A LiDAR power line point cloud segmentation method based on PointNet++ and region growing

Abstract

The application provides an airborne LiDAR power line point cloud segmentation extraction method based on PointNet++ and region growing fusion. The power transmission corridor point cloud collected by a UAV is preprocessed, and the model is subjected to corridor block and class balanced sampling and manual data set labeling. The model is subjected to PointNet++ multi-scale set abstraction coding to extract local features, is subjected to feature propagation recovery to restore point-by-point features, is subjected to a region growing refinement module to repair breakpoints and burrs, and constitutes a PG-Net fusion model. In view of the power line point sparsity and class imbalance, a spatial continuity weighted loss function is designed by fusing a class weighted cross entropy and a neighbor consistency constraint to supervise point-by-point training. The trained model is used to extract power lines, the power line point cloud is obtained through catenary fitting and connectivity analysis, and then sag calculation, tree barrier measurement and hidden danger early warning are completed. The application can extract and quantitatively characterize power lines in different corridor environments, and has important significance for intelligent inspection and hidden danger early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent inspection technology for power transmission lines, specifically to an airborne LiDAR power line point cloud segmentation and extraction method based on the fusion of PointNet++ and region growing. Background Technology

[0002] Safety inspection of transmission line corridors is one of the core tasks of power grid operation and maintenance. The spatial relationship between power lines, towers, vegetation, buildings, and other terrain features directly determines the efficiency of identifying and the reliability of early warnings for safety hazards such as tree obstructions and crossings within the transmission line corridor. Airborne LiDAR can quickly acquire high-density 3D point clouds of transmission corridors and is a core data tool for power line inspection. Power line point clouds are characterized by fine targets, sparse point density, spatial discontinuity, and mixing with vegetation point clouds, posing a significant challenge to the accurate segmentation and extraction of power lines. Accurate segmentation and extraction of power line point clouds have important scientific and engineering value for line sag analysis, tree obstruction distance measurement, and early warning of corridor hazards. Traditional power line point cloud extraction methods typically employ rule-based geometric methods such as elevation thresholding, slope analysis, and Hough transform line detection, which suffer from problems such as parameter sensitivity, poor terrain adaptability, and failure in complex scenarios. In recent years, researchers have successively introduced indirect point cloud processing methods such as voxelization convolution and projection image segmentation for power line extraction. However, voxelization loses fine structural details and projection processing destroys three-dimensional topological relationships. These methods lack robustness to sparse power line points and mixed vegetation interference, making it difficult to achieve automated and high-precision power line point cloud segmentation. With the development of deep learning technology, networks such as PointNet and PointNet++ that directly process disordered point clouds have achieved significant results in the field of three-dimensional semantic segmentation and have been gradually introduced into the interpretation of power line inspection point clouds. However, power line point clouds have characteristics such as extremely imbalanced categories (power line points account for less than one percent) and linear catenary distribution in spatial topology. Single-structure point cloud segmentation networks still have the following shortcomings: (1) During the downsampling process, sparse power line points are easily submerged by a large number of vegetation and ground points, resulting in low recall; (2) Global semantic segmentation does not utilize the spatial prior of continuous catenary of power lines, and the segmentation results are prone to breakpoints and spikes; (3) The loss function does not adequately consider the imbalance between rare power line categories and dominant categories such as vegetation, which easily leads to power line omissions. Therefore, there is an urgent need for an automatic segmentation and extraction method that is tailored to the characteristics of power line point clouds and has the ability to represent local features at multiple scales and constrain spatial continuity. Summary of the Invention

[0003] To address the technical challenges of automating and achieving high-precision segmentation and extraction of airborne LiDAR power line point clouds due to interference from sparse power line point clouds, vegetation point clouds, and spatial discontinuities, existing technologies struggle with this problem. This invention aims to provide a segmentation and extraction method for airborne LiDAR power line point clouds based on a fusion of PointNet++ and region growing. The specific technical solution is as follows: A segmentation and extraction method for airborne LiDAR power line point clouds based on a fusion of PointNet++ and region growing, comprising: acquiring point cloud data of a UAV-borne LiDAR transmission corridor and performing denoising and elevation normalization preprocessing to obtain a standardized point cloud; performing corridor segmentation and balanced sampling data enhancement on the point cloud according to the spatial distribution characteristics of power lines; manually annotating each point in the sampled point cloud to obtain a reference point cloud dataset; and designing a multi-scale set abstract coding module based on PointNet++. The input point cloud is converted into a feature tensor to extract multi-scale local features of power lines. A point-by-point feature decoding module is designed based on feature propagation to recover full-resolution point-by-point features. A region growing refinement module is designed, including a seed point selection submodule and a continuous growth submodule, to refine the segmentation results using the spatial continuity of power lines. A PG-Net fusion model is obtained through fusion. A spatial continuity weighted loss function is designed based on the characteristics of sparse power line points and extreme class imbalance to obtain the loss function weights for rare power line classes. Point-by-point supervised calculation is performed based on the loss function weights to obtain model feedback. The PG-Net fusion model is then trained to obtain a trained model. The trained model is used to segment and extract power lines from the transmission corridor point cloud. The segmentation and extraction results are subjected to catenary fitting and connectivity analysis to obtain the power line point cloud results. Sag calculation, tree obstacle distance measurement, and hazard warning data analysis are performed on the power line point cloud results.

[0004] Furthermore, the point cloud data acquisition and preprocessing method includes: acquiring point cloud data collected by UAV-borne lidar scanning along the power transmission corridor, with a point cloud density of not less than 50 points per square meter, and the point cloud containing three-dimensional coordinates, echo intensity, and echo number attributes; performing statistical filtering to denoise the point cloud, removing outliers that deviate from the neighborhood average distance by more than 2 standard deviations; using cloth simulation filtering to extract ground points and performing elevation normalization on the point cloud, replacing the elevation of each point with the relative ground clearance height to obtain a standardized point cloud.

[0005] Furthermore, the data augmentation and annotation method includes: dividing the standardized point cloud into 150m×150m corridor blocks along the transmission corridor; performing sampling augmentation on each point cloud block according to a category equalization strategy, oversampling the blocks containing power line points and undersampling the blocks containing pure background points, so that the proportion of power line points is not less than 0.5%; augmenting the block point cloud with random rotation, random translation and random scaling around the vertical axis; manually annotating each point of the sampled point cloud, labeling power lines, towers, vegetation, ground and other ground features as different categories to generate reference point cloud data; uniformly sampling the reference point cloud data into 8192 points per block and constructing an N×6 feature tensor, where N is the number of points, and the 6-dimensional features include normalized three-dimensional coordinates, echo intensity, echo count and clearance height, to obtain a standard dataset.

[0006] Furthermore, the method for obtaining the PG-Net fusion model includes: a multi-scale ensemble abstracting encoding module consisting of four levels of ensemble abstracting layers, each level sequentially employing farthest point sampling to select the center point, ball query grouping to construct local neighborhoods, and a shared multilayer perceptron to extract local features, progressively abstracting the number of points to 2048, 512, 128, and 32 and increasing the feature dimension; a point-by-point feature decoding module consisting of four levels of feature propagation layers, each level using inverse distance weighted interpolation to propagate the abstract features back to the dense point set, and concatenating them with the corresponding level of encoded features via skip connections, then fusing them with the shared multilayer perceptron to output full-resolution point-by-point features; a region growing refinement module is designed, where the seed point selection submodule uses points with a power line prediction probability higher than a first threshold as seed points, and the continuous growth submodule uses the seed points as the starting point and iteratively grows along the power line direction according to the spatial distance and tangential consistency criteria, incorporating neighborhood points with a prediction probability higher than a second threshold into the power line category; the multi-scale ensemble abstracting encoding module, the point-by-point feature decoding module, and the region growing refinement module are fused to obtain the PG-Net fusion model.

[0007] Furthermore, the method for obtaining the spatial continuity weighted loss function includes: addressing the problems of sparse power line points, extreme class imbalance, and easy breakpoints in the segmentation results, introducing a spatial continuity constraint term based on the class-weighted cross-entropy loss to obtain the spatial continuity weighted loss function, calculated as follows: L_ce = -Σ w_c·y_c·log(p_c), w_c = log(N_total / N_c), L_cont = Σ_i Σ_j∈Ω(i) ||l_i - l_j||_1·exp(-||x_i - x_j||_2 / σ_d), L = λ1·L_ce + λ2·L_cont, where w_c is the balancing weight of class c; N_c is the number of points of class c in the dataset; N_total is the total number of points; p_c is the predicted class probability; y_c is the ground truth class label; Ω(i) is the set of k nearest neighbors of point i, where k takes the value of 16; l_i and l_j are the predicted label distributions of points i and j, respectively; x_i and x_j are the spatial coordinates of points i and j, respectively; σ_d is the distance decay coefficient, which takes the value of 0.5m; L_cont is the spatial continuity loss; λ1 and λ2 are the balancing coefficients; the loss function is used to perform point-by-point supervised calculation on the PG-Net fusion model to obtain model feedback.

[0008] Furthermore, the model training method includes: using power line points as foreground targets and vegetation, ground, and other ground features as background; pre-training the encoding and decoding modules using class-weighted cross-entropy loss during training, followed by joint fine-tuning using spatial continuity loss; setting the initial learning rate to 1×10^-3 and dynamically adjusting it according to a step decay strategy; setting the batch size to 16 and iterating for at least 100 epochs; stopping training when the cross-union ratio of power line categories on the validation set no longer increases for 10 consecutive epochs, and saving the optimal model weights.

[0009] Furthermore, the catenary fitting and connectivity analysis method includes: segmenting the extracted power line point cloud by span, fitting the catenary equation for each span of power line point cloud using the least squares method, the catenary equation being: y = a·cosh((x - x_0) / a) + b, where a is the catenary shape parameter, and x_0 and b are the horizontal and vertical translation parameters, respectively; calculating the fitting residuals and removing outliers greater than 3 times the standard deviation before iterative refitting; calculating the power line sag and the minimum clearance height based on the catenary model; measuring the minimum spatial distance between the power line model and the vegetation point cloud and building point cloud, generating hazard warnings for sections with distances less than the safe distance threshold, and obtaining the final power line point cloud results and statistical parameters.

[0010] This invention offers the following advantages: Firstly, by acquiring point clouds of power transmission corridors from UAV-borne lidar and performing denoising and elevation normalization, a standardized point cloud is obtained. This point cloud possesses rich morphological and spatial distribution variations, providing ample samples for deep learning models. Secondly, due to the extremely low proportion of power line points, corridor segmentation and class-balanced sampling are employed to augment the point cloud data, improving the model's sensitivity to rare classes while enriching data diversity. Thirdly, point clouds are manually labeled point by point to obtain reference point cloud data. Fourthly, unified sampling and feature construction are performed on the point cloud data to reduce computational load and improve training efficiency, resulting in a standard dataset. Fifthly, because power lines are characterized by their thinness, sparseness, and multi-scale local distribution, the PointNet++ multi-scale set abstract encoding module is used to directly process the unordered point cloud. Multi-scale local features are extracted step-by-step through farthest point sampling and sphere query grouping, avoiding the loss of thin structures during voxelization. For detailed issues, a feature propagation decoding module is used to recover full-resolution point-by-point features through inverse distance weighted interpolation and skip connections. Since power lines are distributed as continuous catenaries in space, a region growing refinement module is designed, starting from high-confidence seed points and growing according to spatial distance and tangential consistency criteria, utilizing spatial continuity priors to repair breakpoints and burrs. Due to the sparse power line points and extreme class imbalance, a spatial continuity weighted loss function is designed, using class balancing weights to strengthen supervision of rare classes and nearest-neighbor label consistency to constrain spatial continuity, improving power line recall without sacrificing overall accuracy. The trained model is used to segment the transmission corridor point cloud. Since the model segmentation results may contain gross errors and breakpoints, catenary fitting and residual iteration removal are used for post-processing to obtain complete power line point cloud results. Finally, data analysis is performed on the power line point cloud for sag calculation, tree obstacle distance measurement, and hazard warning. This invention fully considers the complexity of power line point clouds in terms of density, distribution, and topology, and designs a unique multi-scale set abstract encoding, feature propagation decoding, and region growth refinement module. This enhances the local feature expression capability and spatial continuity constraint capability of the point cloud segmentation model, improves the completeness, continuity, and recall rate of power line segmentation, and provides a reliable technical means for intelligent inspection of transmission lines. Attached Figure Description

[0011] To more clearly illustrate the 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. Figure 1 This is a flowchart of an airborne LiDAR power line point cloud segmentation and extraction method based on PointNet++ and region growing fusion, provided by an embodiment of the present invention. Figure 2 This is a network structure diagram of the PG-Net fusion model provided in one embodiment of the present invention. Detailed Implementation

[0012] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of an airborne LiDAR power line point cloud segmentation and extraction method based on PointNet++ and region growing fusion proposed according to the present invention. In the following description, different embodiments or one embodiment do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The specific scheme of the airborne LiDAR power line point cloud segmentation and extraction method based on PointNet++ and region growing fusion provided by the present invention is described in detail below with reference to the accompanying drawings.

[0013] Please see Figure 1 This paper illustrates an embodiment of the present invention that provides a method for segmenting and extracting airborne LiDAR power line point clouds based on the fusion of PointNet++ and region growing. The method includes the following steps.

[0014] Step S1: Acquire point cloud data of the power transmission corridor using UAV-borne LiDAR and perform denoising and elevation normalization preprocessing to obtain a standardized point cloud. This embodiment of the invention is mainly applied to point cloud segmentation scenarios of 110kV and above overhead power transmission line corridors. To acquire power line point cloud data, point clouds are first collected by scanning along the corridor using UAV-borne LiDAR, and standardized point clouds are obtained after denoising and elevation normalization. Since this embodiment of the invention uses a deep learning model to segment power lines, the model network needs a large number of point cloud samples containing power line morphology to ensure that the model has sufficient generalization ability. Preferably, in one embodiment of the present invention, the point cloud data acquisition and preprocessing method includes: using a rotary-wing UAV equipped with a lidar system to scan and collect point cloud data along the power transmission corridor in a round trip, with a flight altitude of 100m and a point cloud density of not less than 50 points per square meter, the point cloud containing three-dimensional coordinates, echo intensity, and echo number attributes; performing statistical filtering to denoise the point cloud, using the average distance of the 50 nearest neighbors of each point as a statistical measure, and removing outliers that deviate from the mean by more than 2 standard deviations; using cloth simulation filtering to extract ground points and construct a digital elevation model, replacing the elevation of each point with the relative clearance height to the ground; collecting point clouds of corridors with different terrains and voltage levels for no less than 200km to ensure the diversity and representativeness of the dataset.

[0015] Step S2: Based on the spatial distribution characteristics of power lines, corridor segmentation and class-balanced sampling are used to augment the point cloud data. The sampled point cloud is then manually labeled point by point to obtain a reference point cloud dataset. Since power line points account for less than one percent of the overall point cloud, corridor segmentation and class-balanced sampling are used to augment the point cloud data, improving the model's sensitivity to rare classes while enriching data diversity. Because the spatial distribution of power line points and vegetation points differs significantly, point cloud labels are manually applied. Since segmenting and extracting power lines only requires focusing on relative spatial distribution, elevation normalization and unified sampling are performed on the point cloud to reduce computation and improve training efficiency, resulting in a standard dataset. Preferably, in one embodiment of the present invention, the data augmentation and annotation method includes: dividing the standardized point cloud into 150m×150m corridor blocks along the transmission corridor; sampling each point cloud block according to a category equalization strategy, performing 3x oversampling on blocks containing power line points and 0.3x undersampling on pure background blocks, so that the proportion of power line points is not less than 0.5%; and combining the block point cloud with random rotation around the vertical axis from 0-360°, random translation of ±5m, and random scaling of 0.9-1.1x. The dataset was expanded to 6-8 times its original size. Point cloud annotation tools were used to manually annotate each point in the point cloud, and power lines, towers, vegetation, ground, and other features were labeled as categories 1-4 and 0, respectively, to generate reference point cloud data. The reference point cloud data was uniformly sampled into blocks of 8192 points each, and an N×6 feature tensor was constructed. The 6-dimensional features included normalized 3D coordinates, echo intensity, echo count, and clearance height, and were randomly divided into training, validation, and test sets in a ratio of 8:1:1.

[0016] Step S3: Design a multi-scale ensemble abstract coding module based on PointNet++; design a point-by-point feature decoding module based on feature propagation; design a region growing refinement module, and fuse them to obtain the PG-Net fusion model. In one embodiment of the present invention, the method for obtaining the PG-Net fusion model includes: using a multi-scale ensemble abstract coding module to extract multi-scale local features of the power line; using a point-by-point feature decoding module to recover full-resolution point-by-point features; and using a region growing refinement module to refine the segmentation results using spatial continuity. Preferably, in one embodiment of the present invention, the multi-scale set abstract coding module design method includes: the input layer constructs the input point cloud into an 8192×6 feature tensor; the coding module consists of 4 levels of set abstraction layers, the first level uses farthest point sampling to select 2048 center points, queries 32 neighboring points in a sphere with a radius of 0.5m, and outputs 128-dimensional local features through a shared multilayer perceptron; the second level abstracts 512 center points, a radius of 1.0m, and 256-dimensional features; the third level abstracts 128 center points, a radius of 2.0m, and 512-dimensional features; the fourth level abstracts 32 center points, a radius of 4.0m, and 1024-dimensional features. Preferably, in one embodiment of the present invention, the design method of the point-by-point feature decoding module and the region growing refinement module includes: the decoding module consists of 4 levels of feature propagation layers, each level uses inverse distance weighted interpolation of k=3 nearest neighbors to propagate abstract features back to the dense point set, and after being spliced ​​with the corresponding level features of the encoding through skip connections, it is fused by a shared multilayer perceptron, and finally outputs full-resolution point-by-point features of 8192×256; the seed point selection submodule of the region growing refinement module selects points with a power line prediction probability higher than 0.8 as seed points; the continuous growth submodule takes the seed point as the starting point, calculates the angle between the local principal direction and the tangent of the seed point for points with a prediction probability higher than 0.5 in a 0.3m spherical neighborhood, merges points with an angle less than 30° into the power line category and updates the growth queue, iterates until no new points are added, and obtains the PG-Net fusion model.

[0017] Step S4: Based on the characteristics of sparse power line points and extreme class imbalance, a spatial continuity weighted loss function is designed to obtain the loss function weights for rare power line classes. Point-by-point supervised calculation is performed based on these loss function weights to obtain model feedback. The spatial distribution and feature information of power line points and background points are obtained through the above process. Due to the sparse power line points, extreme class imbalance, and the tendency for segmentation results to break, a spatial continuity weighted loss function is designed and calculated based on the aforementioned feature information to obtain model feedback. Preferably, in one embodiment of the present invention, the model feedback acquisition method includes: obtaining the power line rare category loss weight according to the spatial continuity weighted loss function calculation formula, and calculating the model feedback. The calculation formula is as follows: L_ce = -Σ w_c·y_c·log(p_c), w_c = log(N_total / N_c), L_cont =Σ_i Σ_j∈Ω(i) ||l_i - l_j||_1·exp(-||x_i - x_j||_2 / σ_d), L = λ1·L_ce + λ2·L_cont, where Ω(i) is the 16 nearest neighbor set of point i; σ_d takes the value of 0.5m; λ1 and λ2 take the values ​​of 1.0 and 0.3, respectively; and performing point-by-point supervised calculation on the PG-Net fusion model using the loss function to obtain the model feedback.

[0018] Step S5: Train the model using the PG-Net fusion model to obtain a trained model; use the trained model to segment and extract power lines from the power transmission corridor point cloud; perform catenary fitting and connectivity analysis on the segmentation and extraction results to obtain power line point cloud results; perform sag calculation, tree obstacle distance measurement, and hazard warning data analysis on the power line point cloud results. Preferably, in one embodiment of the present invention, training the model using the PG-Net fusion model includes: pre-training the encoding and decoding modules for 60 epochs with class-weighted cross-entropy loss, and then adding spatial continuity loss for joint fine-tuning for 40 epochs; setting the initial learning rate to 1×10^-3, and decreasing it stepwise by 0.1 times at the 60th and 80th epochs; setting the batch size to 16, and using the Adam optimizer; iteratively training for 100 epochs, stopping training when the crossover ratio of power line classes in the validation set no longer increases for 10 consecutive epochs, and saving the optimal model weights. The trained model is used to segment the power transmission corridor point cloud, minimizing incomplete segmentation results and category confusion caused by the similarity of power line points and vegetation points, thus improving segmentation accuracy and completing the segmentation of the power line point cloud. Preferably, in one embodiment of the invention, the catenary fitting and connectivity analysis method includes: segmenting the extracted power line point cloud by span; fitting the catenary equation y = a·cosh((x - x_0) / a) + b for each span of power line point cloud using the least squares method; calculating the fitting residuals and removing outliers greater than 3 times the standard deviation, then iteratively refitting until convergence; calculating the power line sag and minimum clearance height based on the catenary model; measuring the minimum spatial distance between the power line model and the vegetation and building point clouds; generating tree barriers and crossing hazard warnings for sections with distances less than the safety distance threshold; and statistically analyzing the location, type, and degree of exceedance of hazard points. This completes the segmentation and extraction of the power line point cloud.

[0019] In summary, point cloud data of the power transmission corridor obtained by UAV-borne LiDAR was acquired and preprocessed with denoising and elevation normalization to obtain standardized point clouds. Based on the spatial distribution characteristics of power lines, the point cloud was segmented into corridor blocks and subjected to balanced sampling data enhancement. The sampled point clouds were manually annotated point by point to obtain a reference point cloud dataset. A multi-scale ensemble abstract encoding module was designed based on PointNet++. A point-by-point feature decoding module was designed based on feature propagation. A region growing refinement module was designed, and a PG-Net fusion model was obtained. A spatial continuity weighted loss function was designed based on the characteristics of sparse power line points and extreme class imbalance. Point-by-point supervised calculation was performed based on the weights of the loss function to obtain model feedback. The PG-Net fusion model was trained to obtain a trained model. The trained model was used to segment and extract power lines from the power transmission corridor point cloud. The segmentation and extraction results were subjected to catenary fitting and connectivity analysis to obtain the power line point cloud results. The power line point cloud results were then used for sag calculation, tree obstacle distance measurement, and hazard warning data analysis. This invention can segment, extract, and quantitatively characterize power line point clouds of different voltage levels and different terrain corridors, providing technical support for intelligent inspection and early warning of hidden dangers of power transmission lines.

[0020] It should be noted that the order of the above embodiments of the present invention is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. The processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous. The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments.

Claims

1. A method for segmenting and extracting airborne LiDAR power line point clouds based on PointNet++ and region growing fusion, characterized in that, The method includes: acquiring point cloud data of the power transmission corridor from an UAV-borne lidar and performing denoising and elevation normalization preprocessing to obtain a standardized point cloud; performing corridor segmentation and equalized sampling data enhancement on the point cloud according to the spatial distribution characteristics of power lines; manually annotating each point of the sampled point cloud to obtain a reference point cloud dataset; designing a multi-scale set abstract encoding module based on PointNet++ to convert the input point cloud into a feature tensor to extract multi-scale local features of the power lines; designing a point-by-point feature decoding module based on feature propagation to recover full-resolution point-by-point features; and designing a region growing refinement module, including a seed point selection submodule and a continuous growth submodule, to... The segmentation results are refined by utilizing the spatial continuity of power lines; a PG-Net fusion model is obtained through fusion; a spatial continuity weighted loss function is designed based on the characteristics of sparse power line points and extreme class imbalance to obtain the loss function weights for rare power line classes; point-by-point supervised calculation is performed based on the loss function weights to obtain model feedback; the PG-Net fusion model is trained to obtain a trained model; the trained model is used to segment and extract power lines from the transmission corridor point cloud; catenary fitting and connectivity analysis are performed on the segmentation and extraction results to obtain power line point cloud results; sag calculation, tree obstacle distance measurement, and hazard warning data analysis are performed on the power line point cloud results.

2. The method for segmenting and extracting airborne LiDAR power line point clouds based on PointNet++ and region growing fusion as described in claim 1, characterized in that, The point cloud data acquisition and preprocessing method includes: acquiring point cloud data collected by UAV-borne lidar scanning along the power transmission corridor, with a point cloud density of not less than 50 points per square meter, and the point cloud containing three-dimensional coordinates, echo intensity, and echo number attributes; performing statistical filtering to denoise the point cloud, removing outliers that deviate from the neighborhood average distance by more than 2 standard deviations; using cloth simulation filtering to extract ground points and performing elevation normalization on the point cloud, replacing the elevation of each point with the relative ground clearance height to obtain a standardized point cloud.

3. The method for segmenting and extracting airborne LiDAR power line point clouds based on PointNet++ and region growing fusion as described in claim 1, characterized in that, The data augmentation and annotation method includes: dividing the standardized point cloud into 150m×150m corridor blocks along the transmission corridor; performing sampling augmentation on each point cloud block according to a category equalization strategy, oversampling the blocks containing power line points and undersampling the blocks containing pure background points, so that the proportion of power line points is not less than 0.5%; augmenting the block point cloud with random rotation, random translation and random scaling around the vertical axis; manually annotating each point of the sampled point cloud, labeling power lines, towers, vegetation, ground and other features as different categories to generate reference point cloud data; uniformly sampling the reference point cloud data into blocks of 8192 points and constructing an N×6 feature tensor, where N is the number of points, and the 6-dimensional features include normalized three-dimensional coordinates, echo intensity, echo count and clearance height, to obtain a standard dataset.

4. The method for segmenting and extracting airborne LiDAR power line point clouds based on PointNet++ and region growing fusion as described in claim 1, characterized in that, The method for obtaining the PG-Net fusion model includes: a multi-scale ensemble abstracting encoding module consisting of four levels of ensemble abstracting layers. Each level of ensemble abstracting layer sequentially uses farthest point sampling to select the center point, ball query grouping to construct local neighborhoods, and a shared multilayer perceptron to extract local features, progressively abstracting the number of points to 2048, 512, 128, and 32 and increasing the feature dimension; a point-by-point feature decoding module consisting of four levels of feature propagation layers. Each level of feature propagation layer uses inverse distance weighted interpolation to propagate the abstract features back to the dense point set, and then concatenates them with the corresponding level of encoded features via skip connections before being fused by the shared multilayer perceptron to output full-resolution point-by-point features; a region growing and refining module is designed, where the seed point selection submodule uses points with a power line prediction probability higher than a first threshold as seed points, and the continuous growth submodule uses the seed points as the starting point and iteratively grows along the power line direction according to the spatial distance and tangential consistency criteria, incorporating neighborhood points with a prediction probability higher than a second threshold into the power line category; the multi-scale ensemble abstracting encoding module, the point-by-point feature decoding module, and the region growing and refining module are fused to obtain the PG-Net fusion model.

5. The method for segmenting and extracting airborne LiDAR power line point clouds based on PointNet++ and region growing fusion as described in claim 1, characterized in that, The method for obtaining the spatial continuity weighted loss function includes: addressing the problems of sparse power line points, extreme class imbalance, and easy breakpoints in the segmentation results, introducing a spatial continuity constraint term based on the class-weighted cross-entropy loss to obtain the spatial continuity weighted loss function, calculated as follows: L_ce = -Σ w_c·y_c·log(p_c), w_c = log(N_total / N_c), L_cont = Σ_i Σ_j∈Ω(i) ||l_i - l_j||_1·exp(-||x_i - x_j||_2 / σ_d), L = λ1·L_ce + λ2·L_cont, where w_c is the balancing weight of class c; N_c is the number of points of class c in the dataset; N_total is the total number of points; p_c is the predicted class probability; y_c is the ground truth class label; Ω(i) is the set of k nearest neighbors of point i, where k takes the value of 16; l_i and l_j are the predicted label distributions of points i and j, respectively; x_i and x_j are the spatial coordinates of points i and j, respectively; σ_d is the distance decay coefficient, which takes the value of 0.5m; L_cont is the spatial continuity loss; λ1 and λ2 are the balancing coefficients; the loss function is used to perform point-by-point supervised calculation on the PG-Net fusion model to obtain model feedback.

6. The method for segmenting and extracting airborne LiDAR power line point clouds based on PointNet++ and region growing fusion as described in claim 1, characterized in that, The model training method includes: using power line points as foreground targets and vegetation, ground, and other ground features as background; pre-training the encoding and decoding modules with class-weighted cross-entropy loss, followed by joint fine-tuning with spatial continuity loss; setting the initial learning rate to 1×10^-3 and dynamically adjusting it according to a step decay strategy; setting the batch size to 16 and iterating for at least 100 epochs; stopping training when the cross-union ratio of power line categories on the validation set no longer increases for 10 consecutive epochs, and saving the optimal model weights.

7. The method for segmenting and extracting airborne LiDAR power line point clouds based on PointNet++ and region growing fusion as described in claim 1, characterized in that, The catenary fitting and connectivity analysis method includes: segmenting the extracted power line point cloud from the model into sections according to span; fitting the catenary equation for each power line point cloud using the least squares method; the catenary equation is: y = a·cosh((x - x_0) / a) + b, where a is the catenary shape parameter, and x_0 and b are the horizontal and vertical translation parameters, respectively; calculating the fitting residuals and removing outliers greater than 3 times the standard deviation before iterative refitting; calculating the power line sag and the minimum clearance height based on the catenary model; measuring the minimum spatial distance between the power line model and the vegetation point cloud and building point cloud; generating hazard warnings for sections with distances less than the safe distance threshold; and obtaining the final power line point cloud results and statistical parameters.

Images