Three-dimensional scene point cloud data extraction and recognition system based on power distribution network

Abstract

Disclosed in the present invention is a three-dimensional scene point cloud data extraction and recognition system based on a power distribution network. The system comprises: a data collection module used for collecting data for multiple times and fusing same to acquire a complete power distribution network scene; a data processing module used for using the least square method to calibrate closest point matching and perform point cloud registration; a data extraction and segmentation module used for using principal component analysis to calculate a normal vector of the points in a neighborhood of each point cloud data point, measuring the curvature change near each point of a curved surface of a three-dimensional scene of a power distribution network by means of Gaussian curvature, and segmenting and classifying point cloud data by means of a random sampling fitting method and a clustering method; a data recognition and monitoring module used for using PointNet to perform local and global feature learning on the features of each point cloud data point, and acquiring global features by means of a maximum pooling operation for coding and recognition; and a data visual analysis module used for connecting to an open-source point cloud data processing library PCL and a hardware device for visual interaction. The efficiency and sustainability of the power supply system are improved.

Description

A 3D scene point cloud data extraction and recognition system based on distribution network Technical Field

[0001] The present invention belongs to the technical field of distribution network data processing; in particular, it relates to a three-dimensional scene point cloud data extraction and recognition system based on the distribution network. Background Art

[0002] With the increase in electricity demand and the advancement of energy transformation, distribution networks play an important role in the urbanization process. In order to achieve intelligent management and operation and maintenance of distribution networks, it is necessary to accurately extract and identify three-dimensional scene point cloud data of distribution networks.

[0003] In the real world, the three-dimensional point cloud data of distribution networks is subject to various noises and uncertainties, which affect the accuracy and reliability of the data. The three-dimensional scenes of distribution networks contain many different types of equipment and components. Traditional distribution network management relies mainly on manual inspections and record-keeping, which is inefficient and prone to errors. This seriously affects the efficiency and sustainability of the power supply system and makes it impossible to accurately locate and measure power equipment and carry out grid planning and optimization.

[0004] Through point cloud data extraction and recognition, various power equipment in the distribution network can be accurately located. Through the analysis of point cloud data, the inspection of power equipment can be automated, and an accurate three-dimensional scene model can be established to achieve precise positioning and metering of power equipment and carry out grid planning and optimization.

[0005] Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a three-dimensional scene point cloud data extraction and recognition system based on the distribution network to solve the problems of the presence of various types of equipment and components in the three-dimensional scene of the distribution network. Traditional distribution network management mainly relies on manual inspections and manual records, which are inefficient and prone to errors; seriously affects the efficiency and sustainability of the power supply system; and cannot achieve accurate positioning and metering of power equipment and carry out grid planning and optimization.

[0007] The technical solution of the present invention is:

[0008] A three-dimensional scene point cloud data extraction and recognition system based on a distribution network, comprising a data acquisition module, a data processing module, a data extraction and segmentation module, a data recognition and monitoring module, and a data visual analysis module;

[0009] Data acquisition module: Connects to the distribution network laser scanner through the system interface, installs a calibration plate to calculate the error between the actual position of the marker in the point cloud and its position on the calibration plate, adjusts the position and angle of the laser scanner device, and is used to collect and fuse data multiple times to obtain a complete distribution network scenario;

[0010] Data processing module: Statistical filtering is used to perform denoising based on the statistical information between the point in the point cloud and its area, the amount of point cloud data is reduced by voxel downsampling, and the least squares method is used to calibrate the nearest point matching and perform point cloud registration;

[0011] Data extraction and segmentation module: This module uses principal component analysis to calculate the normal vectors of points within the neighborhood of point cloud data points, measures the curvature changes near each point of the three-dimensional scene surface of the distribution network using Gaussian curvature, and segments and classifies the point cloud data using random sampling fitting and clustering methods.

[0012] Data recognition and monitoring module: calls the data extraction and segmentation module and takes the extracted features as input. It uses PointNet to learn local and global features of each point cloud data point, and obtains global features through maximum pooling operation for encoding and recognition.

[0013] Data Visual Analysis Module: Connects to the PCL open-source point cloud data processing library and hardware devices for visual interaction, allowing point cloud data 3D scenes to be rotated, translated, and scaled, highlighting or weakening different features of the point cloud data;

[0014] In a preferred embodiment, the data acquisition module is connected to the distribution network laser scanner through the system interface to provide high-precision point cloud data. The data sampling density is determined according to the complex physical structure, topological connection relationship and distribution data volume of the distribution network, wherein the complex physical structure and topological structure of the distribution network are composed of trunk lines, branch lines and transformers. The data sampling density is determined by the location of the collected data and the number of nodes. The distribution data volume is divided into peak period and trough period. The sampling density is increased by using the distribution data load change area and time period. The higher data density is used to accurately capture and analyze data details, provide more data information, including distribution network node changes, instantaneous changes in distribution parameters and fault node locations, and adjust the position and angle of the laser scanner device to collect and fuse data multiple times to obtain a complete distribution network scenario.

[0015] Furthermore, a calibration board is installed and markers are placed at different positions on the calibration board to record the point cloud position information. By calculating the error between the actual position of the marker in the point cloud and its position on the calibration board, the focal length, principal point coordinates and distortion coefficient of the device are inferred. The error calculation uses the Euclidean distance metric, and the specific formula is:

[0016] Where S represents the error, (X i ,Y i ) represents the coordinates of the marker on the calibration plate, (u i ,v i ) represents the point cloud coordinates detected in the point cloud;

[0017] In a preferred embodiment, the data processing module processes the collected point cloud data, including removing noise points in the point cloud, downsampling operations of the point cloud data, and point cloud registration;

[0018] Furthermore, statistical filtering is used to perform denoising based on the statistical information between the point in the point cloud and its neighborhood. The specific steps are: given point cloud data and defining each point cloud node P and its neighborhood N as a set of points adjacent to the point P, the mean and standard deviation of the coordinate values ​​of all points in the neighborhood are calculated. The specific calculation formula is:

[0019] where mean x ,mean y ,mean z Represents the mean of the x, y, and z coordinates of all points in the neighborhood N, std x ,std y ,std z Represents the standard deviation of the x, y, and z coordinate values ​​of all points in the neighborhood N, N xi ,N yi ,N zi Represent the x, y, and z coordinate values ​​of all points in the neighborhood N, respectively, and n represents the number of points in the neighborhood. It is then determined whether a point is a noise point. The determination process is as follows: if the point cloud node P is not in (mean x -k*std x ,mean x +k*std x ) range, it is judged as a noise point; if the point cloud node P is not within (mean y -k*std y ,mean y +k*std y ) range, it is judged as a noise point; if the point cloud node P is not within (mean z -k*std z ,mean z +k*std z ) range, it is determined to be a noise point, marked as an invalid point and deleted from the point cloud;

[0020] Furthermore, voxel downsampling is used to select the voxel size of the point cloud data as the side length of the cube and create a defined voxel grid. A point in the voxel is selected as the representative point, which includes the center point, the nearest point, and the farthest point. The points in other voxels are discarded and retained to achieve the downsampling effect. The downsampled point cloud is output to form a new point cloud with all the processed representative points.

[0021] Furthermore, the new point cloud dataset is aligned to the same coordinate system. Through iterative optimization between the two point clouds, a rigid body change relationship that minimizes the distance error between the two is found. One point cloud data is selected as the reference point cloud, and the other point cloud data is used as the point cloud to be registered for initial change alignment. The Kd tree is used to stitch the point cloud dataset into two subsets by recursively selecting the coordinate axis when building the tree. The multidimensional search tree is used to find the nearest point in the reference point cloud and establish the correspondence between the point pairs. The correspondence includes the nearest neighbor point correspondence and the point correspondence weight. The least squares method is used to calibrate the nearest point matching. The least squares method includes the translation vector and the rotation matrix. The specific formula of the translation vector is:

[0022] where Q i represents the i-th corresponding point in the reference point cloud, P i Represents the i-th corresponding point in the point cloud to be registered, N represents the number of point pairs, and the specific formula of the rotation matrix is: R = V * U T

[0023] Where U and V are orthogonal matrices that represent the characteristics of the rotation transformation, and T represents the translation vector. The translation vector T and the rotation matrix R are combined to form the rigid body transformation matrix. Repeat the iterative closest point matching and calculation of the transformation until the number of iterations is reached.

[0024] In a preferred embodiment, the data extraction and segmentation module describes the position features of each point in the three-dimensional space coordinate system through the position data of each point in the point cloud data, and uses principal component analysis to calculate the normal vector of the point in the neighborhood of the point cloud data point to describe the direction and curvature of the point cloud surface, and obtain the normal features of each point cloud. For each point cloud node q, the principal component analysis selects a certain number of point sets N in its domain, and makes each point q belong to N p , and calculate the covariance matrix of the decomposed point cloud data. The specific formula is:

[0025] Where C represents the covariance matrix, |N p | represents the number of points in the neighborhood, q i ' represents the point q in the neighborhood after centralization i The coordinate difference relative to point p, q i T Represents the transpose operation of the centered coordinate difference matrix, extracts the eigenvector corresponding to the minimum eigenvalue of C as the normal vector, and measures the curvature change near each point of the three-dimensional scene surface of the distribution network through Gaussian curvature. The specific formula is:

[0026] Where K represents Gaussian curvature, E, F, and G represent the first basic shape parameters of the surface, which are obtained by curvature changes, and L, M, and N represent the second basic shape parameters of the surface, which are obtained by the convexity of the surface. When the Gaussian curvature is positive, it is judged that any point on the three-dimensional scene surface of the distribution network is a convex surface. When the Gaussian curvature is negative, it is judged that the expansion and contraction directions of the three-dimensional scene surface area of ​​the distribution network are opposite. When the Gaussian curvature is 0, it is judged that the curvature of any point on the three-dimensional scene surface of the distribution network remains unchanged.

[0027] Furthermore, three points in the point cloud data are randomly selected as random samples, and the plane model is fitted using the selected three points, and all points whose distance to the model is less than a threshold are marked as inliers. The threshold is selected as 10 mm based on the noise level and sampling density of the point cloud data. The number of inliers is counted. When the number of inliers reaches a predetermined value and exceeds a certain proportion, the fitting result is determined to be valid, and all inliers are used to refit the plane model to obtain the final plane parameters. The steps are repeated ten times, and the set of fitting results with the largest number of inliers is selected as the final plane parameters. The point cloud data is classified using a clustering method, the number of clusters is specified, and k points are randomly selected as cluster centers. Each point in the point cloud data is associated with the nearest cluster center to form k clusters. The distance between the point and each cluster center is calculated, and the point is assigned to the cluster corresponding to the nearest cluster center, and the steps are repeated until convergence. The convergence condition is that the cluster center no longer changes and the maximum number of iterations is reached. The specific calculation formula is:

[0028] Where d(q,p) represents the distance between the point and each cluster center, p i represents the coordinates of point p in the i-th dimension, q i Represents the coordinates of point q in the i-th dimension. Points in the same cluster are considered to be point cloud areas with similar features.

[0029] In a preferred embodiment, the data recognition and monitoring module calls the data extraction and segmentation module to take the extracted features as input, uses PointNet to perform local and global feature learning on the features of each point cloud data point, and maps the point cloud data to a global descriptor of a fixed length. The mapping operation inputs unordered point cloud data, and aggregates local information using the maximum pooling operation based on the extracted feature vector, wherein the maximum pooling operation compresses the local feature vector into a single value, which is the maximum value of all feature vectors, and is used to obtain a point-level global feature descriptor, and splices it into a matrix of size M. The global feature descriptor is mapped to different classification label distributions through a fully connected layer, and a multi-level combination and hierarchical structure are introduced. The multi-level combination and hierarchical structure The structure divides the point cloud data into different hierarchical areas, and constructs a higher-level feature representation by gradually combining features of different levels for point cloud classification tasks. Through layer-by-layer sampling, aggregation and feature propagation, multi-scale local features are gradually extracted. The specific steps are: through sampling operations, the point cloud data is gradually subdivided from a coarse resolution into smaller local areas, and the local features of each point are encoded using a multi-layer perceptron network. The local features are aggregated into global features through pooling operations. The above steps are used to extract features from smaller local areas, and the local features are aggregated and propagated layer by layer to extract feature information of more scales at each layer. The features of all levels are spliced ​​and fused, and sent to the fully connected layer for the final classification task, and finally the global features are obtained.

[0030] In a preferred embodiment, the data visual analysis module is connected to the PCL open source point cloud data processing library, performs reading and writing of various point cloud data formats, loads point cloud data and displays it in a visualization window, calls the data extraction and segmentation module to obtain point cloud data features and attributes, sets different color mapping schemes, connects mouse and keyboard devices for interaction, allows point cloud data three-dimensional scenes to be rotated, translated, and scaled, and is used to view point cloud data of different angles and scales, and uses the visualization window callback function interface to implement point cloud data three-dimensional scene area selection and specific operations, including implementing interaction and event processing, using the visualization window to adjust the size of point cloud nodes, highlighting and weakening different features of point cloud data, where larger points highlight important points and areas, and smaller points reduce visual interference, mapping different data and values ​​to different colors, setting predefined color mapping schemes, including heat maps and gradient colors to understand point cloud data and recognition results, and using the visualization window to adjust the transparency of point cloud data nodes and objects to present point cloud depth information.

[0031] In a preferred embodiment, the method specifically includes the following steps:

[0032] 101. Determine the data sampling density based on the complex physical structure, topological connection relationship, and distribution data volume of the distribution network. Connect the distribution network laser scanner and calibration board to obtain high-precision point cloud data and adjust the device position and angle. Collect and fuse point cloud data multiple times to obtain a complete 3D scene of the distribution network.

[0033] 102. Statistical filtering is used to remove noise based on the statistical information between the midpoints of the point cloud and its area. All the processed representative points are combined into a new point cloud through voxel downsampling. The point cloud dataset is aligned to perform translation eigenvector and rotation matrix for point cloud registration.

[0034] 103. Use principal component analysis to calculate the normal vectors of points in the neighborhood of point cloud data points, measure the curvature changes near each point of the three-dimensional scene surface of the distribution network through Gaussian curvature, and segment and classify the point cloud data using random sampling fitting method and clustering method;

[0035] 104. Call the data extraction and segmentation module and take the extracted features as input. Use PointNet to learn local and global features of each point cloud data point, and splice and fuse features of all levels to achieve global feature recognition and monitoring.

[0036] 105. Connect to the PCL open source point cloud data processing library to read and write various point cloud data formats and load and display them in the visualization window, highlight and weaken different features of point cloud data, connect to hardware devices for interaction, and realize visual analysis operations of point cloud data.

[0037] The beneficial effects of the present invention are:

[0038] The present invention accurately locates various power equipment in the distribution network, including transformers, switchgear, and cables, through point cloud data extraction and identification, which helps operation and maintenance personnel quickly find fault points and perform equipment inspection and maintenance. The recognition technology based on point cloud data monitors the status of power equipment in real time, identifies possible fault points, and automates the inspection of power equipment through analysis of point cloud data. An accurate three-dimensional scene model is established to achieve precise positioning and metering of power equipment and carry out power grid planning and optimization, including line layout, capacity assessment, and load balancing, thereby improving the efficiency and sustainability of the power supply system.

[0039] It solves the problem that there are many different types of equipment and components in the three-dimensional scene of the distribution network. Traditional distribution network management mainly relies on manual inspections and manual records, which are inefficient and prone to errors; seriously affects the efficiency and sustainability of the power supply system; and cannot achieve accurate positioning and metering of power equipment and carry out grid planning and optimization. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] FIG1 is a flow chart of the system of the present invention;

[0041] FIG2 is a block diagram of the system structure of the present invention. DETAILED DESCRIPTION

[0042] Example 1

[0043] This embodiment provides a three-dimensional scene point cloud data extraction and recognition system based on a power distribution network as shown in FIG1 , which specifically includes the following steps:

[0044] 101. Determine the data sampling density based on the complex physical structure, topological connection relationship, and distribution data volume of the distribution network. Connect the distribution network laser scanner and calibration board to obtain high-precision point cloud data and adjust the device position and angle. Collect and fuse point cloud data multiple times to obtain a complete 3D scene of the distribution network.

[0045] 102. Statistical filtering is used to remove noise based on the statistical information between the midpoints of the point cloud and its area. All the processed representative points are combined into a new point cloud through voxel downsampling. The point cloud dataset is aligned to perform translation eigenvector and rotation matrix for point cloud registration.

[0046] 103. Use principal component analysis to calculate the normal vectors of points in the neighborhood of point cloud data points, measure the curvature changes near each point of the three-dimensional scene surface of the distribution network through Gaussian curvature, and segment and classify the point cloud data using random sampling fitting method and clustering method;

[0047] 104. Call the data extraction and segmentation module and take the extracted features as input. Use PointNet to learn local and global features of each point cloud data point, and splice and fuse features of all levels to achieve global feature recognition and monitoring.

[0048] 105. Connect to the PCL open source point cloud data processing library to read and write various point cloud data formats and load and display them in the visualization window, highlight and weaken different features of point cloud data, connect to hardware devices for interaction, and realize visual analysis operations of point cloud data.

[0049] Example 2

[0050] This embodiment provides a three-dimensional scene point cloud data extraction and recognition system based on the power distribution network as shown in FIG2 , which specifically includes: a data acquisition module, a data processing module, a data extraction and segmentation module, a data recognition and monitoring module, and a data visual analysis module;

[0051] Data acquisition module: Connects to the distribution network laser scanner through the system interface, installs a calibration plate to calculate the error between the actual position of the marker in the point cloud and its position on the calibration plate, adjusts the position and angle of the laser scanner device, and is used to collect and fuse data multiple times to obtain a complete distribution network scenario;

[0052] Data processing module: Statistical filtering is used to perform denoising based on the statistical information between the point in the point cloud and its area, the amount of point cloud data is reduced by voxel downsampling, and the least squares method is used to calibrate the nearest point matching and perform point cloud registration;

[0053] Data extraction and segmentation module: This module uses principal component analysis to calculate the normal vectors of points within the neighborhood of point cloud data points, measures the curvature changes near each point of the three-dimensional scene surface of the distribution network using Gaussian curvature, and segments and classifies the point cloud data using random sampling fitting and clustering methods.

[0054] Data recognition and monitoring module: calls the data extraction and segmentation module and takes the extracted features as input. It uses PointNet to learn local and global features of each point cloud data point, and obtains global features through maximum pooling operation for encoding and recognition.

[0055] Data Visual Analysis Module: Connects to the PCL open-source point cloud data processing library and hardware devices for visual interaction, allowing point cloud data 3D scenes to be rotated, translated, and scaled, highlighting or weakening different features of the point cloud data;

[0056] 101. Determine the data sampling density based on the complex physical structure, topological connection relationship, and distribution data volume of the distribution network. Connect the distribution network laser scanner and calibration board to obtain high-precision point cloud data and adjust the device position and angle. Collect and fuse point cloud data multiple times to obtain a complete 3D scene of the distribution network.

[0057] In this embodiment, what needs to be explained specifically is the data acquisition module. The data acquisition module is connected to the distribution network laser scanner through the system interface to provide high-precision point cloud data. The data sampling density is determined according to the complex physical structure, topological connection relationship and distribution data volume of the distribution network. The complex physical structure and topological structure of the distribution network are composed of trunk lines, branch lines and transformers. The data sampling density is determined by the location of the collected data and the number of nodes. The distribution data volume is divided into peak period and trough period. The sampling density is increased by using the area and time period of the load change of the distribution data. The higher data density is used to accurately capture and analyze data details and provide more data information, including distribution network node changes, instantaneous changes in distribution parameters and fault node locations. The position and angle of the laser scanner device are adjusted to collect and fuse data multiple times to obtain a complete distribution network scenario.

[0058] Furthermore, a calibration board is installed and markers are placed at different positions on the calibration board to record the point cloud position information. By calculating the error between the actual position of the marker in the point cloud and its position on the calibration board, the focal length, principal point coordinates and distortion coefficient of the device are inferred. The error calculation uses the Euclidean distance metric, and the specific formula is:

[0059] Where S represents the error, (X i ,Y i ) represents the coordinates of the marker on the calibration plate, (u i ,v i ) represents the point cloud coordinates detected in the point cloud.

[0060] 102. Statistical filtering is used to remove noise based on the statistical information between the midpoints of the point cloud and its area. All the processed representative points are combined into a new point cloud through voxel downsampling. The point cloud dataset is aligned to perform translation eigenvector and rotation matrix for point cloud registration.

[0061] In this embodiment, the data processing module specifically needs to be explained. The data processing module processes the collected point cloud data, including removing noise points in the point cloud, downsampling operations of the point cloud data, and point cloud registration;

[0062] Furthermore, statistical filtering is used to perform denoising based on the statistical information between the point in the point cloud and its neighborhood. The specific steps are: given point cloud data and defining each point cloud node P and its neighborhood N as a set of points adjacent to the point P, the mean and standard deviation of the coordinate values ​​of all points in the neighborhood are calculated. The specific calculation formula is:

[0063] where mean x ,mean y ,mean z Represents the mean of the x, y, and z coordinates of all points in the neighborhood N, std x ,std y ,std z Represents the standard deviation of the x, y, and z coordinate values ​​of all points in the neighborhood N, N xi ,N yi ,N zi Represent the x, y, and z coordinate values ​​of all points in the neighborhood N, respectively, and n represents the number of points in the neighborhood. It is then determined whether a point is a noise point. The determination process is as follows: if the point cloud node P is not in (mean x -k*std x ,mean x +k*std x ) range, it is judged as a noise point; if the point cloud node P is not within (mean y -k*std y ,mean y +k*std y ) range, it is judged as a noise point; if the point cloud node P is not within (mean z -k*std z,mean z +k*std z ) range, it is determined to be a noise point, marked as an invalid point and deleted from the point cloud;

[0064] Furthermore, voxel downsampling is used to select the voxel size of the point cloud data as the side length of the cube and create a defined voxel grid. A point in the voxel is selected as the representative point, which includes the center point, the nearest point, and the farthest point. The points in other voxels are discarded and retained to achieve the downsampling effect. The downsampled point cloud is output to form a new point cloud with all the processed representative points.

[0065] Furthermore, the new point cloud dataset is aligned to the same coordinate system. Through iterative optimization between the two point clouds, a rigid body change relationship that minimizes the distance error between the two is found. One point cloud data is selected as the reference point cloud, and the other point cloud data is used as the point cloud to be registered for initial change alignment. The Kd tree is used to stitch the point cloud dataset into two subsets by recursively selecting the coordinate axis when building the tree. The multidimensional search tree is used to find the nearest point in the reference point cloud and establish the correspondence between the point pairs. The correspondence includes the nearest neighbor point correspondence and the point correspondence weight. The least squares method is used to calibrate the nearest point matching. The least squares method includes the translation vector and the rotation matrix. The specific formula of the translation vector is:

[0066] where Q i represents the i-th corresponding point in the reference point cloud, P i Represents the i-th corresponding point in the point cloud to be registered, N represents the number of point pairs, and the specific formula of the rotation matrix is: R = V * U T

[0067] Among them, U and V are orthogonal matrices, representing the characteristics of the rotation transformation, and T represents the translation vector. The translation vector T and the rotation matrix R are combined to form the rigid body transformation matrix. The closest point matching and transformation calculation are repeated until the number of iterations is reached.

[0068] 103. Use principal component analysis to calculate the normal vectors of points in the neighborhood of point cloud data points, measure the curvature changes near each point of the three-dimensional scene surface of the distribution network through Gaussian curvature, and segment and classify the point cloud data using random sampling fitting method and clustering method;

[0069] In this embodiment, the data extraction and segmentation module specifically needs to be explained. The data extraction and segmentation module describes the position characteristics of each point in the three-dimensional space coordinate system through the position data of each point in the point cloud data, and uses principal component analysis to calculate the normal vector of the point in the neighborhood of the point cloud data point to describe the direction and curvature of the point cloud surface and obtain the normal characteristics of each point cloud. For each point cloud node q, the principal component analysis selects a certain number of point sets N in its domain, and makes each point q belong to N p , and calculate the covariance matrix of the decomposed point cloud data. The specific formula is:

[0070] Where C represents the covariance matrix, |N p | represents the number of points in the neighborhood, q i ' represents the point q in the neighborhood after centralization i The coordinate difference relative to point p, q i T Represents the transpose operation of the centered coordinate difference matrix, extracts the eigenvector corresponding to the minimum eigenvalue of C as the normal vector, and measures the curvature change near each point of the three-dimensional scene surface of the distribution network through Gaussian curvature. The specific formula is:

[0071] Where K represents Gaussian curvature, E, F, and G represent the first basic shape parameters of the surface, which are obtained by curvature changes, and L, M, and N represent the second basic shape parameters of the surface, which are obtained by the convexity of the surface. When the Gaussian curvature is positive, it is judged that any point on the three-dimensional scene surface of the distribution network is a convex surface. When the Gaussian curvature is negative, it is judged that the expansion and contraction directions of the three-dimensional scene surface area of ​​the distribution network are opposite. When the Gaussian curvature is 0, it is judged that the curvature of any point on the three-dimensional scene surface of the distribution network remains unchanged.

[0072] Furthermore, three points in the point cloud data are randomly selected as random samples, and the plane model is fitted using the selected three points, and all points whose distance to the model is less than a threshold are marked as inliers. The threshold is selected as 10 mm based on the noise level and sampling density of the point cloud data. The number of inliers is counted. When the number of inliers reaches a predetermined value and exceeds a certain proportion, the fitting result is determined to be valid, and all inliers are used to refit the plane model to obtain the final plane parameters. The steps are repeated ten times, and the set of fitting results with the largest number of inliers is selected as the final plane parameters. The point cloud data is classified using a clustering method, the number of clusters is specified, and k points are randomly selected as cluster centers. Each point in the point cloud data is associated with the nearest cluster center to form k clusters. The distance between the point and each cluster center is calculated, and the point is assigned to the cluster corresponding to the nearest cluster center, and the steps are repeated until convergence. The convergence condition is that the cluster center no longer changes and the maximum number of iterations is reached. The specific calculation formula is:

[0073] Where d(q,p) represents the distance between the point and each cluster center, p i represents the coordinates of point p in the i-th dimension, q i Represents the coordinates of point q in the i-th dimension, and the points in the same cluster are regarded as point cloud areas with similar features.

[0074] 104. Call the data extraction and segmentation module and take the extracted features as input. Use PointNet to learn local and global features of each point cloud data point, and splice and fuse features of all levels to achieve global feature recognition and monitoring.

[0075] In this embodiment, what needs to be explained specifically is the data recognition and monitoring module. The data recognition and monitoring module calls the data extraction and segmentation module to take the extracted features as input, uses PointNet to perform local and global feature learning on the features of each point cloud data point, and maps the point cloud data to a global descriptor of a fixed length. The mapping operation inputs unordered point cloud data, and aggregates local information using the maximum pooling operation based on the extracted feature vector, wherein the maximum pooling operation compresses the local feature vector into a single value, which is the maximum value of all feature vectors, and is used to obtain a point-level global feature descriptor, and splices it into a matrix of size M. The global feature descriptor is mapped to different classification label distributions through the fully connected layer, and a multi-level combination and hierarchical structure are introduced. The multi-level group The combined and hierarchical structure divides the point cloud data into different hierarchical areas, and constructs a higher-level feature representation by gradually combining features of different levels for point cloud classification tasks. Through layer-by-layer sampling, aggregation and feature propagation, multi-scale local features are gradually extracted. The specific steps are: through sampling operations, the point cloud data is gradually subdivided from a coarse resolution into smaller local areas, and the local features of each point are encoded using a multi-layer perceptron network. The local features are aggregated into global features through pooling operations. The above steps are used to extract features from smaller local areas, and the local features are aggregated and propagated layer by layer to extract feature information of more scales at each layer. The features of all levels are spliced ​​and fused, and sent to the fully connected layer for the final classification task, and finally the global features are obtained.

[0076] 105. Connect to the PCL open source point cloud data processing library to read and write various point cloud data formats and load and display them in the visualization window, highlight and weaken different features of point cloud data, connect to hardware devices for interaction, and realize visual analysis operations of point cloud data;

[0077] In this embodiment, what needs to be specifically explained is the data visual analysis module. The data visual analysis module is connected to the PCL open source point cloud data processing library, reads and writes multiple point cloud data formats, loads point cloud data and displays it in the visualization window, calls the data extraction and segmentation module to obtain point cloud data features and attributes, sets different color mapping schemes, connects the mouse and keyboard devices for interaction, allows the point cloud data three-dimensional scene to be rotated, translated, and zoomed, and is used to view point cloud data of different angles and scales. The visualization window callback function interface is used to realize point cloud data three-dimensional scene area selection and specific operations, including realizing interaction and event processing, using the visualization window to adjust the size of point cloud nodes, highlighting and weakening different features of point cloud data, where larger points highlight important points and areas, and smaller points reduce visual interference, mapping different data and values ​​to different colors, setting predefined color mapping schemes, including heat maps and gradient colors to understand point cloud data and recognition results, and using the visualization window to adjust the transparency of point cloud data nodes and objects to present point cloud depth information.

Claims

1. A 3D scene point cloud data extraction and recognition system based on distribution network, Features: The system comprises: Data acquisition module: connects to the distribution network laser scanner through the system interface, installs the calibration plate to calculate the error between the actual position of the marker in the point cloud and its position on the calibration plate, adjusts the position and angle of the laser scanner device, and is used to collect and fuse data multiple times to obtain a complete distribution network scenario; Data processing module: Statistical filtering is used to perform denoising based on the statistical information between the midpoints of the point cloud and its area, the amount of point cloud data is reduced by downsampling the voxel grid, and the least squares method is used to calibrate the nearest point matching and perform point cloud registration; Data extraction and segmentation module: Use principal component analysis to calculate the normal vector of points in the neighborhood of point cloud data points, measure the curvature changes near each point of the three-dimensional scene surface of the distribution network through Gaussian curvature, and segment and classify the point cloud data through random sampling fitting method and clustering method; Data recognition and monitoring module: Call the data extraction and segmentation module to take the extracted features as input, use PointNet to learn local and global features of each point cloud data point, and obtain global features through maximum pooling operation for encoding and recognition; Data visualization analysis module: connects to the PCL open source point cloud data processing library and hardware devices for visual interaction, allowing point cloud data three-dimensional scenes to be rotated, translated, and scaled to highlight and weaken different features of point cloud data.

2. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 1, Features: The data acquisition module is connected to the distribution network laser scanner through the system interface to provide high-precision point cloud data. The data sampling density is determined according to the physical structure of the distribution network, the topological connection relationship and the amount of distribution data. The data sampling density is determined by the location of the collected data and the number of nodes. The sampling density is increased by using the distribution data load changes in the region and time period. The calibration board is installed in the data acquisition module. Markers at different positions are placed on the calibration board to record the point cloud position information. The error between the actual position of the marker in the point cloud and its position on the calibration board is calculated to reverse the device focal length, principal point coordinates and distortion coefficient. The error calculation uses the Euclidean distance metric. The specific formula is: Where S represents the error, (X i ,Y i ) represents the coordinates of the marker on the calibration plate, (u i ,v i ) represents the point cloud coordinates detected in the point cloud.

3. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 1, Features: The data processing module uses statistical filtering to perform denoising based on the statistical information between the point in the point cloud and its domain, including: given the point cloud data and defining each point cloud node P and its domain N as a set of points adjacent to the point P, calculating the mean and standard deviation of the coordinate values ​​of all points in the domain, the calculation formula is: where mean x ,mean y ,mean z Respectively represent the mean of the x, y, and z coordinates of all points in the neighborhood N, std x ,std y ,std z Represents the standard deviation of the x, y, and z coordinates of all points in the neighborhood N, respectively. xi ,N yi ,N zi Respectively represent the x, y, and z coordinate values ​​of all points in the neighborhood N, and n represents the number of points in the neighborhood.

4. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 1, Features: The least squares method includes the translation vector and the rotation matrix, where the translation vector formula is: Where Q i represents the i-th corresponding point in the reference point cloud, P i represents the i-th corresponding point in the point cloud to be registered, N represents the number of point pairs, and the specific formula of the rotation matrix is: R=V*U T Among them, U and V are orthogonal matrices, representing the characteristics of the rotation transformation, T represents the translation vector, and the translation vector T and the rotation matrix R are combined to form a rigid body transformation matrix. The nearest point matching and calculation transformation are repeated until the number of iterations is reached.

5. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 1, Features: The data extraction and segmentation module uses principal component analysis to calculate the normal vectors of points in the neighborhood of point cloud data points, which is used to describe the direction and curvature of the point cloud surface and obtain the normal features of each point cloud. For each point cloud node q, the principal component analysis selects a certain number of point sets N in its domain, and makes each point q belong to N p , and calculate the covariance matrix of the decomposed point cloud data, the formula is: Where C represents the covariance matrix, |N p | represents the number of points in the neighborhood, q i ' represents the point q in the neighborhood after centralization i The coordinate difference relative to point p, q i T represents the transpose operation of the centered coordinate difference matrix, extracts the eigenvector corresponding to the minimum eigenvalue of C as the normal vector, and measures the curvature change near each point of the three-dimensional scene surface of the distribution network through Gaussian curvature. The specific formula is: Among them, K represents Gaussian curvature, E, F, and G represent the first basic shape parameters of the surface, which are obtained through the change of curvature, and L, M, and N represent the second basic shape parameters of the surface, which are obtained through the convexity of the surface. When the Gaussian curvature is positive, it is judged that any point on the three-dimensional scene surface of the distribution network is a convex surface. When the Gaussian curvature is negative, it is judged that the expansion and contraction directions of the three-dimensional scene surface area of ​​the distribution network are opposite. When the Gaussian curvature is 0, it is judged that the curvature of any point on the three-dimensional scene surface of the distribution network remains unchanged.

6. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 1, Features: The data extraction and segmentation module uses a clustering method to classify point cloud data, specifies the number of clusters and randomly selects k points as cluster centers, associates each point in the point cloud data with the nearest cluster center to form k clusters, calculates the distance between the point and each cluster center, and assigns the point to the cluster corresponding to the nearest cluster center, and repeats the steps until convergence. The convergence condition is until the cluster center no longer changes and the maximum number of iterations is reached. The formula is: Where d(q,p) represents the distance between the point and each cluster center, p i represents the coordinates of point p in the i-th dimension, q i Represents the coordinates of point q in the i-th dimension. Points in the same cluster are regarded as point cloud regions with similar features.

7. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 1, Features: The data recognition and monitoring module uses PointNet to perform local and global feature learning on the features of each point cloud data point, and maps the point cloud data to a global descriptor of a fixed length. The mapping operation inputs unordered point cloud data, and aggregates local information based on the extracted feature vector using the maximum pooling operation, wherein the maximum pooling operation compresses the local feature vector into a single value, which is the maximum value of all feature vectors, and is used to obtain a point-level global feature descriptor and splice it into a matrix of size M; the global feature descriptor is mapped to different classification label distributions through a fully connected layer, and a multi-level combination and hierarchical structure are introduced. The multi-level combination and hierarchical structure divide the point cloud data into different hierarchical areas, and construct a higher-level feature representation by combining features of different levels step by step, which is used for point cloud classification tasks. Multi-scale local features are gradually extracted through layer-by-layer sampling, aggregation and feature propagation.

8. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 7, Features: Through sampling operations, the point cloud data is gradually subdivided from a coarse resolution into smaller local areas. The local features of each point are encoded using a multi-layer perceptron network. The local features are aggregated into global features through pooling operations. The above steps are used to extract features from smaller local areas, and the local features are aggregated and propagated layer by layer to extract feature information of more scales at each layer. The features of all levels are spliced ​​and fused and sent to the fully connected layer for the final classification task, and finally the global features are obtained.

9. The three-dimensional scene point cloud data extraction and recognition system based on the distribution network according to claim 1, Features: The data visualization analysis module connects to the PCL open source point cloud data processing library and hardware devices for visual interaction, loads point cloud data and displays it in the visualization window, calls the data extraction and segmentation module to obtain point cloud data features and attributes, allows point cloud data three-dimensional scenes to be rotated, translated, and scaled, and uses the visualization window callback function interface to implement point cloud data three-dimensional scene area selection and specific operations, including interaction and event processing.

10. A three-dimensional scene point cloud data extraction and recognition system based on a distribution network according to claim 1, Features: Visual interaction implementation methods include:

101. Determine the data sampling density according to the complex physical structure, topological connection relationship and distribution data volume of the distribution network, connect the distribution network laser scanner and calibration board to obtain high-precision point cloud data and adjust the device position and angle, collect and fuse point cloud data multiple times to obtain a complete distribution network three-dimensional scene; 102. Statistical filtering is used to remove noise based on the statistical information between the points in the point cloud and their areas. All processed representative points are combined into a new point cloud through voxel downsampling, and the point cloud dataset is aligned to perform translation feature vector and rotation matrix for point cloud registration; 103. Use principal component analysis to calculate the normal vector of the points in the neighborhood of the point cloud data point, measure the curvature change of the three-dimensional scene surface of the distribution network near each point through Gaussian curvature, and segment and classify the point cloud data by random sampling fitting method and clustering method; 104. Call the data extraction and segmentation module to take the extracted features as input, use PointNet to learn local and global features of each point cloud data point, splice and fuse features of all levels, and realize recognition and monitoring of global features; 105. Connect to PCL open source point cloud data processing library, read and write various point cloud data formats and load and display them in the visualization window, highlight and weaken different features of point cloud data, connect hardware devices for interaction, and realize visual analysis operations of point cloud data.

Images