data;

[0075] In this embodiment of the invention, for point cloud data with conductor category semantic labels, a combined filtering method that combines statistical outlier removal and radius outlier removal is used for noise reduction. The specific noise reduction process is as follows:

[0076] First, for a subset of point clouds whose semantic label category is "wire", calculate the average distance between each point and its k nearest neighbor point clouds. If the average distance of a point cloud exceeds the average distance of all points globally... times (of which, The global mean. Standard deviation, If the threshold is a multiple of 1.5 (preferably 1.5), then the point cloud is determined to be an outlier and removed, thereby effectively filtering out discrete points that significantly deviate from the main structure of the conductor.

[0077] To further eliminate localized noise, this embodiment of the invention counts the number of points within a spherical neighborhood of a specified radius r around each traverse point after the previous step. This radius can be adjusted according to the traverse splitting distance. If the number of points in the neighborhood is less than a set threshold, the point is considered an isolated invalid point and deleted. By iteratively executing the above two steps, the denoising effect is enhanced, ultimately obtaining a traverse point cloud data subset with a clear structure and significantly reduced noise.

[0078] Point cloud data of vegetation category semantic labels are filtered by setting an elevation threshold based on their normalized elevation values ​​relative to the ground.

[0079] Since the main interference in vegetation point clouds comes from low-lying weeds and shrubs that grow close to the ground, which are meaningless for tree barrier analysis, this embodiment of the invention adopts an elevation-based screening method. First, the ground elevation of the current area is obtained or estimated through local plane fitting. Then, for the point cloud data of vegetation category semantic labels, the normalized elevation relative to the ground is calculated: an empirical elevation threshold is set, preferably 1.5 meters, which is the distinguishing height between common low vegetation and meaningful trees / shrubs. Only vegetation points with an elevation value greater than 1.5 meters are retained, thereby directly eliminating noise such as ground-lying weeds.

[0080] For point cloud data with semantic labels for pole / tower categories, adaptive filtering is performed based on the point density distribution in the vertical direction.

[0081] Since the point cloud of the tower may contain hanging objects attached to the tower or clusters of suspended noise points generated by scanning anomalies, this embodiment of the invention adopts adaptive filtering based on vertical density distribution: First, the point cloud data of the tower category semantic label is extracted, and it is divided into a series of thin slices with a set thickness along the vertical direction, each slice being 0.2-0.5 meters thick; for each slice, its point density is calculated, and the mean and standard deviation of the point density of all slices are calculated. A density threshold is set according to the mean and standard deviation, and slices with point densities lower than this threshold are judged as anomalies, and the point cloud contained in them is removed as a whole; Point cloud homogenization and smoothing: the remaining point cloud is downsampled using a voxel grid to make the point distribution more uniform.

[0082] In this embodiment, the ground point cloud mainly serves as the basis for subsequent multi-source reconstruction. Its initial noise reduction focuses more on removing obvious misclassified points. Usually, relatively clean ground points can be obtained in the semantic segmentation stage. Therefore, no additional noise reduction processing is performed in this embodiment of the invention.

[0083] Vertical slicing and hierarchical clustering operations are performed on a subset of the traverse point cloud data to obtain the traverse topology map;

[0084] To determine the spatial orientation of a transmission line in this embodiment, principal component analysis is first performed on a subset of the conductor point cloud data. The first principal component direction vector is extracted, which characterizes the macroscopic orientation of the transmission line, thus obtaining its global orientation. Based on this global orientation, longitudinal slicing is performed with a set thickness to obtain multiple longitudinal slices perpendicular to the global orientation. This set thickness can be adaptively adjusted according to the point cloud density to ensure that each longitudinal slice contains sufficient point cloud data for analysis. Typically, it can be set as follows:

[0085] ;

[0086] In the formula, This refers to setting the thickness.

[0087] Within each longitudinal slice, fine-grained clustering of conductor points is required to separate different conductors. Therefore, this embodiment of the invention uses a density-based clustering algorithm to cluster the conductor point clouds in each longitudinal slice, thereby obtaining multiple conductor point cloud clusters. For each conductor point cloud cluster, the number of points it contains, its elevation span, and the deviation between the cluster centroid elevation and the average elevation of the line are calculated. Based on these, screening conditions are set to filter multiple conductor point cloud clusters and obtain the effective point cloud clusters in each longitudinal slice. The screening conditions can be as follows: if the number of points in a conductor point cloud cluster is less than 8, it is considered to contain too few points and may be isolated noise points; if the elevation span in a conductor point cloud cluster is less than 0.3 meters, it may be horizontal birds or ribbons; if the elevation deviation is greater than 3 meters, it may be a missegmented tower component point cloud. The points retained after screening by the set conditions are taken as the effective point cloud clusters in the longitudinal slice. The number of effective clusters is counted as the number of conductors in the longitudinal slice. If the number of conductors is consistent in multiple consecutive slices, the segmentation type of the section is confirmed to be stable.

[0088] This invention further obtains all valid point cloud clusters in the longitudinal slices as graph nodes, and constructs a traverse topology graph by connecting pairs of nodes in adjacent longitudinal slices, that is, the centroid of each valid traverse cluster. As a graph node, where The slice number. It is a clustered index; for any adjacent vertical slice and nodes and The weights of its connecting edges are defined as follows:

[0089] ;

[0090] in, For nodes and nodes The weight of the connecting edges between them It is a three-dimensional Euclidean distance. The angle between the principal direction vectors of two valid point cloud clusters. This is an indicator of split pattern consistency; when split patterns are consistent, then... =1, otherwise 0; The preset weighting coefficients can be typically set to 1.0, 0.5, or 2.0 to emphasize split consistency.

[0091] A global path search is performed based on the traverse topology map to obtain three-dimensional traverse trajectory data;

[0092] In this embodiment of the invention, based on the obtained conductor topology map, a graph neural network is invoked to perform global optimization on the entire conductor topology map, thereby finding all real conductor paths. This not only correctly distinguishes double-loop and multi-split conductors, but also accurately identifies complex structures such as T-junction branches or crossovers, completely solving the problem of easy misconnection of lines in multi-loop scenarios using traditional methods.

[0093] Specifically, this embodiment of the invention employs a network model based on a graph attention mechanism. The input of this model is a wire topology graph, where the feature vector of each node typically includes information such as its spatial coordinates, principal direction vector, number of splitting roots, and local point cloud density. The network iteratively passes messages between nodes in the graph through multiple layers of graph convolution or graph attention layers. Each layer aggregates information from neighboring nodes, so that the features obtained by each node not only include its own geometric attributes but also incorporate the global topological context. After obtaining the enhanced node features, a learning-based edge classification strategy or a global optimization algorithm can be used to determine which edges in the graph are real connections and which are spurious connections. Finally, an ordered sequence of nodes is output, with each sequence corresponding to a preliminary wire path of a physical wire.

[0094] For the point cloud gaps in the preliminary traverse path caused by occlusion, resulting in broken sections, this embodiment of the invention further provides a data-driven completion scheme. Specifically, for each broken section, several consecutive valid nodes are extended forward and backward from the effective nodes before and after the break, forming a local subgraph. This local subgraph is then input into a lightweight graph neural network regression module. By aggregating the geometric and topological features of the nodes within the subgraph, a three-dimensional unit vector is output as the predicted local growth direction of the broken section. Interpolation is performed using this unit vector to generate a small number of virtual points along a direction perpendicular to the traverse line to simulate the cross-section of the traverse, filling the point cloud gaps in the broken section. Finally, the point cloud is smoothed according to the spatial order of the preliminary traverse path and the virtual traverse point cloud to obtain three-dimensional traverse trajectory data.

[0095] Obtain the minimum Euclidean distance from each vegetation point cloud in the vegetation point cloud data subset to the three-dimensional traverse trajectory data, and compare the minimum Euclidean distance corresponding to each vegetation point cloud with a set threshold. Obtain the vegetation point clouds whose minimum Euclidean distance is less than the set threshold to obtain the tree barrier risk point set.

[0096] In this embodiment of the invention, for any vegetation point cloud in a subset of vegetation point cloud data, the minimum Euclidean distance from it to the three-dimensional traverse trajectory data is calculated. Iterate through all vegetation point clouds in the subset of vegetation point cloud data, and filter out the tree barrier risk point set based on the tree barrier risk assessment criteria. The assessment criteria are as follows:

[0097] ;

[0098] In the formula, For the first A collection of vegetation point clouds Tree barrier risk, To set a threshold, representing a safe distance, it is usually set according to the voltage level.

[0099] Based on the conductor points extracted from the ground point cloud data subset and the 3D conductor trajectory data, a digital ground model of the transmission line is established using a multi-source collaborative reconstruction method.

[0100] In this embodiment of the invention, for a subset of ground point cloud data output by semantic segmentation, especially in severely occluded areas such as canyons and dense forests, high-fidelity digital terrain model (DTM) reconstruction is achieved by fusing multi-source heterogeneous data, as detailed below:

[0101] First, auxiliary terrain data for the transmission line is acquired. This auxiliary terrain data includes a digital surface model (DSM) measured by an airborne satellite and flight trajectory elevation data. Then, based on a subset of ground point cloud data obtained after semantic segmentation, high-quality ground seed points are obtained by filtering out elevation and slope anomalies. From the reconstructed 3D conductor trajectory data, the 3D coordinates of the two ends of each conductor at the pole hanging hardware are accurately extracted. Their elevation values ​​are used as hard geometric constraints that must be strictly followed in subsequent ground reconstruction, as they represent an absolutely accurate known height benchmark above the ground and are crucial for correcting terrain distortion.

[0102] The corridor area of ​​the transmission line is divided into an interpolated grid with adaptive resolution; this adaptive resolution can be dynamically determined based on the spatial density of ground seed points, and is expressed as:

[0103] ;

[0104] in, For adaptive resolution, For ground point cloud density, the grid size is automatically increased in densely vegetated areas and refined in bare areas.

[0105] Obtain the 3D coordinates of the traverse points and combine them with auxiliary terrain data as geometric constraints to establish an optimization objective function, expressed as:

[0106] ;

[0107] in, The elevation values ​​of the difference grid nodes. This represents the elevation value of the ground seed point. For flight trajectory elevation data, The elevation values ​​are for the satellite digital surface model (DSM). = To enable the smoothing regularization term, it should only be used in regions with a slope of 45° or less. Ground curvature is used to match large-scale terrain trends. Norm operators, Denotes the square of the L2 norm. , , as well as The weight is typically set to: =0.6, =0.3, =0.1, and the elevation of the traverse points is forcibly satisfied through methods such as the Lagrange multiplier method.

[0108] For this objective function, the smoothing regularization term needs to be disabled in regions with a slope greater than 45°. Instead, a convexity-preserving irregular triangular network strategy is adopted. In the initial triangular network formed by hard constraints such as suspension points, local densification interpolation is performed according to rules such as slope aspect consistency to maintain the distinctiveness of features such as ridges and valleys.

[0109] The objective function is optimized for each grid node in the interpolation grid to obtain the elevation data of each grid node; a digital ground model of the transmission line is generated based on the interpolation data of each grid node.

[0110] Based on the three-dimensional conductor trajectory data and digital ground model, the distance to the ground of each conductor point is calculated, and a safety assessment report of the transmission line is output by combining the tree obstacle risk point set.

[0111] In this embodiment of the invention, for each traverse point in the three-dimensional traverse trajectory, the elevation value of the corresponding ground position directly below it is obtained in the established digital ground model by methods such as bilinear interpolation. The ground distance of the traverse point is then the elevation difference between its spatial coordinates and the ground position directly below it. By traversing all traverse points, the corresponding ground distance is calculated. Subsequently, the minimum, maximum, and average ground distances of each traverse, each phase, or each span can be statistically analyzed to identify potential hazard sections with insufficient ground distances.

[0112] In this embodiment of the invention, sag is used to reflect conductor tension and operating status. First, the continuous conductor trajectory is divided into independent spans based on the tower positions. Each span is defined by two adjacent tower suspension points. Then, the theoretical elevation of the line connecting the suspension points of the span at the horizontal midpoint is calculated. In the point cloud of the conductor trajectory corresponding to the span, all points located near the midpoint of the span are selected, and the actual elevation of the point with the lowest elevation is found. The sag of the span is then the difference between the theoretical elevation and the actual elevation.

[0113] In this embodiment of the invention, the tree obstacle risk point set is further associated with the spatial location of the line. By associating each tree obstacle risk point with the nearest conductor and the span in which it is located, and integrating the minimum distance value and risk level information of the risk point into the risk overview of the corresponding span or line section, a complete assessment report can be generated by automatically summarizing all analysis results.

[0114] In a specific instance of this invention, a comparative table is provided below, showing the performance of the method of this invention compared with existing technologies in terms of semantic segmentation accuracy, global path search, and ground reconstruction accuracy:

[0115] Table 1. Comparison of semantic segmentation accuracy improvements

[0116] ;

[0117] As can be seen, this invention achieves a breakthrough in the segmentation performance of the most critical conductor-vegetation boundary region by introducing an improved PointTransformer network with boundary distance mapping constraints. Compared with the existing technology using PointNet++, the intersection-union ratio (mIoU) is improved by 13.7%, reaching 96.0%; the average position offset error of the boundary points is significantly reduced by 64%, from 0.22 meters to 0.08 meters, fundamentally solving the problem of conductor "erosion" caused by boundary ambiguity in traditional methods, and providing a reliable premise for subsequent accurate tree barrier analysis.

[0118] Table 2. Comparison of Traverse Topology Reconstruction and Global Path Search

[0119] ;

[0120] As can be seen, this invention achieves outstanding results through the global path search framework of Conductor Topology Graph (CTG) and Graph Neural Network (GNN). Compared with the traditional "slice nearest neighbor matching" method, this invention improves the accuracy of multi-loop conductor connection by 14 percentage points, reaching 96%; the number of misjudgments in complex structures such as crossings is reduced by 89%; and for broken conductors caused by occlusion, the integrity rate of the completed trajectory is improved by 19 percentage points, reaching 97%, effectively eliminating cross-loop erroneous connections and ensuring the topological correctness and geometric integrity of the reconstructed trajectory.

[0121] Table 3. Comparison of Ground Reconstruction Accuracy

[0122] ;

[0123] As can be seen, the multi-source data fusion reconstruction method adopted in this invention demonstrates significant advantages when facing harsh terrains such as canyons and dense forests. Compared with the cloth simulation filtering (CSF) method that relies on single LiDAR data, the root mean square error (RMSE) of the reconstructed digital terrain model (DTM) is reduced by 62% (optimized from 0.55 meters to 0.21 meters), and the maximum elevation error is reduced by 63%. Based on the high-fidelity terrain model and accurate traverse trajectory, the calculation error of the core safety parameters is drastically reduced: the calculation error of the traverse distance to the ground is reduced by 70%, and the calculation error of sag is reduced by 71%.

[0124] The above description is only a preferred 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 within the protection scope of the present invention.

Claims

1. An analysis method of ground based on semantic segmentation and interpolation reconstruction of point cloud, characterized in that, The application relates to a power transmission line safety evaluation method based on improved PointTransformer network structure. The method comprises the following steps: acquiring three-dimensional point cloud data of a power transmission line; establishing a semantic segmentation model based on an improved PointTransformer network structure, and performing semantic segmentation processing on the three-dimensional point cloud data by using the semantic segmentation model to obtain point cloud data with four-class semantic labels of conductors, vegetation, towers and ground; the semantic segmentation model based on the improved PointTransformer network structure comprises: a four-level encoder for performing hierarchical downsampling and feature extraction processing on the input three-dimensional point cloud data; a feature propagation decoder connected to the four-level encoder and performing upsampling processing on the features output by the four-level encoder through difference and jump connection; a semantic classification head connected to the output end of the feature propagation decoder and used for outputting the probability of each three-dimensional point cloud data belonging to the four-class semantic labels; a boundary distance prediction branch connected to the output end of the feature propagation decoder in parallel with the semantic classification head and used for outputting the predicted boundary distance value of each three-dimensional point cloud data; when the semantic segmentation model based on the improved PointTransformer network structure is established, a boundary distance mapping constraint is introduced; the boundary distance mapping constraint comprises a distance mapping definition, a network predicted distance map and a joint loss function; the distance mapping definition is specifically that the shortest Euclidean distance between the three-dimensional point cloud data of each class semantic label and the three-dimensional point cloud data of different classes is calculated as the boundary distance true value of each three-dimensional point cloud data; the network predicted distance map is specifically that the predicted boundary distance value of each three-dimensional point cloud data is output by the boundary distance prediction branch; the joint loss function is specifically that the joint loss function is constructed by using standard cross-entropy loss and distance loss; the point cloud data of the four-class semantic labels are respectively subjected to noise reduction processing to obtain a conductor point cloud data subset, a vegetation point cloud data subset, a tower point cloud data subset and a ground point cloud data subset; the conductor point cloud data subset is subjected to longitudinal slicing and hierarchical clustering to obtain a conductor topology graph; global path searching is performed according to the conductor topology graph to obtain three-dimensional conductor trajectory data; the minimum Euclidean distance of each vegetation point cloud in the vegetation point cloud data subset to the three-dimensional conductor trajectory data is acquired, and the minimum Euclidean distance corresponding to each vegetation point cloud is compared with a set threshold to acquire vegetation point clouds with a minimum Euclidean distance less than the set threshold, thereby obtaining a tree barrier risk point set; a digital ground model of the power transmission line is established by using a multi-source collaborative reconstruction method according to the ground point cloud data subset and the conductor points extracted from the three-dimensional conductor trajectory data; 2. The analysis method of claim 1, wherein: the ground distance of each conductor point in the three-dimensional conductor trajectory data is calculated according to the digital ground model, and a safety evaluation report of the power transmission line is output in combination with the tree barrier risk point set. the point cloud data of the four-class semantic labels are respectively subjected to noise reduction processing, and the noise reduction processing of the point cloud data of the conductor class semantic label is specifically as follows: statistical outlier removal and radius outlier removal are combined to perform noise reduction on the point cloud data of the conductor class semantic label. The point cloud data of the vegetation category semantic label is filtered by setting an elevation threshold based on its normalized elevation value relative to the ground; The point cloud data of the tower category semantic label is adaptively filtered based on its point density distribution in the vertical direction. 3.The analysis method of ground surface based on point cloud semantic segmentation and interpolation reconstruction according to claim 1, characterized in that: The conductor point cloud data subset is longitudinally sliced and hierarchically clustered to obtain a conductor topology graph, specifically: The conductor point cloud data subset is subjected to principal component analysis to obtain the global trend of the transmission line; The global trend of the transmission line is sliced longitudinally with a set thickness to obtain multiple longitudinal slices perpendicular to the global trend; The conductor point cloud in each longitudinal slice is clustered using a density-based clustering algorithm to obtain multiple conductor point cloud clusters; The multiple conductor point cloud clusters are filtered according to a set condition to obtain valid point cloud clusters in each longitudinal slice; The valid point cloud clusters in all longitudinal slices are obtained as nodes, and the conductor topology graph is constructed by connecting two nodes in adjacent longitudinal slices.

4. The analysis method of claim 3, wherein: The connection edge weight between two nodes in the conductor topology graph is calculated according to the following formula: ； wherein, is a connection edge weight between a node and a node , is a three-dimensional Euclidean distance, is an angle between two effective point cloud cluster principal direction vectors, is a split mode consistency index, when the split mode is consistent, then , otherwise 0; , , is a preset weight coefficient.

5. The analysis method of claim 1, wherein: Global path search is performed based on the conductor topology graph to obtain three-dimensional conductor trajectory data, including: The conductor topology graph is input into a graph neural network for global optimal path search, and a preliminary conductor path is output; For the broken sections in the preliminary conductor path where point cloud is missing, interpolation operation is performed according to the point cloud corresponding to the adjacent nodes to generate corresponding virtual conductor point cloud; The preliminary conductor path and the virtual conductor point cloud are smoothed in spatial order to obtain three-dimensional conductor trajectory data.

6. The analysis method of claim 1, wherein: A multi-source collaborative reconstruction method is used to establish a digital ground model of the transmission line based on the ground point cloud data subset and the conductor points extracted from the three-dimensional conductor trajectory data, including: Obtain auxiliary terrain data of the transmission line; the auxiliary terrain data includes: digital surface model measured by airborne satellite and flight trajectory elevation data; Divide the corridor area of the transmission line into interpolation grids with adaptive resolution; Obtain the three-dimensional coordinates of the conductor points and combine them with the auxiliary terrain data as geometric constraints to establish an optimization objective function; Solve the optimization objective function for each grid node in the interpolation grid to obtain the elevation data of each grid node; Generate a digital ground model of the transmission line based on the interpolation data of each grid node.

7. A computer readable storage medium having stored thereon a computer program, characterized in that: The computer program is executed by a processor to implement an analysis method based on point cloud semantic segmentation and interpolation reconstruction of the ground according to any one of claims 1-6.

Images