Tree barrier analysis-oriented power channel lightweight semantic point cloud reconstruction and visualization method and system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ELECTRIC POWER RES INST OF EAST INNER MONGOLIA ELECTRIC POWER
- Filing Date
- 2026-03-30
- Publication Date
- 2026-06-26
Smart Images
Figure CN122289982A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent inspection and safety monitoring technology for power transmission lines, specifically to a method and system based on laser point cloud data processing and analysis. Background Technology
[0002] The safe and stable operation of overhead transmission lines is crucial for ensuring power supply. In complex geographical and climatic environments, the growth of trees within and around the transmission line corridors is one of the main hidden dangers causing transmission line tripping, flashovers, and even large-scale power outages—a phenomenon known as the "line-tree conflict." Traditional manual inspection methods are inefficient, have limited coverage, and pose personal safety risks. In recent years, drone inspection technology, especially inspection methods equipped with LiDAR (Light Detection and Ranging) radar, has become an important means of acquiring three-dimensional information about power transmission corridors. LiDAR can quickly acquire high-precision three-dimensional point cloud data, providing a foundation for analyzing the spatial relationship between transmission lines and vegetation.
[0003] However, current tree obstacle analysis methods based on laser point clouds still face many technical challenges. First, the raw point cloud data is massive, containing a large amount of redundant information. Direct processing and transmission consume enormous computing, storage, and communication resources, making it difficult to meet the needs of real-time or near-real-time monitoring. Second, accurately and efficiently segmenting key features such as power lines and vegetation from massive point clouds is a major challenge, especially in complex backgrounds and with uneven point cloud density. Existing methods either rely on complex manual features or employ computationally intensive deep learning models, making it difficult to achieve a balance between accuracy and efficiency. Furthermore, the segmented power line point clouds are usually only fitted with simple geometric curves, failing to fully consider their true physical form under actual tension and tower support constraints, leading to discrepancies between the reconstructed model and the actual situation, affecting the accuracy of subsequent distance calculations. In addition, existing visualization solutions mostly remain at the level of displaying raw point clouds or simple models, lacking a 3D semantic scene that is deeply integrated with tree obstacle risk analysis results and supports efficient interaction and multi-level detailed browsing, thus restricting the in-depth utilization of inspection data and its decision support effectiveness.
[0004] Therefore, there is an urgent need for a comprehensive solution that can achieve lightweight processing of power transmission line point cloud data, high-precision semantic segmentation, reconstruction of physical reality models, intelligent risk analysis, and strong interactive visualization capabilities, so as to improve the intelligence level and practical value of monitoring tree obstruction hazards on transmission lines. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a lightweight semantic point cloud reconstruction and visualization method and system for power channels oriented towards tree barrier analysis. It generates textured point clouds through multi-source data fusion, achieves accurate semantic segmentation by using a lightweight network, reconstructs a physically realistic power line model by combining support point constraints and mechanical models, and performs adaptive lightweighting of vegetation point clouds based on local geometric features to achieve high-precision tree barrier distance detection and risk assessment. It also constructs a multi-level interactive three-dimensional visualization scene that integrates analysis results.
[0006] In a first aspect, embodiments of this application provide a lightweight semantic point cloud reconstruction and visualization method for power channels oriented towards tree obstacle analysis, the method comprising:
[0007] The drone uses sensors to simultaneously collect laser point cloud and image data of the power channel, and then fuses them to generate an initial point cloud with texture.
[0008] The initial point cloud is subjected to lightweight semantic segmentation. Through a neural network that includes local feature aggregation and attention mechanisms, power line point clouds and vegetation point clouds are identified and segmented.
[0009] For the segmented power line point cloud, the coordinates of the actual support points suspended at both ends of the tower are obtained, and a three-dimensional power line model is generated by fitting and reconstructing based on a preset mechanical model.
[0010] The segmented vegetation point cloud is adaptively thinned according to its local geometric features to generate a lightweight vegetation point cloud that retains key structures.
[0011] Based on the three-dimensional power line model and the lightweight vegetation point cloud, the spatial distance between the vegetation and the power lines is calculated, potentially dangerous vegetation areas are identified, and risk assessment is performed based on the extracted features.
[0012] By integrating the results of semantic segmentation, power line model, vegetation point cloud, and risk assessment, a 3D visualization scene supporting multi-level detail rendering and interactive querying is constructed.
[0013] Secondly, embodiments of this application provide a lightweight semantic point cloud reconstruction and visualization system for power channels oriented towards tree obstacle analysis, applied to the method described in the first aspect, the system comprising:
[0014] The data acquisition and fusion module is used to synchronously acquire laser point cloud and image data of the power channel through sensors carried by the UAV, and fuse them to generate an initial point cloud with texture.
[0015] The lightweight semantic segmentation module is used to perform lightweight semantic segmentation on the initial point cloud. It identifies and segments the power line point cloud and vegetation point cloud through a neural network that includes local feature aggregation and attention mechanisms.
[0016] The physical constraint power line reconstruction module is used to obtain the actual support point coordinates of the power line point cloud suspended at both ends of the tower, and to perform fitting and reconstruction based on a preset mechanical model to generate a three-dimensional power line model.
[0017] The adaptive vegetation point cloud lightweight module is used to adaptively thin the density of the segmented vegetation point cloud according to its local geometric features, and generate a lightweight vegetation point cloud that retains key structures.
[0018] The tree barrier distance and risk assessment module is used to calculate the spatial distance between vegetation and power lines based on the three-dimensional power line model and the lightweight vegetation point cloud, identify potentially dangerous vegetation areas, and conduct risk assessment based on the extracted features.
[0019] The 3D visualization and interaction module is used to integrate the results of semantic segmentation, power line model, vegetation point cloud and risk assessment results to construct a 3D visualization scene that supports multi-level detail rendering and interactive query.
[0020] Thirdly, embodiments of this application provide an electronic device, including:
[0021] processor;
[0022] Memory used to store processor-executable instructions;
[0023] The processor is configured to implement, when executing the instructions, a lightweight semantic point cloud reconstruction and visualization method for power channels oriented towards tree barrier analysis as described in the first aspect.
[0024] Fourthly, embodiments of this application provide a computer-readable storage medium storing a program that instructs a device to execute the lightweight semantic point cloud reconstruction and visualization method for power channels oriented towards tree barrier analysis as described in the first aspect.
[0025] Compared with existing technologies, this invention has the following significant advantages: Through an adaptive lightweight strategy, while preserving the point cloud details of key areas in tree obstacle analysis (such as the canopy and near-line areas), it significantly reduces the overall data volume and improves processing and transmission efficiency. The lightweight network structure, combined with an attention mechanism, enhances the feature learning ability for specific targets such as power lines and vegetation, reducing model complexity while maintaining high segmentation accuracy, which is beneficial for practical deployment. The introduction of tower support point coordinates and mechanical models for reconstruction makes the generated 3D power line model more consistent with its actual physical form under tension and gravity, providing a more accurate benchmark for subsequent distance calculations. In addition to distance detection, it also integrates tree geometry and spatial features for risk level prediction; the constructed 3D semantic scene deep fusion analysis results support multi-level, multi-perspective interactive queries and historical backtracking, greatly improving the analytical depth and decision support capabilities of inspection data. Attached Figure Description
[0026] Figure 1 This is a schematic diagram of a lightweight semantic point cloud reconstruction and visualization method for power channels oriented towards tree obstacle analysis, provided as an embodiment of this application.
[0027] Figure 2 The system architecture diagram for lightweight semantic point cloud reconstruction and visualization of power channels for tree obstacle analysis provided in this application.
[0028] Figure 3 A schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation
[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.
[0030] It should be noted that in the embodiments of this application, "at least one" refers to one or more, and "more than one" refers to two or more. 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 application belongs. The terminology used in the specification of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application.
[0031] Based on the embodiments described in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0032] Example 1
[0033] Figure 1This is a schematic diagram illustrating a lightweight semantic point cloud reconstruction and visualization method for power channels based on tree obstacle analysis, provided as an embodiment of this application. Figure 1 As shown, a lightweight semantic point cloud reconstruction and visualization method for power channels oriented towards tree barrier analysis includes:
[0034] S1. Laser point cloud and image data from the power channel are simultaneously acquired using sensors mounted on a drone, and fused to generate an initial point cloud with texture. Its function is to provide a high-precision, textured initial 3D data foundation for the entire method. This step, through simultaneous acquisition and fusion of multiple sensors, combines the geometric accuracy of the laser point cloud with the color and texture information of the image, forming a unified data source for all subsequent analyses.
[0035] Specifically, in this embodiment, the step of synchronously acquiring laser point cloud and image data of the power channel using sensors mounted on the UAV, and fusing them to generate an initial point cloud with texture, specifically includes:
[0036] S1.1. Simultaneous acquisition of multi-source data via a drone platform. In this embodiment, a DJI Matrice M300 RTK drone equipped with a Zenmuse L1 LiDAR module and a Zenmuse P1 full-frame aerial survey camera is used for data acquisition. The LiDAR acquires point clouds at a rate of 240,000 points per second, the RGB-D camera simultaneously captures high-resolution color images and records depth information, and the IMU continuously records flight attitude data. The above three types of data are timestamped using a unified hardware clock to ensure strict synchronization of the data in time sequence.
[0037] S1.2. Joint calibration of the lidar, RGB-D camera, and IMU. This step is completed at a dedicated calibration field before the flight operation. A high-precision three-dimensional calibration target is used, which has multiple feature points (such as checkerboard corner points) with known precise dimensions and spatial distribution. The UAV is controlled to hover at different positions and angles, allowing all sensors to collect data from the calibration target from multiple perspectives. By analyzing the correspondence between the target feature points in the lidar point cloud and camera images, and combining nonlinear optimization algorithms (such as bundle adjustment), the precise rotation and translation parameters between the lidar coordinate system and the camera coordinate system, i.e., the spatial transformation extrinsic parameter matrix, are calculated. Simultaneously, by analyzing the angular velocity measured by the IMU at the same moment and the pixel movement of feature points in the camera image, the time delay between the IMU and the camera is precisely calibrated, obtaining microsecond-level time synchronization parameters to ensure the consistency of the spatiotemporal reference during subsequent data fusion.
[0038] S1.3. Based on the time synchronization parameters and the IMU motion data, motion distortion correction is performed on the laser point cloud sequence. During the acquisition of the original laser point cloud, due to the continuous flight of the UAV, the position and attitude of the sensors change within a single laser scan cycle, causing the geometry of objects in the point cloud to be stretched or distorted. This error is called motion distortion. This embodiment uses a tightly coupled inertial navigation calculation method for correction. Specifically, using calibrated time synchronization parameters, high-frequency IMU data (e.g., 200Hz) is strictly aligned with the emission time of each laser pulse from the lidar. Then, based on the instantaneous acceleration and angular velocity recorded by the IMU, the precise six-degree-of-freedom pose (position and orientation) of the UAV at the moment of laser point acquisition is calculated in real time through integration. Finally, using this instantaneous pose information, all laser points originally located in different instantaneous coordinate systems are transformed and reprojected onto a unified global coordinate system, thereby eliminating the geometric deformation introduced by the carrier motion and obtaining a corrected point cloud with accurate position and correct shape.
[0039] Specifically, motion distortion correction of laser point clouds can be achieved through the following mathematical model:
[0040] ,
[0041] in, Indicates the first The coordinates of a laser point in the local coordinate system of the lidar; This represents the corrected coordinates in the global coordinate system. Indicates the first The acquisition time of each laser point Indicates the reference start time. , represents the transformation matrix from the lidar coordinate system to the global coordinate system; This represents the rotation matrix obtained by IMU integration; Represents the translation vector, obtained from IMU acceleration data. By performing a second integral, we obtain the expression:
[0042] ,
[0043] in, To achieve zero bias in the accelerometer, These are the initial velocity and position, respectively. It is the integral variable.
[0044] S1.4. Using the spatial transformation extrinsic matrix, the corrected laser point cloud is registered and fused with the corresponding color image. This step aims to color the geometrically accurate laser point cloud. First, based on the image acquisition timestamp and the laser point timestamp, the closest one or more color images in time are matched for each frame of the corrected point cloud. Then, using the spatial transformation extrinsic matrix obtained in step (2.2), the coordinates of the three-dimensional laser points are transformed to the camera coordinate system. Next, using the camera's intrinsic parameter model (which describes parameters such as lens focal length and principal point), the three-dimensional points are projected onto the two-dimensional image plane to obtain the pixel coordinates of each laser point in the image. Through bilinear interpolation, red, green, and blue color values are sampled from the image pixels surrounding the pixel coordinates. Finally, the sampled RGB color values are assigned to the corresponding three-dimensional laser points to generate a colored point cloud that has both accurate three-dimensional geometric coordinates and realistic surface color texture, i.e., an initial point cloud with texture, providing key visual features for subsequent semantic recognition.
[0045] S2. Lightweight semantic segmentation is performed on the initial point cloud. A neural network incorporating local feature aggregation and attention mechanisms is used to identify and segment power line point clouds and vegetation point clouds. Its function is to intelligently understand complex 3D scenes and separate key targets. This step utilizes a lightweight neural network integrating local feature aggregation and attention mechanisms to efficiently and accurately identify and extract power line point clouds and vegetation point clouds, providing a prerequisite for subsequent targeted professional processing.
[0046] Specifically, in this embodiment, the neural network for lightweight semantic segmentation of the initial point cloud adopts a network architecture based on point cloud local neighborhood construction and feature aggregation, and includes an attention mechanism for enhancing the target feature response.
[0047] The neural network is implemented based on the improved RandLA-Net architecture and incorporates a channel attention module to achieve high-precision semantic segmentation with lightweight design.
[0048] First, an initial point cloud with texture is input. The network reduces the massive point cloud to about 1 / 4 of the original number of points through random sampling, serving as the input point set for subsequent layers, thus providing a lightweight design starting point. In each layer's feature encoding stage, for each sampled center point, K-nearest neighbor search (KNN) is used to quickly find its K nearest neighbors in space (K=16 in this embodiment), constructing a local neighborhood. Then, a local feature aggregation module processes each neighborhood: a small multilayer perceptron (MLP, a fully connected feedforward neural network) independently processes the coordinates and color features of each neighborhood point, and then max pooling is used to aggregate the features of all points within the neighborhood, obtaining a feature vector representing the local region. This process progresses layer by layer, constructing hierarchical features from details to abstractions.
[0049] Secondly, to enhance the recognition capability of key targets such as slender power lines and complex-shaped vegetation, this embodiment embeds a Squeeze-and-Excitation (SE) attention module at the encoder-decoder connection of the network. The module's workflow is as follows: First, the high-dimensional feature map output by the encoder is compressed spatially, and a global descriptor vector is obtained through global average pooling. Then, through an excitation process consisting of two fully connected layers, the correlation between each feature channel is learned, and a weight vector is generated. Finally, this weight vector is recalibrated by channel-level multiplication with the original feature map, automatically improving the response of feature channels important to the segmentation task and suppressing unimportant channels. This allows the network to focus more on the discriminative features of power lines and vegetation, such as the linear continuity of power lines and the color-texture patterns of vegetation.
[0050] Finally, the feature decoder gradually restores the spatial resolution of the point cloud through nearest neighbor interpolation upsampling and combines shallow features passed through skip connections, ultimately outputting the semantic label for each original point. This completes the task of segmenting the scene into multiple categories such as power lines, vegetation, towers, ground, and background, thereby accurately extracting power line point clouds and vegetation point clouds. The entire network model has fewer than 1 million parameters and can run in real time on drones or ground stations equipped with edge computing devices such as NVIDIA Jetson Xavier NX, meeting the engineering requirements of lightweight design and high efficiency.
[0051] S3. For the segmented power line point cloud, obtain the coordinates of the actual support points suspended at both ends of the tower. Based on a preset mechanical model, perform fitting and reconstruction to generate a three-dimensional power line model. Its function is to establish an accurate three-dimensional reference model of the line that conforms to the laws of engineering physics. This step, by introducing the actual support point coordinates of the tower and fitting the preset mechanical model, reconstructs the discrete point cloud into a continuous and realistic conductor shape, providing a highly reliable spatial reference for the core distance calculation.
[0052] Specifically, in this embodiment, obtaining the coordinates of the actual support points suspended at both ends of the tower in step S3 includes:
[0053] S3.1. Identify the power transmission tower point cloud from the initial point cloud based on the semantic segmentation result. In this embodiment, use the semantic label point cloud generated in step S2 to extract all point cloud clusters labeled as power transmission tower categories. Since the tower structure usually has regular and vertical characteristics, first use the Euclidean distance clustering algorithm, an algorithm that groups adjacent points according to the spatial distance between points, to cluster the point cloud belonging to the tower category and separate the point cloud sets belonging to each independent tower. For the situation where the point cloud of a single tower is divided into multiple small clusters due to occlusion or incomplete segmentation, through analyzing the continuity of each cluster in the vertical direction and the proximity of the projection positions on the horizontal plane, perform merging processing to finally obtain a complete tower point cloud object corresponding one by one to the actual towers on site.
[0054] S3.2. Determine the three-dimensional coordinates of the wire suspension points at the top of the tower by analyzing the structural characteristics of the tower point cloud. For each independent tower point cloud object, first determine the ground elevation at the bottom and the highest elevation at the top of the tower by statistically analyzing the height values of all its points in the vertical direction (Z-axis). Then, apply the plane fitting algorithm based on random sample consensus, an algorithm that can robustly fit a geometric model (such as a plane) from data containing noise (such as point clouds of branches and auxiliary equipment), to perform multiple samplings and plane model fittings on the point cloud in the top area of the tower point cloud. Since the crossarm (the horizontal structure supporting the conductors) of a high-voltage transmission tower is usually in an approximately horizontal state, the optimal plane fitted can be regarded as the plane where the crossarm is located. Finally, on this fitted plane, find the point located on the outermost side of the tower structure (along the line direction) and with the Y coordinate (corresponding to the direction perpendicular to the line) close to the center position of the tower body, and determine the average value of its three-dimensional coordinates as the three-dimensional coordinates of the wire suspension point on this side. For towers with multiple layers of crossarms such as dry-type or portal-type towers, identify the wire suspension points of each layer of crossarms according to this method layer by layer.
[0055] S3.3. Combining the power line alignment information, match the coordinates of the hanging points at both ends of each power line as the coordinates of the actual support points. After obtaining the coordinates of the hanging points at the top of all towers, combine the macroscopic alignment information of the power line (which can be extracted from the spatial position sequence of continuous towers). First, sort and number the towers according to their spatial distribution order (e.g., from substation A to substation B). Then, for each independent power line point cloud cluster extracted through semantic segmentation, calculate the main direction of its overall point cloud on the horizontal plane. Compare this main direction with the connecting direction of adjacent tower pairs, while considering the position of the power line point cloud cluster in three-dimensional space (e.g., located on the upper, middle, or lower layer of the tower), and match it with the two hanging points on the corresponding tower pair at the corresponding elevation level. The specific matching process uses a dual judgment of distance and direction consistency: the two ends of the power line point cloud cluster should be spatially closest to their two hanging points, and the overall alignment of the power line point cloud cluster should be basically consistent with the connecting direction of these two hanging points. The coordinates of the two successfully matched hanging points are determined as the actual support points of the power line.
[0056] Specifically, in this embodiment, step S3, which involves fitting and reconstructing a three-dimensional electric field model based on a preset mechanical model, specifically includes:
[0057] S3.4. The coordinates of the actual support points are used as boundary conditions. Specifically, for a power line within a span, the coordinates of its two actual support points (denoted as points A and B) serve as key constraints in the reconstruction process. The core of this constraint is that the final reconstructed three-dimensional power line curve must strictly pass through these two points that have been precisely measured in three-dimensional space. Before the model calculation begins, the coordinates are usually preprocessed, for example, by constructing a local coordinate system so that the projections of points A and B on a specific plane are aligned. This simplifies the complex three-dimensional curve fitting problem into a two-dimensional curve fitting problem within that plane, while simultaneously satisfying the spatial position condition of fixed ends.
[0058] S3.5. Using the catenary or parabolic equation as the mechanical model, curve fitting is performed on the electric field line point cloud. In this embodiment, the catenary equation is preferred to simulate the true shape of the conductor under its own weight. A catenary is a mathematical description of a flexible curve that sags naturally under gravity, and its shape is determined by specific parameters. For scenarios with higher computational efficiency requirements or relatively small sag, the parabolic equation can be used for approximate fitting. The fitting process involves projecting the electric field line point cloud onto the aforementioned two-dimensional fitting plane and using the planar coordinates of these projection points to approximate the selected mechanical model curve. The essence of this step is to find a continuous mechanical curve that best conforms to the distribution law of the discrete point cloud.
[0059] The fitted catenary function equation is as follows: The curve is fitted by minimizing the sum of squared perpendicular distances from points to the curve:
[0060] ,
[0061] in, The shape parameters of the catenary are related to the tension. and linear density related: ; These are the parameters for curve translation; Let be the projected coordinates of the power line point cloud in the fitting plane. The horizontal translation parameter of the catenary determines the horizontal position of the lowest point of the catenary. For the first The horizontal coordinates of each power line point; It is the acceleration due to gravity; For the number of power line point clouds, The vertical translation parameter of the catenary determines the vertical position of the lowest point of the catenary. For the first The vertical coordinates of the power line points.
[0062] The fitting process is as follows: 1. Coordinate projection: Project the three-dimensional electric field point cloud onto a vertical plane to obtain a two-dimensional point set. 2. Parameter solution: Solving for parameters using an optimization algorithm. , , This makes the catenary curve as close as possible to the point cloud distribution. Related to conductor tension and linear density; , Related to the location of the support points. If the coordinates of the support points at both ends of the power line are known... and Therefore, the fitted catenary must pass through these two points. and It will be constrained by these two boundary conditions. In actual fitting, the support point is often used as the initial condition or constraint.
[0063] S3.6. Solve for the parameters of the mechanical model using an optimization algorithm to generate a continuous three-dimensional electric field model. Specifically, the Levenberg-Marquardt optimization algorithm is used for parameter solving. This algorithm is a numerical method widely used to solve nonlinear least squares problems. First, a reasonable initial estimate of the parameters of the catenary or parabolic model is set based on the boundary conditions. The core of the algorithm is iterative optimization: in each iteration, the sum of squares of the vertical distances from all electric field projection points to the model curve under the current parameters (i.e., the objective function) is calculated, and the sensitivity of the objective function to each parameter (gradient information) is used to determine how to adjust the parameters to reduce this sum of squares. The algorithm intelligently balances two search strategies to ensure a fast and stable approximation of the optimal solution. When the iteration reaches the preset convergence condition (such as the change in the objective function value being less than a certain minimum threshold), a set of optimal model parameters is obtained. Substituting these parameters back into the model equation yields a curve that is accurate on the fitting plane. Transforming this curve back into the original three-dimensional world coordinate system ultimately generates a smooth, continuous three-dimensional electric field model that strictly passes through the two end support points.
[0064] S4. For the segmented vegetation point cloud, density is adaptively thinned based on its local geometric features to generate a lightweight vegetation point cloud that retains key structures. Its function is to perform task-oriented intelligent compression of vegetation data while ensuring analytical accuracy. This step adaptively adjusts the thinning strategy based on local geometric features, aiming to significantly reduce the amount of data while selectively preserving structural features crucial for tree barrier distance analysis.
[0065] Specifically, in this embodiment, the density thinning in step S4, which adaptively performs density thinning based on local geometric features, specifically includes:
[0066] S4.1 Calculate the local surface curvature of each point in the vegetation point cloud. In practice, the eigenvalue estimation method based on covariance analysis is used to calculate the local surface curvature of each point. The principle of this method is as follows: For each target point in the vegetation point cloud, the K-nearest neighbor algorithm is first used to search for a certain number (e.g., 30) of its nearest neighbors in space, forming a local neighborhood point set. Then, the three-dimensional coordinate covariance matrix of all points within this neighborhood is calculated. By performing eigenvalue decomposition on this covariance matrix, three eigenvalues are obtained in order of magnitude, reflecting the dispersion of the point set in the three principal directions. Among them, the smallest eigenvalue effectively reflects the flatness or curvature of the local surface along the normal direction. The curvature value can be calculated from these eigenvalues according to a specific formula. The larger the value, the more drastic the curvature change of the local surface (such as the edge of a leaf or the tip of a branch); conversely, the smaller the curvature value, the gentler the local surface (such as the surface of a tree trunk or flat ground).
[0067] Among them, point and Nearest neighbor set Local covariance matrix for:
[0068] ,
[0069] right Perform eigenvalue decomposition:
[0070] ,
[0071] curvature Defined as:
[0072] ,
[0073] in, The first in vegetation point clouds One point; Let be the centroid of the neighborhood point set; For point The neighborhood point set, These are the eigenvalues of the covariance matrix. , For the first eigenvectors and eigenvalues of ( ) =1,2,3), To prevent small positive numbers from being divided by zero, point The local curvature, This represents the index of the point cloud point currently being processed. Indicates the ordinal index of the feature value.
[0074] S4.2. Divide the point cloud region into high-curvature and low-curvature regions according to a preset curvature threshold. After obtaining the curvature values of all points, a threshold needs to be set for region division. In this embodiment, by statistically analyzing the curvature distribution of the entire vegetation point cloud (e.g., drawing a curvature histogram), the top 20% quantiles of the curvature values are selected as the threshold to divide the point cloud into high-curvature and low-curvature regions. For example, points with curvature values greater than the threshold can be classified as high-curvature regions. These points typically correspond to geometrically complex and detailed areas such as the outer edge of the canopy, leaf edges, and small branches. Points with curvature values less than or equal to the threshold are classified as low-curvature regions. These points typically correspond to relatively flat areas such as the main trunk, the surface of thick branches, and ground vegetation. The selection of the threshold requires a balance between the goal of detail preservation and data simplification, and can be pre-calibrated through experiments.
[0075] S4.3. A low thinning rate is applied to high-curvature areas, and a high thinning rate is applied to low-curvature areas. After partitioning, different uniform grid downsampling strategies are used to perform thinning operations on the two types of areas. For high-curvature areas, a smaller grid size (corresponding to a low thinning rate) is used for downsampling, for example, setting the grid side length to 0.05 meters. This operation retains a relatively large number of original points in the area, thereby maximizing the preservation of the fine geometric structure and edge contour information of the vegetation. For low-curvature areas, a larger grid size (corresponding to a high thinning rate) is used for downsampling, for example, setting the grid side length to 0.2 meters. This operation retains only one representative point (such as the centroid or the first point) in each grid, thereby significantly reducing the number of point clouds in flat areas and significantly reducing the overall data volume. Through this curvature-based adaptive processing, the vegetation point cloud is effectively lightweighted while ensuring the complete preservation of key morphological features.
[0076] Specifically, in this embodiment, the generation of a lightweight vegetation point cloud that preserves key structures in step S4 further includes the following steps:
[0077] S4.4 Before the thinning process, based on the semantic segmentation results, identify the near-line vegetation area within a certain range of the three-dimensional power line model.
[0078] In practice, the vegetation point cloud obtained after semantic segmentation and the reconstructed 3D power line model are first acquired. Near-line region identification employs a spatial buffer analysis method. Its core is calculating the shortest spatial distance from each vegetation point to the nearest power line model. This embodiment defines a distance threshold, such as 3 meters, as the boundary of the safety buffer. Next, all vegetation points are traversed, and the spatial Euclidean distance from their 3D coordinates to the power line model is calculated. All vegetation points with a distance less than the threshold are marked, and density-based spatial clustering algorithms are used to aggregate these spatially adjacent marked points into several continuous clusters. Each cluster represents a near-line vegetation region spatially adjacent to the power line. These areas are high-risk zones for tree obstruction, and the integrity of their point cloud is crucial for subsequent accurate distance calculations.
[0079] S4.5. For the near-line vegetation area, a thinning rate lower than or equal to that of the high curvature area is applied.
[0080] After identifying the near-line vegetation areas, a protective thinning strategy is employed. Specifically, these areas are processed first before the aforementioned curvature-based partitioned thinning. For each near-line vegetation area, regardless of whether its internal point cloud is classified as a high-curvature or low-curvature region based on curvature calculations, a uniformly low global thinning rate, not exceeding the grid size of the high-curvature region, is applied. For example, a very small grid size (e.g., 0.02 meters) is set for uniform grid downsampling, or in some applications where extreme accuracy is paramount, the thinning step is even skipped entirely, preserving the entire original point cloud of the area. This strategy ensures that the geometric details of the vegetation closest to the power lines and potentially posing a direct threat are preserved to the greatest extent possible in the final lightweight vegetation point cloud, thus providing a reliable data foundation for subsequent accurate distance calculations and risk assessments.
[0081] S5. Based on the three-dimensional power line model and the lightweight vegetation point cloud, calculate the spatial distance between the vegetation and the power lines, identify potentially hazardous vegetation areas, and conduct a risk assessment based on the extracted features. Its function is to achieve intelligent analysis from spatial distance quantification to comprehensive risk assessment. This step first calculates the precise spatial distance between the vegetation and the power lines to locate potential hazards, and then assesses the risk level based on the extracted multi-dimensional features, thereby providing decision support information that goes beyond simple threshold alarms.
[0082] Specifically, in this embodiment, step S5 identifies potentially hazardous vegetation areas and performs a risk assessment based on the extracted features, specifically including:
[0083] S5.1 Calculate the shortest distance from each point in the lightweight vegetation point cloud to the three-dimensional power line model, and mark all vegetation points with a distance less than a first safety threshold as primary danger points. In specific implementation, an efficient nearest neighbor search algorithm based on spatial indexing is used for calculation. This algorithm organizes the power line model by establishing a three-dimensional spatial grid (such as an octree), accelerating the distance query process. For each point in the lightweight vegetation point cloud, the algorithm quickly finds the nearest power line segment in space and calculates the three-dimensional Euclidean distance from the point to that segment. In this embodiment, a strict first safety distance threshold is pre-set based on the voltage level and operating procedures of the transmission line (for example, it may be set to 3.5 meters for a 220kV line). All vegetation points are traversed, and their calculated shortest distances are compared with this threshold. All vegetation points with a distance less than this threshold are filtered out and marked as primary danger points.
[0084] S5.2. Spatial clustering is performed on the primary hazard points to aggregate points belonging to the same vegetation object into a hazardous vegetation cluster. After obtaining the discrete primary hazard points, a density-based noise-applied spatial clustering algorithm is used for processing. This algorithm is based on a core idea: a cluster consists of the largest set of density-connected points, which can effectively discover clusters of arbitrary shapes and filter out noise points. In specific implementation, two key parameters are set: the neighborhood search radius and the minimum number of points required to form a cluster. The algorithm starts from any unprocessed primary hazard point, searches for its neighboring points within a given radius, and if the number of neighboring points reaches the minimum point requirement, the density of the area is considered sufficient, and a cluster is formed by expanding from this core; then, the same density check and expansion are recursively performed on the points in the neighborhood until the cluster can no longer expand. Finally, all primary hazard points that are spatially adjacent and belong to the same tree or vegetation clump will be aggregated into an independent hazardous vegetation cluster, while scattered, isolated points are regarded as noise and are removed.
[0085] S5.3. For each hazardous vegetation cluster, extract a set of assessment features. For each aggregated hazardous vegetation cluster, extract a set of multi-dimensional assessment features related to tree barrier risk. These features include: 1) Geometric features: such as the elevation of the highest point of the cluster (reflecting tree height), the volume of the cluster's three-dimensional axial bounding box (approximately reflecting tree size), and the location of the cluster's centroid. 2) Spatial relationship features: such as the average and minimum shortest distances from all points in the cluster to power lines, and the vertical distance (horizontal offset) of the cluster's centroid projected onto the nearest power line on the horizontal plane. 3) Density features: such as the average point density of the original point cloud forming the cluster, which can indirectly reflect the density of the vegetation. These features together constitute a quantitative description of the hazardous vegetation cluster.
[0086] S5.4. Based on the aforementioned assessment features, assess the risk level of the hazardous vegetation cluster. The risk assessment employs a gradient boosting decision tree model based on the XGBoost framework. This model is an ensemble learning algorithm that sequentially constructs a series of decision trees, each attempting to correct the prediction error of the previous tree. Finally, the predictions from all trees are weighted and combined to form a powerful predictive capability. During the model training phase, a large amount of historical tree obstacle case data (including the various features extracted above as input and manually labeled real risk levels as labels) is used to train the model. In this step, the assessment feature vector extracted from each hazardous vegetation cluster is input into the trained model. Internally, the model comprehensively judges the input features based on the learned complex nonlinear relationships, ultimately outputting a quantified risk level (e.g., divided into low, medium, and high risk levels). This risk level comprehensively reflects the urgency and potential severity of the threat posed by the vegetation cluster to the line.
[0087] S6. Integrate the semantic segmentation results, power line model, vegetation point cloud, and risk assessment results to construct a 3D visualization scene that supports multi-level detail rendering and interactive querying. Its function is to integrate all data, models, and analysis results into an interactive decision support platform. This step, by constructing a 3D visualization scene that supports multi-level detail rendering and interactive querying, enables the analysis results to be presented intuitively and explored in depth, completing a closed loop from data processing to business application.
[0088] Specifically, in this embodiment, the construction of a 3D visualization scene supporting multi-level detail rendering and interactive querying in step S6 includes:
[0089] S6.1. Spatial indexing and hierarchical organization are performed on the semantically segmented point cloud, 3D power line model, and lightweight vegetation point cloud to establish a multi-level detail structure. Specifically, a pyramid-style multi-level detail technique is used to construct the 3D scene. This technique achieves smooth rendering interaction by creating multiple resolution versions of the same scene data. First, the semantically segmented point cloud (such as ground and background) and lightweight vegetation point cloud are spatially indexed using an octree structure. An octree is a data structure that recursively divides 3D space into eight sub-cubes, efficiently managing massive point clouds. Based on this, multiple detail levels are generated from high resolution (original point cloud or lightweight point cloud) to low resolution (such as a meshed model) through layer-by-layer aggregation and downsampling. For the 3D power line model, its vector data is directly stored as an independent high-precision layer. Finally, a scene graph containing multiple LOD (Level of Detail) nodes is established. When the user's viewpoint is far away, the low-detail model is automatically loaded; when the viewpoint is zoomed in, the high-detail point cloud and the complete vector model are dynamically loaded and switched.
[0090] S6.2. Associate the risk level information from the risk assessment results with the corresponding hazardous vegetation clusters in three-dimensional space. In the system backend, create a unique data identifier for each hazardous vegetation cluster object. This identifier not only stores its three-dimensional geometric information (such as point cloud or bounding box) but also serves as a smart object container, associating all its analysis attributes. Specifically, the structured data calculated in step S5, such as the risk level (e.g., high risk), specific distance values (minimum clearance distance), the corresponding line section, tower number, and analysis timestamp, are strongly bound to the three-dimensional spatial location of the hazardous vegetation cluster (such as its bounding box or centroid coordinates) through this unique identifier. This association is typically achieved by adding custom attribute fields to spatial databases or scene graph nodes. Therefore, in the three-dimensional scene, each hazardous vegetation cluster is not only a visual geometry but also a data-spatial fusion entity carrying complete risk assessment information.
[0091] S6.3. Provide interactive query functionality in the visualization scene to display the associated attribute information and historical data comparison of the selected target. In the 3D visualization client, implement interactive query functionality based on light pickup or spatial range selection. When a user clicks or selects a hazardous vegetation cluster in the scene (which is usually rendered on the screen with a highlighted color or special icon), the client program immediately obtains the object's unique data identifier. Using this identifier, the client initiates a query request to the backend database, retrieving and displaying an attribute information panel in real time. This panel not only displays its current risk level, precise distance, and other core information, but also provides a historical comparison function button. Clicking this button retrieves all assessment records of the same spatial location (or the same tree) from the historical database within different inspection cycles. These records can be displayed side-by-side in the form of a timeline chart, intuitively presenting the distance change trend and risk evolution process of the hazard point, providing dynamic and long-term reference for operation and maintenance decisions.
[0092] Example 2
[0093] like Figure 2 As shown, this application provides a lightweight semantic point cloud reconstruction and visualization system architecture for power channels oriented to tree barrier analysis, which is applied to the lightweight semantic point cloud reconstruction and visualization system for power channels oriented to tree barrier analysis as described in Embodiment 1. It includes: a data acquisition and fusion module 210, a lightweight semantic segmentation module 220, a physical constraint power line reconstruction module 230, an adaptive vegetation point cloud lightweight module 240, a tree barrier distance and risk assessment module 250, and a three-dimensional visualization and interaction module 260.
[0094] The data acquisition and fusion module 210 is used to synchronously acquire laser point cloud and image data of the power channel through sensors carried by the UAV, and fuse them to generate an initial point cloud with texture.
[0095] The lightweight semantic segmentation module 220 is used to perform lightweight semantic segmentation on the initial point cloud. Through a neural network that includes local feature aggregation and attention mechanisms, it identifies and segments the power line point cloud and the vegetation point cloud.
[0096] The physical constraint power line reconstruction module 230 is used to obtain the actual support point coordinates of the segmented power line point cloud suspended at both ends of the tower, and to perform fitting and reconstruction based on a preset mechanical model to generate a three-dimensional power line model.
[0097] The adaptive vegetation point cloud lightweight module 240 is used to adaptively thin the density of the segmented vegetation point cloud according to its local geometric features, and generate a lightweight vegetation point cloud that retains key structures.
[0098] The tree barrier distance and risk assessment module 250 is used to calculate the spatial distance between vegetation and power lines based on the three-dimensional power line model and the lightweight vegetation point cloud, identify potentially dangerous vegetation areas, and conduct risk assessment based on the extracted features.
[0099] The 3D visualization and interaction module 260 is used to integrate the results of semantic segmentation, power line model, vegetation point cloud and risk assessment results to construct a 3D visualization scene that supports multi-level detail rendering and interactive query.
[0100] The system modules are connected sequentially, and the data flow is as follows: the initial point cloud output by the data acquisition and fusion module 210 is input to the lightweight semantic segmentation module 220; the segmented power line point cloud and vegetation point cloud are input to the physical constraint power line reconstruction module 230 and the adaptive vegetation point cloud lightweight module 240, respectively; the reconstructed power line model and lightweight vegetation point cloud are input to the tree barrier distance and risk assessment module 250; finally, all data and assessment results are input to the 3D visualization and interaction module 260 to complete scene construction and interactive display.
[0101] Figure 3 This is an electronic device provided in one embodiment of this application. For example... Figure 3 As shown, the electronic device includes at least the following components: processor 301 and memory 300, communication interface 303, and bus 302.
[0102] In this embodiment of the application, memory 300 is used to store executable instructions of processor 301, which, when configured to execute instructions, implements the method as described in the first aspect.
[0103] In embodiments of this application, a computer-readable storage medium includes instructions that instruct a device to perform the method as described in the first aspect. For example, the instructions instruct the device to perform... Figure 1 The method is shown in the process steps.
[0104] In one embodiment of this application, the program operating in the electronic device may be a program that controls a central processing unit (CPU) or similar device to achieve the functions of the above-described embodiments of the present invention (a program that enables the computer to function). Information processed by these systems is then temporarily stored in random access memory (RAM) during processing, and subsequently stored in various ROMs such as read-only memory (FlashROM) and hard disk drives (HDDs), and read, corrected, and written by the CPU as needed.
[0105] It should be noted that a portion of the electronic device described above can also be implemented using a computer. In this case, the program for implementing the control function can be recorded on a computer-readable recording medium, and the program recorded on the recording medium can be read into the computer and executed.
[0106] It should be noted that the computer mentioned here refers to a computer built into an electronic device, employing hardware including an operating system and peripheral devices. Furthermore, computer-readable recording media refers to removable media such as floppy disks, magneto-optical disks, ROMs, and CD-ROMs, as well as storage systems such as hard drives built into the computer.
[0107] Furthermore, computer-readable recording media can include: media that dynamically stores programs for short periods of time, such as communication lines used when transmitting programs via networks like the Internet or communication lines like telephone lines; and media that store programs for fixed periods of time, such as volatile memory inside a computer that serves as a server or client in this case. In addition, the aforementioned program can be a program used to implement the above-mentioned functions, or it can be a program that can implement the above-mentioned functions by combining them with programs already recorded in the computer.
[0108] Furthermore, the electronic device in the above embodiments can also be implemented as an assembly (system group) composed of multiple systems. Each system constituting the system group can possess some or all of the functions or functional blocks of the electronic device in the above embodiments. As a system group, it is sufficient to have all the functions or functional blocks of the electronic device.
[0109] Those skilled in the art should recognize that the above embodiments are only used to illustrate this application and are not intended to limit this application. Any appropriate changes and variations made to the above embodiments within the essential spirit and scope of this application fall within the scope of protection claimed in this application.
Claims
1. A lightweight semantic point cloud reconstruction and visualization method for power transmission channels oriented towards tree obstacle analysis, characterized in that, Includes the following steps: The drone uses sensors to simultaneously collect laser point cloud and image data of the power channel, and then fuses them to generate an initial point cloud with texture. The initial point cloud is subjected to lightweight semantic segmentation. Through a neural network that includes local feature aggregation and attention mechanisms, power line point clouds and vegetation point clouds are identified and segmented. For the segmented power line point cloud, the coordinates of the actual support points suspended at both ends of the tower are obtained, and a three-dimensional power line model is generated by fitting and reconstructing based on a preset mechanical model. The segmented vegetation point cloud is adaptively thinned according to its local geometric features to generate a lightweight vegetation point cloud that retains key structures. Based on the three-dimensional power line model and the lightweight vegetation point cloud, the spatial distance between the vegetation and the power lines is calculated, potentially dangerous vegetation areas are identified, and risk assessment is performed based on the extracted features. By integrating the results of semantic segmentation, power line model, vegetation point cloud, and risk assessment, a 3D visualization scene supporting multi-level detail rendering and interactive querying is constructed.
2. The method according to claim 1, characterized in that, The step of synchronously acquiring laser point cloud and image data of the power channel using sensors mounted on a drone, and fusing them to generate an initial point cloud with texture, specifically includes: Multi-source data is collected synchronously through a drone platform. The multi-source data includes: a laser point cloud sequence collected by a solid-state lidar, a high-resolution color image sequence with depth information collected by an RGB-D camera, and IMU motion data collected by an inertial measurement unit. The lidar, RGB-D camera, and IMU are jointly calibrated to obtain the spatial transformation extrinsic matrix and time synchronization parameters between the sensors. Based on the time synchronization parameters and the IMU motion data, motion distortion correction is performed on the laser point cloud sequence; Using the spatial transformation extrinsic matrix, the corrected laser point cloud is registered and fused with the color image at the corresponding time, thereby giving the laser point cloud texture information and generating an initial point cloud with texture.
3. The method according to claim 1, characterized in that, The neural network that performs lightweight semantic segmentation on the initial point cloud adopts a network architecture based on point cloud local neighborhood construction and feature aggregation, and includes an attention mechanism to enhance the target feature response.
4. The method according to claim 1, characterized in that, Obtaining the coordinates of the actual support points suspended at both ends of the tower specifically includes: Based on the semantic segmentation results, the power pole point cloud is identified from the initial point cloud; By performing structural feature analysis on the point cloud of the tower, the three-dimensional coordinates of the hanging point at its top are determined; By combining the route information of the power lines, the coordinates of the hanging points at both ends of each power line are matched as the coordinates of the actual support points.
5. The method according to claim 1 or 4, characterized in that, The process of fitting and reconstructing a three-dimensional electric field model based on a preset mechanical model specifically includes: The actual support point coordinates are used as boundary conditions; The catenary or parabolic equation is used as the mechanical model to perform curve fitting on the electric field line point cloud; The parameters of the mechanical model are solved by an optimization algorithm to generate a continuous three-dimensional electric field model.
6. The method according to claim 1, characterized in that, The adaptive density thinning based on its local geometric features specifically includes: Calculate the local surface curvature of each point in the vegetation point cloud; The point cloud region is divided into a high curvature region and a low curvature region based on a preset curvature threshold. A low thinning rate is used for high curvature regions, and a high thinning rate is used for low curvature regions.
7. The method according to claim 6, characterized in that, The generation of a lightweight vegetation point cloud that preserves key structures also includes the following steps: Before the thinning process, based on the semantic segmentation results, the near-line vegetation area within a certain range of the three-dimensional power line model is identified; The near-line vegetation area is treated with a thinning rate lower than or equal to that of the high curvature area.
8. The method according to claim 1, characterized in that, The process of identifying potentially hazardous vegetation areas and conducting risk assessments based on extracted features specifically includes: Calculate the shortest distance from each point in the lightweight vegetation point cloud to the three-dimensional power line model, and mark all vegetation points whose distance is less than the first safety threshold as primary danger points; Spatial clustering is performed on the primary hazard points, and points belonging to the same vegetation object are grouped into a hazard vegetation cluster; For each hazardous vegetation cluster, extract a set of assessment features; Based on the aforementioned assessment characteristics, the risk level of the hazardous vegetation cluster is assessed.
9. The method according to claim 1, characterized in that, The construction of a 3D visualization scene that supports multi-level detail rendering and interactive querying specifically includes: Spatial indexing and hierarchical organization are performed on the semantically segmented point cloud, the 3D power line model and the lightweight vegetation point cloud to establish a multi-level detailed structure. The risk level information in the risk assessment results is associated with the corresponding dangerous vegetation clusters in three-dimensional space; Interactive query functionality is provided in the visualization scenario to display the associated attribute information and historical data comparison of the selected target.
10. A lightweight semantic point cloud reconstruction and visualization system for power transmission channels oriented towards tree obstacle analysis, characterized in that, The system includes: The data acquisition and fusion module is used to synchronously acquire laser point cloud and image data of the power channel through sensors carried by the UAV, and fuse them to generate an initial point cloud with texture. The lightweight semantic segmentation module is used to perform lightweight semantic segmentation on the initial point cloud. It identifies and segments the power line point cloud and vegetation point cloud through a neural network that includes local feature aggregation and attention mechanisms. The physical constraint power line reconstruction module is used to obtain the actual support point coordinates of the power line point cloud suspended at both ends of the tower, and to perform fitting and reconstruction based on a preset mechanical model to generate a three-dimensional power line model. The adaptive vegetation point cloud lightweight module is used to adaptively thin the density of the segmented vegetation point cloud according to its local geometric features, and generate a lightweight vegetation point cloud that retains key structures. The tree barrier distance and risk assessment module is used to calculate the spatial distance between vegetation and power lines based on the three-dimensional power line model and the lightweight vegetation point cloud, identify potentially dangerous vegetation areas, and conduct risk assessment based on the extracted features. The 3D visualization and interaction module is used to integrate the results of semantic segmentation, power line model, vegetation point cloud and risk assessment results to construct a 3D visualization scene that supports multi-level detail rendering and interactive query.