Electric power laser point cloud semantic segmentation method based on improved RandLA-Net

Abstract

The invention relates to a power laser point cloud semantic segmentation method based on improved RandLA-Net, and the method comprises the steps: obtaining the point cloud data of a power transmission line as a sample library, and dividing the data of the sample library into a training set, a verification set and a test set according to a certain proportion; eliminating outliers and reducing noise for the point cloud data in the sample library by adopting a point cloud preprocessing algorithm of a straight-through filter; performing point cloud voxelization operation on the preprocessed point cloud data, and mapping the point cloud data to discrete voxel grids in the voxelization process; constructing a point cloud semantic segmentation network of the improved RandLA-Net, encoding voxel positions and intensity features on voxel grids through the point cloud semantic segmentation network, and calculating voxel categories; and assigning voxel categories to the points in the voxels according to the affiliation relationship between the point cloud and the voxels so as to obtain the category of each point. The point cloud voxelization operation is directly carried out on the point cloud data, and the preprocessing time is shortened.

Description

Technical Field

[0001] The invention relates to the field of three-dimensional laser point cloud data processing, and in particular to an electric laser point cloud semantic segmentation method. Background Art

[0002] In recent years, 3D laser scanning technology has developed rapidly in the digital modeling of power transmission and transformation equipment, highlighting a very broad application prospect. In terms of line hidden danger inspection, it provides an efficient inspection method for the identification and analysis of dangerous points such as tree barriers, crossing distances, and air gap clearance distances. In terms of autonomous drone inspections, it provides a high-precision modeling tool for drone track planning. At present, laser point cloud data classification relies heavily on manual data processing, which has problems such as low efficiency and accuracy.

[0003] Due to the problems of complex scenes, rich targets, uneven point density and noise points in airborne laser point cloud data, it is always a challenge to extract effective information from a large amount of laser radar point cloud data for scene classification. Traditional point cloud classification methods need to use various machine learning classifiers to classify the feature descriptions extracted from manually labeled point clouds. Usually, the correlation between point cloud data is ignored due to the inability to fully connect the context information. Noise is easily generated during the classification process, and the model generalization ability is poor. The application in real scenes is limited, and the efficiency and accuracy of point cloud classification are not high.

[0004] In recent years, more and more researchers have applied deep learning to point cloud classification, and the classification efficiency and accuracy have been greatly improved compared with traditional point cloud classification algorithms. Methods for implementing point cloud classification based on deep learning can be divided into four categories, namely, classification methods based on original point clouds, voxel-based point cloud classification methods, range image-based point cloud classification methods, and fusion-based point cloud classification methods. Among them, the classification method based on original point clouds, RandLA-Net, is an efficient and lightweight network structure that can directly process large-scale point clouds and obtain the segmentation results of each point. RandLA-Net uses random sampling technology to save a lot of computing resources and memory resources. The local feature aggregation module can effectively retain the geometric detail features of the point cloud and improve the accuracy of recognition by gradually increasing the receptive field of each 3D point.

[0005] Among the classification methods that directly process point clouds, the classification method based on RandLA-Net has relatively high classification accuracy and low computing and memory resource usage. However, the raster sampling and KD-Tree query in the data preprocessing stage are time-consuming. In addition, the minimum likelihood sampling method used by RandLA-Net in the data loading stage does not support data parallel sampling, which further affects the efficiency of point cloud semantic segmentation calculations. Summary of the invention

[0006] The purpose of the present invention is to solve the deficiencies of the above-mentioned prior art, thereby providing an electric laser point cloud semantic segmentation method based on improved RandLA-Net to solve the problem of low computational efficiency of the prior art.

[0007] A method for semantic segmentation of electric laser point cloud based on improved RandLA-Net includes the following steps:

[0008] Obtain the transmission line point cloud data as a sample library, and divide the sample library data into a training set, a validation set, and a test set according to a certain ratio;

[0009] A point cloud preprocessing algorithm using a straight-through filter is used on the point cloud data in the sample library to remove outliers and reduce noise;

[0010] Performing a point cloud voxelization operation on the preprocessed point cloud data. During the voxelization process, the point cloud data is mapped onto a discrete voxel grid.

[0011] The point cloud semantic segmentation network model is trained based on the training set, the validation set and the test set to obtain the point cloud semantic segmentation network of the improved RandLA-Net. After encoding the voxel position and intensity features on the voxel grid through the improved RandLA-Net point cloud semantic segmentation network, the voxel category is calculated.

[0012] According to the relationship between the point cloud and the voxel, the voxel category is assigned to the point within the voxel, thereby obtaining the category of each point.

[0013] The transmission line includes 1000kV, ±800kV, ±500kV, 500kV, and 220kV voltage levels, with each voltage level accounting for 20%; of which double-circuit and single-circuit voltage levels account for about 50%; the line in mountainous areas and plains accounts for about 50% of each voltage level, and the total length of the transmission line is 1000km;

[0014] The labeling categories of the sample library include 7 categories: poles, ground wires, conductors, insulators, ground, vegetation and buildings.

[0015] The ratios of the training set, validation set and test set are 60%, 20% and 20% respectively.

[0016] The point cloud data in the sample library is processed using a point cloud preprocessing algorithm with a straight-through filter to remove outliers and reduce noise. Specifically:

[0017] A straight-through filter is used to filter out points that are not within the given value range in the three dimensions of x, y, and z; outliers are removed and noise is reduced to ensure the purity of the input sample.

[0018] A straight-through filter is used to filter out points that are not within the given value range in the three dimensions of x, y, and z; outliers are removed and noise is reduced. The specific steps are as follows:

[0019] A straight-through filter is used to remove points that are not within the given value range in the three dimensions of x, y, and z to obtain a processed point cloud dataset R = {r i ,i=1,2,3,...,n}, where r i represents the i-th point cloud data, and n is the total number of point clouds in the sample;

[0020] For any point r i , calculate the average distance between the point and the k points closest to it, denoted as A i :

[0021]

[0022] Where j = 1, 2, 3, ..., k, represents the point r i The closest point to j, d ij Represents point r i The distance from point j;

[0023] Calculate A i The mean μ and standard deviation σ are:

[0024]

[0025]

[0026] A KD tree is constructed for the point cloud data, and then each point in the point cloud is traversed. The KD tree is used to query its k nearest neighbor points, and then the Euclidean distance is used to calculate the distance to its k neighboring points. Points whose distances are outside the range of (μ-f·σ,μ+f·σ) are defined as discrete points and eliminated. f is the proportional coefficient used to adjust the noise reduction effect. The smaller f is, the more points are eliminated, and the larger f is, the fewer points are eliminated.

[0027] The preprocessed point cloud data is subjected to point cloud voxelization. During the voxelization process, the point cloud data is mapped to a discrete voxel grid, specifically:

[0028] Based on the voxel size of 0.1 meters, the voxel size is recorded as v, and a 3D grid covering the entire point cloud data range is created. The 3D grid is composed of voxels of uniform size, and each voxel can be regarded as a small cube in the 3D space;

[0029] Traverse each point in the point cloud, determine the voxel to which it belongs according to the coordinates of the point, and finally obtain a voxel grid with depth, height and width respectively d, h, w;

[0030] For each non-empty voxel, its position and intensity features are used to represent the voxel features. The average position of all points in the voxel is recorded as the position feature of the voxel, and the average intensities of all points in the voxel is recorded as the intensity feature of the voxel. If there is no point in the voxel, the feature values ​​of the voxel are set to 0. Finally, the voxel features are represented as a feature map of d*h*w*4, where 4 represents the number of feature channels for each voxel.

[0031] The point cloud semantic segmentation network of the improved RandLA-Net consists of an encoder and a decoder. The encoder is mainly composed of three-dimensional convolution and maximum pooling, and the decoder is mainly composed of deconvolution and activation function.

[0032] After encoding the voxel position and intensity features on the voxel grid through the point cloud semantic segmentation network, the voxel category is calculated as follows:

[0033] The position and intensity features of the voxel are encoded by the encoder of the point cloud semantic segmentation network, and then the voxel category is calculated by the decoder of the point cloud semantic segmentation network.

[0034] The encoder encodes the position and intensity features of the voxel, specifically:

[0035] The encoder has five layers. The first layer uses two 3D convolutions to calculate the input voxel features to obtain a feature map of d*h*w*32. Then, the feature map is downsampled using maximum pooling to obtain a feature map of (d / 2)*(h / 2)*(w / 2)*32.

[0036] The second, third and fourth layers have the same structure as the first layer. The feature map of (d / 4)*(h / 4)*(w / 4)*64 is obtained through the second layer, the feature map of (d / 8)*(h / 8)*(w / 8)*128 is obtained through the third layer, and the feature map of (d / 16)*(h / 16)*(w / 16)*256 is obtained through the fourth layer.

[0037] The fifth layer performs feature extraction on the output of the fourth layer by two three-dimensional convolutions to obtain a feature map of (d / 16)*(h / 16)*(w / 16)*512.

[0038] The decoder calculates the voxel class as:

[0039] The decoder has five layers. The first layer is a deconvolution that upsamples the output of the fifth layer of the encoder to obtain a feature map of (d / 8)*(h / 8)*(w / 8)*256;

[0040] The second layer concatenates the upsampled features with the fourth layer encoded features, extracts features through two 3D convolutions, and then obtains a feature map of (d / 4)*(h / 4)*(w / 4)*128 through upsampling.

[0041] The third, fourth and fifth layers have the same structure as the second layer. The feature map of (d / 2)*(h / 2)*(w / 2)*64 is obtained through the third layer, and the feature map of d*h*w*32 is obtained through the fourth layer.

[0042] Then, in the fifth layer, the output features of the fourth layer are concatenated with the encoding features of the first layer, and then a three-dimensional convolution layer and a softmax activation function are passed to obtain the voxel semantic segmentation result.

[0043] The point cloud data is voxelized to obtain point cloud voxelized data, and the voxels are segmented through a pre-built point cloud semantic segmentation network to obtain the segmentation result, thereby improving the efficiency of point cloud classification.

[0044] RandLA-Net needs to perform grid sampling and KD-Tree query on point clouds in the data preprocessing stage, which is time-consuming. However, the point cloud semantic segmentation algorithm based on the improved RandLA-Net directly performs point cloud voxelization operations on point cloud data, shortening the preprocessing time.

[0045] The electric power point cloud semantic segmentation network uses three-dimensional convolution to extract point cloud voxel features. This feature extraction method can extract features from point cloud voxel data of different sizes. Therefore, multiple copies of point cloud data can be loaded in parallel during the data loading phase, and there is no fixed requirement for the point cloud size. Data parallel loading can accelerate the model training and reasoning process. However, the minimum likelihood sampling method used by RandLA-Net during the data loading phase does not support data parallel sampling, which reduces the training and reasoning speed. In addition, RandLA-Net performs minimum likelihood sampling through nearest neighbor queries. This sampling method takes a long time, further increasing the training and reasoning time.

[0046] The network in this paper adopts a unique encoder-decoder structure, which helps the network retain spatial information in the point cloud segmentation task. Through skip connections, the network is able to fuse the features in the encoder with the features in the decoder at each step of the decoder. This fusion helps to retain more contextual information during the point cloud segmentation process, thereby improving the accuracy of segmentation.

[0047] Point cloud voxelization combined with 3D convolution and encoding-decoding structure provides an effective method for processing 3D point cloud data. Through voxelization, the original irregular point cloud data can be converted into regular grid data, and 3D convolution further enhances the network's processing ability for 3D spatial data. The application of encoding-decoding structure enables this voxelized 3D convolutional neural network to maintain high accuracy while also having good operating efficiency and generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A flowchart of the semantic segmentation method for electric laser point cloud;

[0049] Figure 2 This is a schematic diagram of the network structure for semantic segmentation of power point clouds;

[0050] Figure 3 The semantic segmentation effect of towers and insulators on 500kV power lines;

[0051] Figure 4 The semantic segmentation effect of conductors, ground wires, ground, and vegetation on 500kV power lines;

[0052] Figure 5 This is the semantic segmentation effect of towers and insulators on 1000kV power lines;

[0053] Figure 6 The semantic segmentation effect of conductors, ground wires, ground, and vegetation on 1000kV power lines;

[0054] Figure 7 The semantic segmentation effect of poles and insulators on power lines in mountainous terrain;

[0055] Figure 8 The semantic segmentation effect of conductors, ground wires, vegetation, and ground on power lines in mountainous terrain;

[0056] Fig. 9 This is the semantic segmentation effect of point cloud on power lines in plain terrain. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0058] like Figure 1 ,An electric laser point cloud semantic segmentation method based on ,improved RandLA-Net consists of three parts, including: sample library establishment, point cloud data ,preprocessing and point cloud semantic segmentation algorithm.

[0059] 1. Sample library establishment

[0060] The point cloud data of transmission lines are obtained as a sample library. The constructed sample library contains the point cloud data of transmission lines of 1000kV, ±800kV, ±500kV, 500kV, and 220kV voltage levels, with each voltage level accounting for 20%; double-circuit and single-circuit voltage levels account for about 50%; the voltage levels of lines in mountainous areas and plains account for about 50%, with a total of 1000km of transmission line point cloud sample library. There are 7 sample annotation categories, including towers, ground wires, conductors, insulators, ground, vegetation, and buildings.

[0061] The sample library data of the transmission line point cloud is divided into training set, validation set and test set according to a certain ratio, with the ratios of 60%, 20% and 20% respectively.

[0062] 2. Point cloud data preprocessing

[0063] A point cloud preprocessing algorithm with a straight-through filter is used for the point cloud data in the sample library to remove outliers (external interference points that are too far away from the center point) and reduce noise to ensure the purity of the input samples;

[0064] A straight-through filter is used to filter out points that are not within the given value range in the three dimensions of x, y, and z, thereby eliminating outliers and reducing noise. The specific method and steps are described as follows:

[0065] Step 2.1: Use a straight-through filter to remove points that are not within the given value range in the three dimensions of x, y, and z to obtain the processed point cloud dataset R = {r i ,i=1,2,3,...,n}, where r i represents the i-th point cloud data, and n is the total number of point clouds in the sample;

[0066] Step 2.2: For any point r i , calculate the average distance between the point and the k points closest to it, denoted as A i :

[0067]

[0068] Where j = 1, 2, 3, ..., k, represents the point r i The closest point to j, d ij Represents point r i The distance from point j;

[0069] Step 2.3: Calculate A i The mean μ and standard deviation σ are:

[0070]

[0071] Step 2.4: The points in the target point cloud file whose distance from the point to its neighborhood is outside the range of (μ-f·σ,μ+f·σ) are considered discrete points and are removed. Where f is the proportional coefficient, which is used to adjust the noise reduction effect and is determined by the density of the noise points.

[0072] 3. Point cloud semantic segmentation algorithm based on improved RandLA-Net

[0073] The point cloud data is subjected to a point cloud voxelization operation. During the voxelization process, the point cloud data is mapped onto a discrete voxel grid.

[0074] The basic operation process of point cloud voxelization:

[0075] (1) Create a voxel grid: Create a 3D grid covering the entire point cloud data range based on a voxel size of 0.1 meters, denoted by voxel size v. This grid consists of voxels of uniform size, and each voxel can be regarded as a small cube in 3D space.

[0076] (2) Assigning point cloud to voxels: By calculating the distance between the point cloud point and the center of each voxel in the voxel grid, each point cloud point is assigned to the corresponding voxel whose distance on the X-axis, Y-axis, and Z-axis is less than v / 2. If there is no point in the voxel, the eigenvalue of the voxel is set to 0; otherwise, the average value of the features of the points in the voxel is recorded as the eigenvalue of the voxel. The coordinate range of the point cloud on the X-axis, Y-axis, and Z-axis is recorded as x, y, and z, respectively, and finally a voxel grid with a depth, height, and width of d, h, and w is obtained. The calculation formula is as follows:

[0077]

[0078] (3) Voxel feature representation: For each non-empty voxel, its position and intensity features are used to represent the voxel feature. The average value of the positions of all points in the voxel is recorded as the position feature of the voxel, and the average value of the intensities of all points in the voxel is recorded as the intensity feature of the voxel. Finally, the voxel feature is represented as a matrix with depth, height, width, and number of channels d, h, w, and 4 respectively.

[0079] like Figure 2 , build an improved RandLA-Net point cloud semantic segmentation network, and train the point cloud semantic segmentation network model based on the training set, validation set and test set to obtain the improved RandLA-Net point cloud semantic segmentation network.

[0080] The improved RandLA-Net point cloud semantic segmentation network consists of an encoder and a decoder. The encoder is mainly composed of three-dimensional convolution and maximum pooling, and the decoder is mainly composed of deconvolution and activation function. The encoder encodes the voxel position and intensity features, and then the decoder calculates the voxel category.

[0081] The encoder has five layers. The first layer uses two 3D convolutions to extract features from the input voxels with depth, height, width, and number of channels d, h, w, 4, respectively, to obtain a feature matrix of d*h*w*32. Then, the feature matrix is ​​downsampled using maximum pooling to obtain a feature matrix of (d / 2)*(h / 2)*(w / 2)*32. The second, third, and fourth layers have the same structure as the first layer. The second layer obtains a feature matrix of (d / 4)*(h / 4)*(w / 4)*64, the third layer obtains a feature matrix of (d / 8)*(h / 8)*(w / 8)*128, and the fourth layer obtains a feature matrix of (d / 16)*(h / 16)*(w / 16)*256. The fifth layer uses two 3D convolutions to extract features from the output of the fourth layer to obtain a feature matrix of (d / 16)*(h / 16)*(w / 16)*512.

[0082] The decoder has five layers. The first layer uses a deconvolution to upsample the output of the fifth layer of the encoder to obtain a feature matrix of (d / 8)*(h / 8)*(w / 8)*256. The second layer concatenates the upsampled features with the encoded features of the fourth layer, and then extracts features by two three-dimensional convolutions. Then, a feature matrix of (d / 4)*(h / 4)*(w / 4)*128 is obtained by upsampling. The third, fourth, and fifth layers have the same structure as the second layer. The feature matrix of (d / 2)*(h / 2)*(w / 2)*64 is obtained through the third layer, and the feature matrix of d*h*w*32 is obtained through the fourth layer. Finally, in the fifth layer, the output features of the fourth layer are concatenated with the encoded features of the first layer, and then a three-dimensional convolution layer and a softmax activation function are used to obtain the voxel semantic segmentation result.

[0083] Finally, according to the relationship between the point cloud and the voxel, the voxel category is assigned to the point in the voxel to obtain the category of each point.

[0084] The experiment was conducted in Windows 10 environment, based on Pytorch programming, using Intel 11th Gen Intel (R) Core (TM) i9-11900K 8-core processor, 128GB memory, and NVIDIA 48 GB memory graphics card Quadro RTX8000 to train the power point cloud semantic segmentation model.

[0085] The loss function used in training is cross entropy loss, and the weight of the model is updated through the Adam optimizer, and the segmentation categories classes = 7. The power point cloud semantic segmentation network model is trained based on the training set, and the learned power point cloud semantic segmentation network model is evaluated based on the validation set and the test set. By continuously adjusting and optimizing the relevant parameters, the optimal hyperparameter settings are as follows: the initial learning rate is 0.001, the momentum is 0.95, and the epoch is 100.

[0086] In order to verify the generalization and accuracy of the point cloud semantic segmentation model, the point cloud semantic segmentation model effect was tested on power point cloud data of different voltage levels and different terrains. The test was conducted using an Intel 11th Gen Intel(R) Core(TM) i9-11900K 8-core processor, 128GB memory, and NVIDIA 48 GB memory graphics card Quadro RTX8000.

[0087] 1. 500kV power line

[0088] The semantic segmentation effect of point cloud on 500kV power line is as follows Figure 3 , 4 As shown, Figure 3 (a) is the semantic segmentation effect diagram of the tachymeter tower and umbrella-shaped insulator on the 500kV power line. Figure 3 (b) is the semantic segmentation effect diagram of the wine glass-shaped tower and type I and type V hybrid insulators on the 500kV power line. The segmentation results show that the recognition accuracy of the point cloud semantic segmentation model on the 500kV power line is above 80%, and it has a high recognition accuracy for the thousand-shaped tower, wine glass-shaped tower, umbrella-shaped insulator, type I and type V hybrid insulator, conductor, ground wire, vegetation, ground, and buildings.

[0089] On this power line, the prediction time of the point cloud semantic segmentation algorithm based on the improved RandLA-Net was 245s, and the prediction time of RandLA-Net was 972s. The prediction time of RandLA-Net was much higher than that of the point cloud semantic segmentation algorithm based on the improved RandLA-Net.

[0090] 2. 1000kV power lines

[0091] The semantic segmentation effect of point cloud on 1000kV power line is as follows Figure 5 , 6 As shown, Figure 5 (a) is the semantic segmentation effect diagram of drum-shaped tower and V-type insulator on 1000kV power line. Figure 5(b) is the semantic segmentation effect diagram of drum-shaped towers and double umbrella-shaped insulators on 1000kV power lines. The segmentation results show that the recognition accuracy of the point cloud semantic segmentation model on 1000kV power lines is above 80%, and it has a high recognition accuracy for drum-shaped towers, V-type insulators, double umbrella-shaped insulators, conductors, ground wires, vegetation, and the ground.

[0092] On this power line, the prediction time of the point cloud semantic segmentation algorithm based on the improved RandLA-Net was 1027s, and the prediction time of RandLA-Net was 6249s. The prediction time of RandLA-Net was much higher than that of the point cloud semantic segmentation algorithm based on the improved RandLA-Net.

[0093] 3. Power lines in mountainous terrain

[0094] The semantic segmentation effect of point cloud on power lines in mountainous terrain is as follows Figure 7 , 8 As shown, Figure 7 (a) is the semantic segmentation effect of drum-shaped towers and umbrella-shaped insulators on power lines in mountainous terrain. Figure 8 (b) is the semantic segmentation effect diagram of the ram-horn tower and type I insulator on the power line in mountainous terrain. The segmentation results show that the recognition accuracy of the point cloud semantic segmentation model on the power line in mountainous terrain is above 80%, and it has a high recognition accuracy for drum-shaped towers, ram-horn towers, type I insulators, umbrella-shaped insulators, conductors, ground wires, vegetation, and the ground.

[0095] On this power line, the prediction time of the point cloud semantic segmentation algorithm based on the improved RandLA-Net was 549s, and the prediction time of RandLA-Net was 3146s. The prediction time of RandLA-Net was much higher than that of the point cloud semantic segmentation algorithm based on the improved RandLA-Net.

[0096] 4. The segmentation effect of the electric power point cloud in plain terrain is as follows:

[0097] The semantic segmentation effect of point cloud on power lines on plain terrain is as follows Fig. 9 The segmentation results show that the point cloud semantic segmentation model has an accuracy rate of over 80% in identifying power lines on plain terrain, and has a high recognition accuracy for towers, insulators, conductors, ground wires, vegetation, ground, and buildings.

[0098] On this power line, the prediction time of the point cloud semantic segmentation algorithm based on the improved RandLA-Net was 158s, and the prediction time of RandLA-Net was 743s. The prediction time of RandLA-Net was much higher than that of the point cloud semantic segmentation algorithm based on the improved RandLA-Net.

[0099] The test results show that the point cloud semantic segmentation algorithm based on the improved RandLA-Net has achieved high accuracy on power point cloud data of different voltage levels and different terrains. The prediction time of the improved algorithm is much lower than that of RandLA-Net, which shows that the point cloud semantic segmentation algorithm based on the improved RandLA-Net has significantly improved in computational efficiency.

[0100] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents. Any modification or equivalent replacement that does not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A semantic segmentation method for electric laser point cloud based on improved RandLA-Net, characterized by: The following steps are involved: Obtain the transmission line point cloud data as a sample library, and divide the sample library data into a training set, a validation set, and a test set according to a certain ratio; A point cloud preprocessing algorithm using a straight-through filter is used on the point cloud data in the sample library to remove outliers and reduce noise; Performing a point cloud voxelization operation on the preprocessed point cloud data. During the voxelization process, the point cloud data is mapped onto a discrete voxel grid. The point cloud semantic segmentation network model is trained based on the training set, the validation set and the test set to obtain the point cloud semantic segmentation network of the improved RandLA-Net. After encoding the voxel position and intensity features on the voxel grid through the improved RandLA-Net point cloud semantic segmentation network, the voxel category is calculated. According to the relationship between the point cloud and the voxel, the voxel category is assigned to the point within the voxel, thereby obtaining the category of each point.

2. According to claim 1, a method for semantic segmentation of electric laser point cloud based on improved RandLA-Net, characterized in that: The transmission line includes 1000kV, ±800kV, ±500kV, 500kV, and 220kV voltage levels, with each voltage level accounting for 20%; of which double-circuit and single-circuit voltage levels account for about 50%; the line in mountainous areas and plains accounts for about 50% of each voltage level, and the total length of the transmission line is 1000km; The labeling categories of the sample library include 7 categories: poles, ground wires, conductors, insulators, ground, vegetation and buildings.

3. The electric laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 1, characterized in that: The ratios of the training set, validation set and test set are 60%, 20% and 20% respectively.

4. The electric laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 1, characterized in that: The point cloud data in the sample library is processed using a point cloud preprocessing algorithm with a straight-through filter to remove outliers and reduce noise. Specifically: A straight-through filter is used to filter out points that are not within the given value range in the three dimensions of x, y, and z; outliers are removed and noise is reduced to ensure the purity of the input sample.

5. The electric laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 5, characterized in that: A straight-through filter is used to filter out points that are not within the given value range in the three dimensions of x, y, and z; outliers are removed and noise is reduced. The specific steps are as follows: A straight-through filter is used to remove points that are not within the given value range in the three dimensions of x, y, and z to obtain a processed point cloud dataset R = {r i ,i=1,2,3,...,n}, where r i represents the i-th point cloud data, and n is the total number of point clouds in the sample; For any point r i , calculate the average distance between the point and the k points closest to it, denoted as A i : Where j = 1, 2, 3, ..., k, represents the point r i The jth point closest to the ij Represents point r i The distance from point j; Calculate A i The mean μ and standard deviation σ are: A KD tree is constructed for the point cloud data, and then each point in the point cloud is traversed. The KD tree is used to query its k nearest neighbor points, and then the Euclidean distance is used to calculate the distance to its k neighboring points. Points whose distances are outside the range of (μ-f·σ,μ+f·σ) are defined as discrete points and eliminated. f is the proportional coefficient used to adjust the noise reduction effect. The smaller f is, the more points are eliminated, and the larger f is, the fewer points are eliminated.

6. The electric laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 1, characterized in that: The preprocessed point cloud data is subjected to point cloud voxelization. During the voxelization process, the point cloud data is mapped to a discrete voxel grid, specifically: Based on the voxel size of 0.1 meters, the voxel size is recorded as v, and a 3D grid covering the entire point cloud data range is created. The 3D grid is composed of voxels of uniform size, and each voxel can be regarded as a small cube in the 3D space; Traverse each point in the point cloud, determine the voxel to which it belongs according to the coordinates of the point, and finally obtain a voxel grid with depth, height and width respectively d, h, w; For each non-empty voxel, its position and intensity features are used to represent the voxel features. The average position of all points in the voxel is recorded as the position feature of the voxel, and the average intensities of all points in the voxel is recorded as the intensity feature of the voxel. If there is no point in the voxel, the feature values ​​of the voxel are set to 0. Finally, the voxel features are represented as a feature map of d*h*w*4, where 4 represents the number of feature channels for each voxel.

7. The electric laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 1, characterized in that: The point cloud semantic segmentation network of the improved RandLA-Net consists of an encoder and a decoder. The encoder is mainly composed of three-dimensional convolution and maximum pooling, and the decoder is mainly composed of deconvolution and activation function.

8. The electric laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 1 or 7, characterized in that: After encoding the voxel position and intensity features on the voxel grid through the point cloud semantic segmentation network, the voxel category is calculated as follows: The position and intensity features of the voxel are encoded by the encoder of the point cloud semantic segmentation network, and then the voxel category is calculated by the decoder of the point cloud semantic segmentation network.

9. The electric power laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 8, characterized in that: The encoder encodes the position and intensity features of the voxel, specifically: The encoder has five layers. The first layer uses two 3D convolutions to calculate the input voxel features to obtain a feature map of d*h*w*32. Then, the feature map is downsampled using maximum pooling to obtain a feature map of (d / 2)*(h / 2)*(w / 2)*32. The second, third and fourth layers have the same structure as the first layer. The feature map of (d / 4)*(h / 4)*(w / 4)*64 is obtained through the second layer, the feature map of (d / 8)*(h / 8)*(w / 8)*128 is obtained through the third layer, and the feature map of (d / 16)*(h / 16)*(w / 16)*256 is obtained through the fourth layer. The fifth layer performs feature extraction on the output of the fourth layer by two three-dimensional convolutions to obtain a feature map of (d / 16)*(h / 16)*(w / 16)*512.

10. The electric power laser point cloud semantic segmentation method based on improved RandLA-Net according to claim 8, characterized in that: The decoder calculates the voxel class as: The decoder has five layers. The first layer is a deconvolution that upsamples the output of the fifth layer of the encoder to obtain a feature map of (d / 8)*(h / 8)*(w / 8)*256; The second layer concatenates the upsampled features with the fourth layer encoded features, extracts features through two 3D convolutions, and then obtains a feature map of (d / 4)*(h / 4)*(w / 4)*128 through upsampling. The third, fourth and fifth layers have the same structure as the second layer. The feature map of (d / 2)*(h / 2)*(w / 2)*64 is obtained through the third layer, and the feature map of d*h*w*32 is obtained through the fourth layer. Then, in the fifth layer, the output features of the fourth layer are concatenated with the encoding features of the first layer, and then a three-dimensional convolution layer and a softmax activation function are passed to obtain the voxel semantic segmentation result.

Images