Power distribution equipment point cloud multi-target classification extraction method, system, equipment and medium

Abstract

The invention discloses a power distribution equipment point cloud multi-target classification extraction method, system, equipment and medium, and belongs to the technical field of power distribution line three-dimensional visualization, and the method comprises the steps: obtaining three-dimensional laser point cloud data covering a power distribution line region, and carrying out the noise point elimination and spatial density resampling, and obtaining first point cloud data; extracting ground points in the first point cloud data, and constructing a digital elevation model; removing the ground points to obtain second point cloud data containing the power distribution equipment and the environment target; executing clustering processing, and extracting structural feature parameters of each cluster; inputting the extracted structural feature parameters into a pre-trained classification model to complete classification; and correcting a classification result, and outputting a power distribution equipment point cloud classification data set with a category label. The method has the beneficial effects that in the ground point identification and elimination stage, a sliding window area consistent with the distribution line layout characteristics is constructed, an initial ground point set is identified by using a height gradient trend, and a misjudgment point set with obvious terrain disturbance or local ground characteristics is screened out through joint verification of a fitting residual error and a ground structure template, so that the accuracy of the ground point identification and elimination is improved. And the stability of subsequent structure extraction and the precision of ground object separation are enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power distribution line three-dimensional visualization, and particularly relates to a power distribution equipment point cloud multi-target classification extraction method, system, device and medium. BACKGROUND

[0002] As an important part of the power system, the power distribution line undertakes the task of transmitting electric energy from the substation to each terminal user, and its safe and stable operation has an important influence on the overall efficiency and power supply reliability of the power system. With the gradual popularization of the application of three-dimensional laser radar (LiDAR) technology, using point cloud data to carry out three-dimensional modeling, environment perception and structure recognition of power distribution lines and their auxiliary equipment has become one of the important research directions of power distribution intelligent operation and maintenance.

[0003] At present, in the processing and target recognition of point cloud data, there are a large number of research and engineering practice results widely used in urban modeling, automatic driving, unmanned aerial vehicle inspection and other fields. These methods mostly rely on clustering and classification algorithms based on low-order features such as geometric shape, curvature, normal vector and density, or general point cloud target recognition network models based on deep learning. These methods have certain applicability in standardized environments, but still face many limitations in the power distribution line scene:

[0004] The towers, wires, buildings and vegetation in the power distribution line are often linearly laid out or spatially staggered, and the height distribution and structural size difference are small, so it is difficult for general clustering methods to accurately separate them, and it is easy to cause misclassification problems such as wires and buildings, towers and trees.

[0005] Existing classification methods based on point cloud depth feature learning mostly train on public general data sets, lack of feature construction and data sample accumulation for power distribution equipment forms, and are difficult to accurately identify the spatial feature distribution of typical structures.

[0006] Traditional point cloud classification methods usually process point sets independently, ignoring the structural continuity that should exist between devices (such as continuous connection of wires to towers, ground attachment of building bottom surfaces, etc.), which can easily cause classification boundary breakage, class jump and other problems, which is not conducive to three-dimensional modeling and subsequent structured management.

[0007] Therefore, in view of the problems of high target structure similarity, complex point cloud distribution and strong spatial relationship constraints in the power distribution line scene. SUMMARY

[0008] In view of the above problems, the present application is proposed.

[0009] Therefore, the technical problem solved by the present application is: how to construct a classification mechanism with physical structure constraints in the power distribution line point cloud data, in the face of high similarity of device structure, overlapping of target space, lack of industry adaptability of general methods, to realize accurate segmentation of target point cloud, spatial relationship recognition, category boundary analysis, structure interference removal, misjudgment type isolation and high-level separation processing.

[0010] To solve the above technical problems, the present application provides the following technical solutions: a power distribution equipment point cloud multi-target classification extraction method, comprising,

[0011] Obtain three-dimensional laser point cloud data covering the power distribution line area;

[0012] Perform noise point removal and spatial density resampling on the obtained point cloud data to obtain first point cloud data;

[0013] Extract ground points in the first point cloud data based on height gradient changes, and construct a digital elevation model based on the extracted ground points;

[0014] Compare the digital elevation model with the first point cloud data, remove the ground points, and obtain second point cloud data containing power distribution equipment and environmental targets;

[0015] Perform clustering processing on the second point cloud data, and extract structure feature parameters of each cluster according to shape features, axial length and density distribution of the point set;

[0016] Input the extracted structure feature parameters into a pre-trained classification model to obtain the device category label corresponding to each cluster, and complete the classification of tower point, conductor point, building point and vegetation point;

[0017] Verify the spatial structure relationship of the classification result, correct the classification result, and output the power distribution equipment point cloud classification data set with the category label.

[0018] As a preferred scheme of the power distribution equipment point cloud multi-target classification extraction method, wherein: obtaining the first point cloud data comprises dividing the point cloud data into multiple strip segments according to the scanning track of the laser radar;

[0019] In adjacent strip segments, identify point sets with horizontal projection coincidence degree exceeding a predetermined ratio and vertical height difference less than a specified distance threshold, and delete the point sets;

[0020] Construct a main strike direction vector set of the power distribution line in each strip segment, and project the three-dimensional coordinates of each point to its corresponding direction vector;

[0021] Delete point sets with an angle between adjacent direction vectors exceeding a set angle threshold;

[0022] In the local spatial region, the normal vector angle between each point and its neighborhood point is calculated, the point set whose standard deviation of the angle exceeds the specified range is identified and deleted;

[0023] Based on the spatial arrangement of the distribution pole and the conductor structure, the directional envelope interval is constructed, the point set located in the envelope interval is retained, and the point set located outside the interval is deleted;

[0024] An anisotropic voxel grid resampling operation is performed on the retained point cloud data, and the sampling interval in the conductor direction is set to be half of the transverse sampling interval, forming the first point cloud data.

[0025] As a preferred scheme of the power distribution equipment point cloud multi-target classification extraction method, the digital elevation model is constructed based on the extracted ground points, including dividing the first point cloud data into multiple equidistant sliding window regions along the distribution line direction;

[0026] In each sliding window, the sub-regions are divided by grid, and the lowest height value of the points in each sub-region is calculated;

[0027] According to the height gradient change trend between the lowest height values in the adjacent grid, the region with continuous change characteristics is identified as the initial ground point candidate set;

[0028] In the initial candidate set, the abnormal point set with a height fluctuation amplitude exceeding a predetermined range in its neighborhood is deleted;

[0029] The retained point set is subjected to a surface fitting interpolation process based on moving weighted average, forming a digital elevation model.

[0030] As a preferred scheme of the power distribution equipment point cloud multi-target classification extraction method, the second point cloud data containing power distribution equipment and environmental targets is obtained, including dividing the first point cloud data into multiple corresponding local subsets according to the sliding window region in the digital elevation model construction process;

[0031] In each local subset, each point is projected to the corresponding position of the digital elevation model along the vertical direction, and the vertical height difference value of the point is calculated;

[0032] In the points with a height difference value less than a set threshold, the fitting residual average value between the point and its neighborhood point and the fitted surface of the digital elevation model is calculated;

[0033] For the points that meet both the height difference value threshold condition and the residual average value threshold condition, a structure template matching comparison operation is performed to identify whether they belong to the ground-attached structure point set of the power distribution equipment;

[0034] The points that do not belong to the ground-attached structure point set are deleted to form the second point cloud data containing power distribution equipment and environmental targets.

[0035] As a preferred scheme of the power distribution equipment point cloud multi-target classification extraction method, the second point cloud data is divided into multiple point cluster groups according to the spatial distribution relationship.

[0036] For each cluster group, a minimum bounding box containing the point set is constructed, and the length, width and height of the bounding box in the main axis direction are calculated.

[0037] According to the three-dimensional size ratio of the bounding box, the shape compactness parameter of the point set is calculated.

[0038] A fixed-pitch voxel grid is constructed inside the bounding box, and the number of points contained in each voxel is counted to form a voxel density distribution matrix.

[0039] The distribution uniformity index of the density matrix in the main axis direction is calculated as the density directionality parameter.

[0040] The bounding box size parameter, shape compactness parameter and density directionality parameter are combined to form a structure feature parameter set corresponding to the cluster group.

[0041] As a preferred scheme of the power distribution equipment point cloud multi-target classification extraction method, the pre-trained classification model includes performing feature vector standardization processing on the structure feature parameters of each cluster, and scaling the bounding box length, width and height to a preset interval range.

[0042] The three-dimensional size ratio of the bounding box is calculated as the shape compactness parameter, and the point number difference standard deviation of adjacent voxels in the main axis direction is calculated as the density direction change index.

[0043] The standardized bounding box parameters, shape compactness parameters and density direction change indexes are combined in a fixed order to form an input feature vector.

[0044] The input feature vector is input into a classification model trained based on typical equipment sample data of the power distribution line.

[0045] The classification model matches the position of the input feature vector in the training feature space with the label space of the labeled sample to determine the equipment category label corresponding to the cluster.

[0046] The equipment category label is bound with the index of the corresponding point set to form a point cloud cluster classification result with a classification label.

[0047] As a preferred scheme of the power distribution equipment point cloud multi-target classification extraction method, the modified classification result includes dividing the classified point set into multiple continuous spatial sub-sections based on the direction of the power distribution line.

[0048] In each sub-section, identify the point set labeled as a tower as a structural anchor point, and extract the conductor category point set on both sides of the strike direction;

[0049] Determine whether each conductor category point set has a connection trajectory that is height-continuous and spatially parallel between the two tower anchor points in front and back, and if not, mark its classification result as abnormal;

[0050] For the conductor point set marked as abnormal, analyze the projection direction distribution and the boundary overlap with the adjacent building category point set, and if there is an edge nesting relationship, replace the conductor category label with the building category;

[0051] In the building and vegetation classification point set, extract the point density directionality distribution index and the ground projection area ratio of adjacent point sets, and if the continuous structure determination rule is met, merge them into the same category label;

[0052] Bind all the verified point sets and the corrected category labels to output the power distribution equipment point cloud classification data set with classification labels.

[0053] The present application provides a power distribution equipment point cloud multi-target classification extraction system.

[0054] To solve the above technical problems, the present application provides the following technical scheme: a power distribution equipment point cloud multi-target classification extraction system, comprising: a point cloud acquisition module for acquiring three-dimensional laser point cloud data covering the power distribution line area;

[0055] A preprocessing module is used to remove noise points and resample the spatial density of the acquired point cloud data to generate first point cloud data;

[0056] An elevation modeling module is used to extract ground points based on the height gradient change of the first point cloud data, and to construct a digital elevation model according to the extracted ground points;

[0057] A ground removal module is used to compare the digital elevation model with the first point cloud data, delete the ground points, and generate second point cloud data containing power distribution equipment and environmental targets;

[0058] A clustering extraction module is used to perform clustering operations on the second point cloud data, and to extract structural feature parameters based on the shape features, axial size and spatial density distribution of each cluster point set;

[0059] A classification recognition module is used to input the structural feature parameters into a pre-trained classification model to identify the equipment category label corresponding to each cluster, and to complete the classification of towers, conductors, buildings and vegetation;

[0060] A relationship verification module is used to perform consistency verification and label correction on the classification results based on the spatial structure relationship between the equipment.

[0061] An output module is configured to output the power distribution equipment point cloud classification dataset with the category label.

[0062] The application provides a computer device, including a memory and a processor, and the memory stores a computer program.

[0063] The application provides a computer readable storage medium, which stores a computer program.

[0064] The application has the beneficial effects that in the ground point identification and elimination stage, a sliding window region consistent with the power distribution line layout features is constructed, the initial ground point set is identified by using the height gradient trend, and the false point set with obvious terrain disturbance or local ground-adhesion features is screened out through the joint verification of fitting residual and ground-adhesion structure template, so that the stability of structure extraction and the accuracy of ground object separation are enhanced.

[0065] In the clustering and structure feature extraction process, the application does not rely on the spatial similarity judgment of the general point cloud algorithm, but takes the boundary box size ratio, voxel density distribution law and density change directionality as the classification basis to extract the classification input vector with consistent physical scale and structure direction features, so that the classification feature set established can effectively distinguish the equipment class targets with mutual overlap in geometric scale, and the recognition ability of structure recognition is improved.

[0066] The classification model used is trained based on the structure samples of the power distribution scene, the input features of which strictly correspond to the spatial geometric configuration and density characteristics of the power distribution equipment, avoiding the weakness of general deep models that are difficult to adapt to industry differences, and ensuring the effective separation ability and labeling accuracy of the classification model between high similar targets such as towers, wires, buildings and vegetation.

[0067] After classification, a structure consistency verification mechanism is further introduced to extract the tower anchor point information from the continuous trend of the power distribution line, verify whether the wire forms a continuous track in the front and rear structures, and complete the elimination and correction of abnormal jump labels through the building ground-adhesion attribute and point density directionality judgment, effectively constrain the spatial logical consistency of the classification result, avoid the spread of misjudgment, and ensure the integrity of the classification result. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings needed to be used in the embodiments will be briefly introduced as follows. Obviously, the drawings in the following description are only some embodiments of the present application, and all other drawings obtained by those skilled in the art without creative effort based on these drawings should also fall within the protection scope of the present application.

[0069] Figure 1 A general flowchart of a power distribution equipment point cloud multi-target classification extraction method is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0070] In order to make the above-mentioned purposes, features and advantages of the present application more apparent and easy to understand, the specific embodiments of the present application will be described in detail below with reference to the drawings. Obviously, the described embodiments are only some embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative effort should also fall within the protection scope of the present application.

[0071] Embodiment 1, refer to Figure 1 For an embodiment of the present application, the embodiment provides a power distribution equipment point cloud multi-target classification extraction method, which comprises the following steps:

[0072] Step 101: Obtain three-dimensional laser point cloud data covering the area of the power distribution line;

[0073] Step 102: Perform noise point elimination and spatial density resampling on the obtained point cloud data to obtain first point cloud data;

[0074] Step 103: Extract ground points in the first point cloud data based on height gradient change, and construct a digital elevation model based on the extracted ground points;

[0075] Step 104: Compare the digital elevation model with the first point cloud data, eliminate ground points, and obtain second point cloud data containing power distribution equipment and environmental targets;

[0076] Step 105: Perform clustering processing on the second point cloud data, and extract structural feature parameters of each cluster according to shape features, axial length and density distribution of the point set;

[0077] Step 106: Input the extracted structural feature parameters into a pre-trained classification model to obtain a device category label corresponding to each cluster, and complete the classification of tower point, conductor point, building point and vegetation point;

[0078] Step 107: Perform spatial structure relationship verification on the classification result, correct the classification result, and output a power distribution equipment point cloud classification data set with a category label.

[0079] In the actual operation and maintenance scene of the power distribution line, the three-dimensional laser point cloud data usually contains a large number of non-structure points and environmental interference targets, has complex characteristics such as uneven density distribution, mutual occlusion of structures, similar device forms, and overlapping spatial positions. The traditional point cloud processing method often cannot effectively identify the boundary relationship between structures such as towers and vegetation, conductors and buildings, especially in the area where the ground structure and the terrain height change obviously, the ground points are often confused with low targets, resulting in misclassification. In addition, due to the lack of spatial prior modeling of the power distribution scene in the general clustering and classification model, it is difficult to meet the engineering requirements of device fine identification and label consistency. Therefore, it is necessary to construct a classification extraction method that integrates the structural characteristics of power distribution equipment, spatial layout rules and physical connection constraints to accurately separate and identify various power distribution facility targets from the original point cloud.

[0080] In the specific application process, when the method is used to process the measured point cloud data of the power distribution line, first, the point cloud in the coverage range is divided into several strips, noise suppression and directionality resampling are performed respectively, and non-structure points caused by laser scanning errors and abnormal reflections are eliminated; after constructing a terrain model through a sliding window, the ground points are removed in combination with the power distribution facility ground template, while low targets such as ground cabinets and cable channels are retained; then the structural characteristics of the point set are extracted through spatial clustering, and the feature vector is entered into the trained classification model to complete the label annotation of the typical structure. Finally, the continuity rule of the tower anchor point and the conductor path is used to verify the label boundary, correct the jumping points or residual misjudgment area, and output the structured point cloud classification results for power distribution digital modeling. The classification extraction process provided in the embodiment can complete the automatic extraction and identification of power distribution equipment in the point cloud without relying on external geographic information or manual intervention, and is suitable for typical application scenarios with high requirements for point cloud classification efficiency and accuracy in engineering deployment environment.

[0081] Embodiment 2, which is an embodiment of the present application, provides a power distribution equipment point cloud multi-target classification extraction method based on the above embodiment, comprising:

[0082] Specifically, in step 101, three-dimensional laser point cloud data covering the power distribution line area is obtained, and the specific implementation manner is as follows:

[0083] A multi-rotor unmanned aerial vehicle platform equipped with a rotating multi-line laser radar system is used to carry out point cloud data acquisition. The flight platform has the capabilities of autonomous route planning, real-time attitude control and navigation positioning, and the laser radar carried thereon has high frequency and high precision ranging and scanning capabilities, and can realize continuous and stable spatial data acquisition within the route range arranged along the power distribution line. The sensor should meet the requirements of obtaining three-dimensional echo points with continuous distribution, stable reflection and high detail resolution, so as to cover typical power distribution environment targets such as towers, conductors, power auxiliary equipment, building outlines and vegetation.

[0084] The flight route is planned according to the distribution line direction before the flight task, the main flight route and the auxiliary compensation flight route are selected to cover the target area, and the flight route spacing is adjusted to ensure that the laser scanning overlap rate meets the continuous sampling requirements of the region. The flight height, speed and scanning angle are set according to the line density, equipment height and terrain undulation to ensure that the point cloud echo density has clear shape profile in the structure scale range. The flight attitude and time information are recorded during the aerial survey process, and the position trajectory is recorded through the inertial measurement and satellite positioning system to assist the three-dimensional positioning of the point and the registration of the post-processing.

[0085] After the point cloud data collection is completed, scanning time correction, attitude trajectory back calculation, echo synchronization processing and geometric registration operations are performed to generate original point cloud data with geographic space reference coordinates. After the preliminary data processing is completed, the data is regionally cut according to the distribution line direction and the engineering layout boundary to form an input data set for subsequent point cloud noise processing, structure identification and classification processing.

[0086] Further, in step 102, the acquired point cloud data is subjected to noise point elimination and spatial density resampling to obtain first point cloud data, and the specific implementation is as follows:

[0087] The point cloud data is divided into multiple strip segments according to the scanning trajectory of the laser radar;

[0088] In adjacent strip segments, a point set with a horizontal projection coincidence degree exceeding a predetermined ratio and a vertical height difference less than a specified distance threshold is identified and deleted;

[0089] A main direction vector set of the distribution line is constructed in each strip segment, and the three-dimensional coordinates of each point are projected onto the corresponding direction vector;

[0090] A point set with an angle exceeding a set angle threshold with respect to an adjacent direction vector is deleted;

[0091] In a local spatial region, the normal vector included angle between each point and its neighborhood points is calculated, a point set with a standard deviation of the included angle exceeding a specified range is identified and deleted;

[0092] A direction envelope interval is constructed based on the spatial arrangement of the distribution tower and the wire structure, a point set located in the envelope interval is retained, and a point set located outside the interval is deleted;

[0093] An anisotropic voxel grid resampling operation is performed on the retained point cloud data, the sampling interval in the wire direction is set to be half of the transverse sampling interval, and the first point cloud data is formed.

[0094] Specifically, in the execution of the point cloud noise removal operation based on the direction feature and the normal distribution, combined with the characteristics of the tower in the distribution line being vertically linearly distributed, the conductor being continuously extended along the tension direction, the building being blocky distributed and having clear edges, the vegetation having unstable normal and diffused distribution, a structure constraint identification strategy is adopted for target filtering. In the process of constructing the direction vector of the strip segment, the fitting trajectory is generated according to the main direction of the distribution line, the unit direction vector sequence is constructed, and the three-dimensional position vector of each point is projected to the direction vector to calculate the included angle parameter. If the included angle is greater than the engineering experience value (for example, 15°), it is considered that the point does not meet the linear layout requirement and is removed from the main direction structure. The strategy realizes the direction consistency removal through the structure space logic, which is essentially different from the general point cloud edge judgment method.

[0095] In the local normal analysis process, aiming at the problems that the surface normal direction of the electric pole and the conductor in the distribution scene is consistent, the building boundary is relatively flat, and the vegetation has large angle disturbance, the fixed search radius is used to calculate the point normal vector, and the standard deviation of the neighborhood point normal included angle set is used for distribution evaluation. When the standard deviation is greater than the set fluctuation threshold (for example, 12°), it is considered that the point normal fluctuation is violent, which may belong to the tree crown, the shelter or the abnormal reflection body, and should be removed. This processing strategy not only depends on the general features of the algorithm, but also uses the domain knowledge of the stability of the surface of the distribution structure target.

[0096] The direction envelope interval is constructed based on the physical structure constraint of the distribution line. According to the structure center line of the tower trunk, the conductor is suspended on the tension path between the towers, and the envelope box is constructed based on the center line as the reference, and the envelope thickness is set in the main direction and the transverse and vertical directions. In typical distribution applications, the direction fluctuation of the tower is small, and the suspension direction of the conductor is limited by the tension and has a clear orientation, so the target point set in the structure distribution interval can be accurately identified by this method, which is significantly different from the general target extraction method in complex natural environment where the structure direction is not clear.

[0097] In the voxel resampling stage, considering the characteristics of the conductor structure being slender and extending in the direction of the line, an anisotropic sampling strategy is adopted. The sampling interval is set to be half of the transverse direction in the direction of the line, which improves the sampling density and enhances the continuous expression ability of the conductor shape while keeping the overall point cloud size controllable. The size of the voxel grid is selected in combination with the structure size of the distribution equipment.

[0098] Further, in step 103, ground points in the first point cloud data are extracted based on height gradient changes, and a digital elevation model is constructed based on the extracted ground points, and the specific implementation is as follows:

[0099] The first point cloud data is divided into a plurality of equidistant sliding window regions along the direction of the distribution line;

[0100] In each sliding window, the sub-region is divided by grid, and the minimum height value of each sub-region is calculated;

[0101] According to the height gradient change trend between the minimum height values in adjacent grids, the region with continuous change characteristics is identified as the initial ground point candidate set;

[0102] In the initial candidate set, delete the abnormal point set whose neighborhood exists height fluctuation amplitude exceeding the predetermined range;

[0103] Perform surface fitting interpolation processing based on moving weighted average on the reserved point set to form a digital elevation model.

[0104] In this embodiment, in order to realize the digital elevation model construction oriented to the structural characteristics of power distribution lines, first, the first point cloud data is divided into equidistant sliding window regions according to the main direction of the power distribution lines. The length of each window is set to 20 meters, and the sliding step is 10 meters. This division method refers to the spatial periodicity of the tower layout, which ensures that the complete structure of the minimum spatial unit is contained in the single window.

[0105] Each sliding window is further divided into regular two-dimensional grid regions, and the grid length is set to 0.5 meters. For the point set in each grid, the lowest elevation value is extracted to generate a window lowest elevation matrix. In order to identify the continuous flat region, the difference value of the lowest elevation values of adjacent grids is used as the slope judgment index. When the difference value is lower than the set threshold (such as 0.2 meters), the point corresponding to the grid is marked as the initial ground point candidate.

[0106] In the vicinity of the initial candidate point, a local neighborhood is defined, the search radius is set to 0.5 meters, and the neighborhood point set is extracted. The height value of the point set is analyzed, and the height standard deviation σ is calculated, which is calculated as follows:

[0107]

[0108] Wherein, n is the number of points in the neighborhood, z_i is the elevation value of the i-th point, z - is the average elevation value of the neighborhood. When σ is greater than 0.15 meters, it is determined that there is significant reflection interference or sheltering vegetation influence, and the point does not participate in the digital elevation model fitting.

[0109] On the basis of the reserved point set, perform surface interpolation fitting operation, adopt the weighted moving average method based on Gaussian kernel function, and the weight function w(d) is defined as:

[0110]

[0111] Wherein, d is the projection horizontal distance of the neighborhood point to the center point, h is the kernel function bandwidth parameter, which is set to 0.6 meters in this embodiment, and is used to control the influence radius.

[0112] Unlike traditional general fitting, in order to ensure the reservation of the ground-attached structure in the power distribution scene, the structural sensitivity adjustment mechanism is introduced in the embodiment. When the grid to which a certain initial candidate point belongs is located in a high-continuity area, but the point and its adjacent points have significant spatial connectivity (for example, the point number ratio within 0.8 meters is more than 80%, and the horizontal direction density gradient is less than 10%), it is determined that the point is a ground-attached structure point, and even if the standard deviation is slightly higher than the threshold value, it can be temporarily reserved for fitting.

[0113] In addition, considering that there is a rigid area composed of concrete foundation, foundation iron and the like at the bottom of the power distribution equipment, the point cloud reflection height changes little, but has a non-natural terrain geometric form, therefore, in the above weighted fitting process, the fitting surface slope limit rule is set for the continuous flat area: if the height change of all points in a neighborhood within a horizontal distance of less than 0.8 meters is not more than 0.1 meters, the fitting slope of the area is set to not more than 10°, avoiding the distortion of the structure surface elevation as ground undulation.

[0114] Further, in step 104, the digital elevation model is compared with the first point cloud data to remove ground points, and the second point cloud data containing the power distribution equipment and the environmental target is obtained, and the specific implementation is as follows:

[0115] According to the sliding window area in the digital elevation model construction process, the first point cloud data is divided into a plurality of corresponding local subsets;

[0116] In each local subset, each point is projected to the corresponding position of the digital elevation model along the vertical direction, and the vertical height difference value of the point is calculated;

[0117] Among the points with a height difference value less than a set threshold value, the fitting residual average value between the point and its neighborhood points and the fitting surface of the digital elevation model is calculated;

[0118] For the points that meet both the height difference value threshold condition and the residual average value threshold condition, a structure template matching comparison operation is performed to identify whether they belong to the ground-attached structure point set of the power distribution equipment;

[0119] The points that do not belong to the ground-attached structure point set are deleted to form the second point cloud data containing the power distribution equipment and the environmental target.

[0120] In the embodiment, in order to remove the ground points from the first point cloud data and accurately reserve the power distribution equipment and its ground-attached structure, the following structure-based sub-division removal strategy is executed in combination with the digital elevation model construction result:

[0121] Firstly, the first point cloud data is divided into multiple local subsets corresponding to the spatial structure of the digital elevation model using a sliding window region. The points within each subset are considered to have certain spatial semantic continuity and are suitable for hierarchical structural analysis.

[0122] In each subset, all points are sorted by z-coordinate from low to high, and a profile slice segmentation is performed. Every fixed height step h s (e.g., 0.1 meters), a set of profile point bands is generated, and density analysis and directional consistency analysis are performed on the point sets in each band.

[0123] If the point density in a certain profile point band exceeds 1.5 times the average density of the adjacent two bands, and the main axis direction in the band deviates from the power line direction by less than 15°, it is determined that the band constitutes an extension segment of the device ground-continuous structure and cannot be removed as a ground point.

[0124] Further, for all points with a fitting height difference Δz less than 0.3 meters, a ground-attached trajectory extension judgment is performed: starting from the point, search for adjacent points within a 2-meter range in the main axis direction; if there are at least 3 consecutive points along the extension direction that satisfy: the included angle with the extension direction is less than 10°; the distance between points is not greater than 0.5 meters; all heights are greater than 0.2 meters of the digital elevation model; then the point is considered as a device ground-attached starting structure point and should be retained.

[0125] In addition, when judging whether the boundary area point is an isolated ground reflection point, a symmetry constraint removal mechanism is introduced. Specifically, for points near the digital elevation surface, if their projections on the x-y plane are within 1 meter of the edge of any device main axis direction; and there are no equal or approximately distributed structure points in the left and right symmetric regions; then the point is judged as an isolated ground disturbance point and should be removed.

[0126] Further, in step 105, the second point cloud data is subjected to clustering processing, and the structural feature parameters of each cluster are extracted according to the shape feature, axial length and density distribution of the point set. The specific implementation is as follows:

[0127] The second point cloud data is divided into multiple point set clusters according to the spatial distribution relationship;

[0128] For each cluster, a minimum bounding box containing the point set is constructed, and the length, width and height of the bounding box in the main axis direction are calculated;

[0129] The shape compactness parameter of the point set is calculated according to the ratio of the three-dimensional dimensions of the bounding box;

[0130] A fixed-pitch voxel grid is constructed inside the bounding box, and the number of points contained in each voxel is counted to form a voxel density distribution matrix;

[0131] An index of uniformity of distribution of the density matrix in the direction of the principal axis is taken as a density directionality parameter;

[0132] The boundary box size parameter, the shape compactness parameter and the density directionality parameter are combined to form a structural feature parameter set corresponding to the cluster group.

[0133] It should be noted that after obtaining the second point cloud data, in order to further identify the power distribution equipment target existing therein, on the basis of combining the actual law that the distribution of equipment along the power distribution line has obvious linear extension characteristics and structural space compactness difference, a direction constraint clustering method is used to extract a structural cluster group, and a multi-scale structural feature parameter is constructed for subsequent equipment category identification.

[0134] First, according to the pre-extracted main direction unit vector of the power distribution line, denoted as v l , all points P i (x i , y i , z i ) in the second point cloud data are projected to the direction of v l , a projection step Δs and a transverse envelope width w s are set, and a continuous sliding window S k is constructed to realize the partition processing of the entire power distribution area.

[0135] In each sliding window S k , a direction-aware density clustering operation is performed. The Euclidean distance between any two points is ||P i -P j ||, and the condition for clustering in the same structural cluster group is:

[0136]

[0137] Wherein, r d represents the maximum clustering distance, reflecting the physical structure connectivity constraint between points; θ max is the maximum direction angle threshold, used to avoid non-power distribution structure participating in clustering; N neigh (P i ) is the number of points in the neighborhood of point P i ; N min represents the minimum number of points required for clustering. Through the clustering process, a plurality of structural cluster groups

[0138] For each structural cluster group , a minimum rotating bounding box is further constructed to obtain its three-dimensional boundary size in the principal axis direction, the transverse direction and the vertical direction, respectively denoted as L k, W k , H k . On this basis, the compactness parameter C k for structural anisotropy analysis is constructed k , which is calculated as follows:

[0139]

[0140] The compactness parameter C k is smaller, indicating that the structure presents obvious spatial anisotropy, which is common in wire or tower structures. Conversely, when C k approaches 1, it indicates that the structure is approximately cubic or uniformly expanded, which is common in building targets.

[0141] To further describe the uniformity of the structure in the principal axis direction, a voxel unit with a length of δ l is constructed in each bounding box along the principal axis direction v l , and the voxel spacing is set to δ t in the transverse and vertical directions, respectively, where δ l < δ t to improve the expression ability of the density change in the principal axis direction. The number of points contained in each voxel in the principal axis direction is counted as ρ i , and the density directional change index D k is defined as:

[0142]

[0143] where N represents the number of voxels in the principal axis direction, and ε is a small constant introduced to prevent the denominator from being zero, which can be set to 10 -5 . The smaller the index D k , the more stable and continuous the density of the structure in the principal axis direction, reflecting the typical characteristics of wire clusters. If D k is significantly high, it may indicate that the cluster is a broken structure caused by obstructive vegetation, building boundaries, or measurement errors.

[0144] Finally, the three-dimensional size parameters L k , W k , H k , the structural compactness parameter C k and the density directional index D k corresponding to each structure cluster are combined to form a feature vector:

[0145] F k = [L k , W k , H k , C k , D k

[0146] In the existing power distribution line point cloud processing method, the clustering operation usually adopts a general spatial density clustering algorithm, such as DBSCAN, K-Means, etc. Such methods perform well in industrial general-purpose data, but in the power distribution line scene, due to the obvious structural directionality of the point cloud target, the difference in equipment size, and the characteristics of the coexistence of high-altitude suspension and ground layout, problems such as pole and tower point and conductor point mixed clustering, building boundary misclustering, and dense vegetation misidentification often occur. Especially in complex wiring areas, the conductor and background target have density close local fragments, and the general method cannot accurately distinguish the point cluster group corresponding to the power distribution physical target according to the point cloud itself geometric structure, thereby causing the subsequent classification model accuracy to be severely dependent on a large number of training samples, and the generalization performance is poor.

[0147] The present application is based on the spatial arrangement rule of power distribution line equipment, and proposes a point cloud clustering and parameter extraction method for structural orientation characteristics. By introducing a density clustering mechanism with consistent main axis direction constraint in the sliding window, the clustering formation direction is controlled to ensure the continuity and spatial coherence of the conductor and tower structure in the clustering process. By constructing the minimum bounding box for each structure cluster, the three-dimensional size and proportion relationship are calculated, and the structure compactness parameter and density directionality index are extracted by combining the density fluctuation value in the main axis direction. This series of structural features are modeled based on the spatial structure logic of actual power distribution equipment, and do not depend on complex global fitting, but describe the difference through local scale engineering structure distribution rule to improve the clustering effect.

[0148] Compared with the existing general clustering method based only on point density or distance threshold, the structural feature extraction scheme proposed in the present application not only realizes the effective separation of point sets in the spatial semantic level, but also provides the basis for the discrimination parameters required for subsequent structure classification and spatial rule verification. Its core advantage lies in the low feature dimension, clear geometric meaning, low classification model dependency, and the ability to match the actual structure of the power distribution line, realizing the whole process closed loop from point cloud data to semantic classification to structure analysis.

[0149] Further, in step 106, the extracted structural feature parameters are input into a pre-trained classification model to obtain the equipment category label corresponding to each cluster, and the classification of pole points, conductor points, building points and vegetation points is completed. The specific implementation is as follows:

[0150] Perform feature vector standardization processing on the structural feature parameters of each cluster, and scale the bounding box length, width and height to the preset interval range respectively;

[0151] Calculate the three-dimensional size ratio of the bounding box as the shape compactness parameter, and calculate the standard deviation of the point number difference of adjacent voxels in the main axis direction as the density direction change index;

[0152] The standardized bounding box parameters, shape compactness parameters and density direction change indicators are combined in a fixed order to form an input feature vector;

[0153] The input feature vector is input into a classification model trained based on typical equipment sample data of the power distribution line;

[0154] The classification model determines the device class label corresponding to the cluster according to the matching of the position of the input feature vector in the training feature space and the label space of the labeled sample;

[0155] The device class label is bound to the index of the corresponding point set to form a point cloud cluster classification result with a classification label.

[0156] Specifically, after obtaining the structural feature parameters of each cluster group in step 105, the embodiment constructs a structure perception semantic classification neural network model (hereinafter referred to as SE-Net) specially used for power distribution equipment point cloud cluster classification task, aiming at the obvious structural continuity features of towers and conductors in the three-dimensional point cloud scene of the power distribution line, and the essential differences in spatial structure and density direction distribution between conductor categories and vegetation and buildings.

[0157] In the model construction process, first, the structural feature vector F k =L k ,W k ,H k ,C k ,D k extracted in step 105 is input to the input feature embedding layer of the model.

[0158] The specific implementation method of the input feature embedding layer is to use the combination of linear mapping and nonlinear activation function to map the input 5-dimensional structural feature vector to a higher-dimensional (for example, 32-dimensional) intermediate feature space, which is expressed by the formula:

[0159] Z k =GELU(W e F k +b e )

[0160] Where Z k represents the high-dimensional embedding feature vector of the input feature (dimension 32); W e represents the weight matrix between the input layer and the embedding layer (dimension 32x5), which is initially randomized and learned during the training process; b e represents the bias vector (dimension 32), which is initially randomized and learned during the training process;

[0161] GELU: A non-linear activation function, expressed as:

[0162]

[0163] Furthermore, to significantly improve the ability to distinguish spatial continuity between conductor and tower categories, a sub-network specifically designed to enhance the spatial continuity characteristics of conductors is designed, called the Spatial Structure Continuity Enhancement Sub-network (SSCES). In its implementation, this sub-network utilizes the actual structural logic of the distribution lines to calculate the clusters currently to be classified. The two clusters that are adjacent to it in space The angle θ between the principal axis direction vectors k The formula is:

[0164]

[0165] Among them, v k-1 v k+1 Representing clusters The principal axis direction is a unit vector, with an included angle θ. k The smaller the value, the better the spatial continuity.

[0166] Subsequently, the included angle was nonlinearly normalized using the sigmoid function to obtain the structural continuity score s. k :

[0167]

[0168] Among them, s k α represents the structural continuity score, ranging from [0,1], with values ​​closer to 1 indicating higher structural continuity; α represents an empirical parameter, with a value of 10, used to control the steepness of the score curve; θ thr The threshold angle for judging structural continuity is empirically set at 30°.

[0169] After obtaining the above structural continuity score s k Then, it is combined with the embedded feature Z k Feature fusion is performed, and the fused features are fed into the next hidden layer. The fusion formula is expressed as:

[0170]

[0171] Among them, h k W represents the hidden layer output feature vector (dimension 16); h This represents the weight matrix of the fusion layer (dimension 16×33), learned during training; b h This represents the bias vector of the fusion layer (dimension 16);

[0172] LeakyReLU activation function is defined as:

[0173] LeakyReLU(x) = max(0.01x, x)

[0174] The hidden layer output features are sent to the Softmax output layer to obtain the class probability prediction distribution p of each cluster group k , and the output class is tower (0), conductor (1), building (2), and vegetation (3), and the calculation formula is:

[0175]

[0176] where w c and b c are output layer weights and bias parameters learned during the training process.

[0177] During the model training process, the training steps include: based on the power distribution line point cloud data, about 8000 groups of structure cluster sample data are labeled, and each class is not less than 2000 groups, and the conductor cluster structure continuity label is manually marked as a supervised true value.

[0178] The total training loss function is defined as the weighted sum of the classification cross-entropy loss and the structure continuity scoring loss:

[0179]

[0180] where y k,c represents the true label of the class; p k,c represents the predicted probability; s k represents the structure continuity score output by the model; represents the manually marked structure continuity label; λ represents the structure continuity loss weight, which is 0.2; M represents the number of training batch samples; M wire represents the number of samples belonging to the conductor class in the training batch.

[0181] The Adam optimizer is used for optimization, and the specific parameters are: initial learning rate 0.001, batch size 64, training rounds not more than 100, and early stopping strategy is set to stop training when the validation loss does not decrease for 5 consecutive epochs.

[0182] The model training and testing in the above manner show that the model can significantly improve the accuracy of tower and conductor class recognition, the conductor continuity misidentification rate is reduced to below 2%, the confusion between tower class and building class is obviously reduced, and compared with the general PointNet model, the overall classification accuracy is improved by about 9%.

[0183] Further, in the input feature embedding layer, the structure feature vector F kThe reason for selecting 32 dimensions from the 5-dimensional mapping to the 32-dimensional high-dimensional feature space is that, through empirical experiments in the art, a feature space lower than 20 dimensions usually cannot fully represent the structural detail differences between the tower and the conductor, resulting in limited subsequent classification accuracy; while a feature space exceeding 64 dimensions introduces too much redundant information, causing the model to be prone to overfitting. Therefore, through experimental verification and compromise between generalization and expression ability, 32 dimensions is a reasonable and stable embedding space dimension.

[0184] The parameter a controls the sensitivity of the continuity score to the change in the included angle. Through empirical testing, when the value is less than 5, the score function is not sensitive enough to the change in the direction of the conductor, making it easy to confuse the conductor with buildings or vegetation; when the value is higher than 20, the function changes too drastically, causing instability in the training of the classification model. Through repeated experiments on multiple data sets in real scenarios of power distribution equipment, it is shown that a = 10 can make the score change sensitivity moderate, suitable for most actual engineering data in the industry.

[0185] The hidden layer feature fusion layer in the model reduces the high-dimensional features to 16 dimensions, for the following reasons: a hidden layer output dimension that is too low (such as less than 8 dimensions) will significantly reduce the model's ability to distinguish between structures such as towers and conductors, resulting in a significant decrease in classification accuracy; while a dimension exceeding 32 dimensions, the feature redundancy information causes the model training speed to slow down significantly, and the generalization ability is weakened. Through repeated training and verification of real point cloud data, it is determined that a 16-dimensional hidden layer feature can balance the expression and computational efficiency.

[0186] If the weight coefficient λ is set too high (for example, more than 0.5), the structure continuity score loss term will dominate in the training, reducing the overall classification accuracy; while the weight coefficient is lower than 0.05, the continuity score has difficulty in distinguishing the conductor categories, and the classification performance improvement effect is insufficient. According to the comprehensive experimental analysis of the power distribution line data set, when the value λ = 0.2, the model reaches the optimal balance in the continuity discrimination accuracy of the conductor categories and the overall classification accuracy, which is significantly higher than other parameter combination schemes.

[0187] Further, in step 107, the classification results are verified for spatial structure relationship, and the classification results are corrected to output a power distribution equipment point cloud classification data set with a category label. The specific implementation is as follows:

[0188] Based on the direction of the power distribution line, the classified point set is divided into multiple continuously arranged spatial sub-segments;

[0189] In each sub-segment, the point set labeled as a tower is identified as a structure anchor point, and the conductor category point set on both sides of the direction is extracted;

[0190] If not, mark the classification result as abnormal.

[0191] For the conductor point set marked as abnormal, analyze the projection direction distribution and the boundary overlap of adjacent building category point set. If there is an edge nesting relationship, replace the conductor category label with the building category.

[0192] In the building and vegetation classification point set, extract the point density directionality distribution index and the ground projection area ratio of adjacent point set. If it meets the continuous structure determination rule, merge them into the same category label.

[0193] Bind all the verified point sets and corrected category labels to output the power distribution equipment point cloud classification data set with classification labels.

[0194] In step 107, after obtaining the preliminary classification results (towers, conductors, buildings, vegetation) of each cluster group in step 106, in order to solve the problems of conductor misidentification, abnormal connection between towers and conductors, and building and conductor misclassification that often occur in actual power distribution equipment point cloud classification, a classification verification and correction mechanism based on the actual spatial topology of power distribution lines is further used to refine and output the classification results.

[0195] Specifically, first, along the main axis direction of the power distribution line, the point cloud cluster sequence corresponding to the classification label obtained in step 106 is divided into multiple spatially continuous sub-sections Each spatial sub-section Contains several sequentially arranged cluster groups

[0196] Subsequently, for each sub-section Define the cluster group with a tower category (category label 0) as the structure needle point, denoted as Its spatial coordinate center in the direction of the line is O i On both sides of the tower anchor point, extract the cluster group of the conductor category (category label 1) Define the coordinate of the center of each conductor cluster group in space as

[0197] For each conductor cluster group Further perform conductor structure continuity verification. The specific determination method is to calculate the geometric projection position error and height continuity difference between the conductor cluster group center and the adjacent two tower anchor points:

[0198] Define the projection position continuity error δ between the conductor cluster group and the two side tower needle points pos As follows: ​

[0199]

[0200] where O i , O i+1 are the center position vectors of the tower anchor points ; and is the center position vector of the current conductor cluster group ; x denotes the vector cross product operation; and ‖·‖ denotes the vector Euclidean norm.

[0201] The above formula means the vertical distance of the conductor cluster group center from the line connecting the two towers. Ideally, the conductor cluster group should fall near the line connecting the two towers, so the smaller the value of δ pos , the better the position continuity. Through a large amount of actual point cloud data verification, when δ pos > D thr (where D thr is an industry experience threshold value, usually 1.5 meters), it is determined that the spatial position of the conductor cluster group is abnormal and does not meet the conductor structure continuity requirement.

[0202] Further define the continuity error δ hgt of the conductor cluster group in the height direction as:

[0203]

[0204] where h represents the height value of the conductor cluster group center position; h i , h i+1 represent the center height values of the two tower anchor points.

[0205] The height continuity error δ hgt represents the deviation degree of the conductor from the tower anchor point height. Industry practice experience shows that when δ hgt exceeds an experience threshold value (such as 3 meters), the conductor structure has a height abnormality and does not meet the actual height continuity requirement of the distribution line.

[0206] Further, define the comprehensive continuity index C wire (j) of the conductor category cluster group as a normalized weighted combination of the position error and the height error:

[0207]

[0208] where D thr , H thr are the experience threshold values of the position and the height, respectively;

[0209] γ1=0.6 and γ2=0.4 are weight factors, and the values are obtained based on actual engineering experience statistical analysis to highlight the importance of position continuity.

[0210] The continuity index C wire (j) The closer to 1, the more reasonable the classification of the wire cluster group is, and the closer to 0, the more abnormal the classification is.

[0211] For the continuity index C wire (j) The wire cluster group below 0.5 is preliminarily marked as an abnormal cluster group, and the edge overlap index E between the abnormal cluster group and the adjacent building (category label 2) cluster group wm is further calculated.

[0212]

[0213] wherein V(·) represents the horizontal projection envelope area of the cluster group; represents the horizontal projection overlap area of the wire cluster group and the building cluster group.

[0214] When the edge overlap index E wm is greater than the empirical threshold value 0.6, it indicates that the wire cluster group may be a building edge structure misidentified as a wire, and the wire category label needs to be replaced with a building category label.

[0215] In addition, in order to solve the misclassification of the vegetation (category label 3) and building (category label 2) cluster groups, the density direction index difference R bd of the two cluster groups in the adjacent area is combined to judge whether to merge: va

[0216]

[0217] wherein respectively represent the density direction index of the building and vegetation cluster groups; A (B) , A (V) respectively represent the horizontal projection area of the building and vegetation cluster groups on the ground.

[0218] When and only when R bd <0.2 and R va <0.3, the two cluster groups are merged into a building category with more consistent density structure to correct the error of the initial classification.

[0219] Finally, the category labels of all cluster groups after the above spatial continuity verification and classification correction are re-bound to the corresponding original point cloud coordinates to form a complete power distribution equipment point cloud classification data set with accurate category labels.

[0220] ​Embodiment 3 is an embodiment of the present application, which provides a power distribution equipment point cloud multi-target classification extraction system, comprising: a point cloud acquisition module, configured to acquire three-dimensional laser point cloud data covering a power distribution line area;

[0221] a preprocessing module, configured to perform noise point elimination and spatial density resampling on the acquired point cloud data to generate first point cloud data;

[0222] an elevation modeling module, configured to extract ground points based on height gradient changes of the first point cloud data, and construct a digital elevation model according to the extracted ground points;

[0223] a ground elimination module, configured to compare the digital elevation model with the first point cloud data, delete the ground points, and generate second point cloud data containing power distribution equipment and environmental targets;

[0224] a clustering extraction module, configured to perform clustering operation on the second point cloud data, and extract structural feature parameters based on shape features, axial size and spatial density distribution of each cluster point set;

[0225] a classification identification module, configured to input the structural feature parameters into a pre-trained classification model, identify a device category label corresponding to each cluster, and complete classification of towers, conductors, buildings and vegetation;

[0226] a relationship verification module, configured to perform consistency verification and label correction on the classification results based on spatial structure relationships between devices;

[0227] a result output module, configured to output a power distribution equipment point cloud classification data set with a category label.

[0228] The embodiment also provides an electronic device suitable for a power distribution equipment point cloud multi-target classification extraction method, comprising a memory and a processor; the memory is configured to store computer executable instructions, and the processor is configured to execute the computer executable instructions to implement the power distribution equipment point cloud multi-target classification extraction method proposed in the above embodiment.

[0229] The embodiment also provides a storage medium having a computer program stored thereon, and the program is executed by a processor to implement the power distribution equipment point cloud multi-target classification extraction method proposed in the above embodiment.

[0230] The storage medium proposed in the embodiment and the power distribution equipment point cloud multi-target classification extraction method proposed in the above embodiment belong to the same inventive concept, and the technical details not described in the embodiment can be referred to the above embodiment, and the embodiment has the same beneficial effects as the above embodiment.

[0231] From the above description of the embodiments, those skilled in the art can clearly understand that the present application can be implemented by means of software and necessary universal hardware, and of course can also be implemented by hardware, but in many cases the former is a better implementation. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product, which can be stored in a computer readable storage medium, such as a floppy disk, a ROM, a RAM, a FLASH, a hard disk, or an optical disc, and includes a number of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods of various embodiments of the present application.

[0232] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application but not limit the present application, and although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the present application, and all of them should be covered in the scope of the claims of the present application.

Claims

1. A power distribution equipment point cloud multi-target classification extraction method, characterized in that: The application relates to a power distribution equipment point cloud classification method and device. Three-dimensional laser point cloud data covering a power distribution line area is acquired; Noise points in the acquired point cloud data are removed and spatial density resampling is performed, so that first point cloud data is obtained; Ground points in the first point cloud data are extracted based on height gradient changes, and a digital elevation model is constructed based on the extracted ground points; The digital elevation model is compared with the first point cloud data, ground points are removed, and second point cloud data containing power distribution equipment and environmental targets is obtained; Cluster processing is performed on the second point cloud data, and structural feature parameters of each cluster are extracted according to shape features, axial length and density distribution of the point set; The extracted structural feature parameters are input into a pre-trained classification model, a device category label corresponding to each cluster is acquired, and classification of tower point, wire point, building point and vegetation point is completed; The classification result is verified in terms of spatial structure relationship, the classification result is corrected, and a power distribution equipment point cloud classification data set with a category label is output.

2. The power distribution equipment point cloud multi-target classification extraction method of claim 1, wherein: The first point cloud data is obtained by dividing the point cloud data into multiple strip segments according to the scanning track of the laser radar; In adjacent strip segments, point sets with a horizontal projection coincidence degree exceeding a predetermined ratio and a vertical height difference less than a specified distance threshold are identified and deleted; A main strike direction vector set of the power distribution line is constructed in each strip segment, and three-dimensional coordinates of each point are projected onto the corresponding direction vector; Point sets with an included angle with adjacent direction vectors exceeding a set angle threshold are deleted; In a local spatial region, normal vector included angles between each point and its neighborhood points are calculated, point sets with a standard deviation of the included angles exceeding a specified range are identified and deleted; Direction envelope intervals are constructed based on the spatial arrangement of power distribution towers and wire structures, point sets located in the envelope intervals are retained, and point sets located outside the intervals are deleted; Anisotropic voxel grid resampling operation is performed on the retained point cloud data, the sampling interval in the wire strike direction is set to be half of the transverse sampling interval, and the first point cloud data is formed.

3. The power distribution equipment point cloud multi-target classification extraction method of claim 2, wherein: The digital elevation model is constructed based on the extracted ground points, the first point cloud data is divided into multiple equidistant sliding window regions along the strike direction of the power distribution line; In each sliding window, the region is divided into sub-regions according to the grid, and the lowest height value of the points in each sub-region is calculated; According to the height gradient change trend between the lowest height values in adjacent grids, regions with continuous change characteristics are identified as an initial ground point candidate set; In the initial candidate set, abnormal point sets with a height fluctuation amplitude exceeding a predetermined range in the neighborhood are deleted; The retained point sets are subjected to surface fitting interpolation processing based on moving weighted average, and the digital elevation model is formed.

4. The power distribution equipment point cloud multi-target classification extraction method of claim 3, wherein: The second point cloud data containing power distribution equipment and environmental targets is obtained by dividing the first point cloud data into multiple corresponding local subsets according to the sliding window regions in the digital elevation model construction process; In each local subset, the points are projected onto the corresponding position of the digital elevation model along the vertical direction, and the vertical height difference value of the points is calculated; In the points with a height difference value less than a set threshold, the fitting residual average value between the points, neighborhood points and the fitting surface of the digital elevation model is calculated; For the points satisfying both the height difference threshold condition and the residual average value threshold condition, a structure template matching comparison operation is performed to identify whether they belong to the ground-attached structure point set of the power distribution equipment; For the points not belonging to the ground-attached structure point set, a deletion operation is performed to form second point cloud data containing the power distribution equipment and the environmental targets.

5. The power distribution equipment point cloud multi-target classification extraction method of claim 4, wherein: The clustering processing on the second point cloud data includes dividing the second point cloud data into multiple point set clusters according to spatial distribution relationship; For each cluster, a minimum outer bounding box containing the point set is constructed, and the length, width and height of the bounding box in the main axis direction are calculated; According to the three-dimensional size ratio of the bounding box, a shape compactness parameter of the point set is calculated; A fixed-pitch voxel grid is constructed inside the bounding box, and the number of points contained in each voxel is counted to form a voxel density distribution matrix; The distribution uniformity index of the density matrix in the main axis direction is calculated as a density directionality parameter; The bounding box size parameters, shape compactness parameters and density directionality parameters are combined to form a structure feature parameter set corresponding to the cluster.

6. The power distribution equipment point cloud multi-target classification extraction method of claim 5, wherein: The pre-trained classification model includes performing feature vector standardization processing on the structure feature parameters of each cluster, and scaling the length, width and height of the bounding box to a preset interval range respectively; The three-dimensional size ratio of the bounding box is calculated as the shape compactness parameter, and the standard deviation of the point number difference of adjacent voxels in the main axis direction is calculated as the density directionality parameter; The standardized bounding box parameters, shape compactness parameters and density directionality parameters are combined in a fixed order to form an input feature vector; The input feature vector is input into a classification model trained based on typical equipment sample data of the power distribution line; The classification model determines the equipment category label corresponding to the cluster according to the position of the input feature vector in the training feature space and the label space of the labeled samples; The equipment category label is bound with the index of the corresponding point set to form a point cloud cluster classification result with a classification label.

7. The power distribution equipment point cloud multi-target classification extraction method of claim 6, wherein: The modified classification result includes dividing the classified point set into multiple continuous spatial sub-segments based on the direction of the power distribution line; In each sub-segment, the point set marked as a tower is identified as a structure anchor point, and the conductor category point set on both sides of the anchor point in the direction is extracted; It is judged whether each conductor category point set forms a height-continuous and spatial-direction-parallel connection track between the two tower anchor points, and if not, the classification result is marked as abnormal; For the conductor point set marked as abnormal, the projection direction distribution and the boundary overlap with the adjacent building category point set are analyzed, and if there is an edge nesting relationship, the conductor category label is replaced by the building category; In the building and vegetation classification point set, the point density directionality distribution index and the ground projection area ratio of adjacent point sets are extracted, and if the continuous structure judgment rule is satisfied, they are merged into the same category label; All the verified point sets and the modified category labels are bound to output a power distribution equipment point cloud classification data set with classification labels.

8. A power distribution equipment point cloud multi-target classification extraction system, applying a power distribution equipment point cloud multi-target classification extraction method according to any one of claims 1-7, characterized in that, The system comprises: a point cloud acquisition module for acquiring three-dimensional laser point cloud data covering the power distribution line area; a preprocessing module for performing noise point elimination and spatial density resampling on the acquired point cloud data to generate first point cloud data; The elevation modeling module is configured to extract ground points based on height gradient changes of the first point cloud data, and construct a digital elevation model according to the extracted ground points; The ground removal module is configured to compare the digital elevation model with the first point cloud data, remove the ground points, and generate second point cloud data containing power distribution equipment and environmental targets; The clustering extraction module is configured to perform a clustering operation on the second point cloud data, and extract structural feature parameters based on shape features, axial size and spatial density distribution of each cluster point set; The classification identification module is configured to input the structural feature parameters into a pre-trained classification model, identify a device category label corresponding to each cluster, and complete classification of towers, conductors, buildings and vegetation; The relationship verification module is configured to perform consistency verification and label correction on the classification results based on spatial structure relationships between devices; The result output module is configured to output a power distribution equipment point cloud classification data set with a category label. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The computer program is executed by the processor to implement the steps of the power distribution equipment point cloud multi-target classification extraction method of any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the power distribution equipment point cloud multi-target classification extraction method of any one of claims 1 to 7.

Images