A Substation Equipment Point Cloud Segmentation Method Based on Improved RandLA-Net

Abstract

The present invention discloses a substation equipment point cloud segmentation method based on an improved RandLA-Net, including: acquiring point cloud data and image data; making a training data set using the point cloud data and the image data, performing MIX3D data augmentation on the point cloud data, importing it into a point cloud segmentation dual-stream network to obtain point cloud features and image color texture features; according to the mapping relationship between the image data and the point cloud data, fusing and splicing the point cloud features and the image color texture features to obtain a feature vector, importing it into a segmentation head network to output a test value of the point cloud segmentation result; substituting it into a loss function to calculate the cross-entropy loss, and performing backpropagation derivation; optimizing the point cloud segmentation dual-stream network according to the gradient of the backpropagation derivation, and outputting the segmentation result of the substation equipment point cloud based on the optimized point cloud segmentation dual-stream network. The present invention effectively balances the global semantic information and local structure information of the substation, introduces an image branch network to supplement the color texture features of the point cloud, and improves the accuracy of substation equipment point cloud segmentation.

Description

Technical Field

[0001] The present invention belongs to the fields of holographic panoramic digital substation and point cloud semantic segmentation, and particularly relates to a substation equipment point cloud segmentation method based on an improved RandLA-Net. Background Art

[0002] Semantic segmentation, as one of the basic tasks of computer vision, has rich downstream applications. With the development of deep learning, the semantic segmentation task of two-dimensional images has achieved very excellent results. The ultimate manifestation of computer vision is three-dimensional vision. In recent years, research on autonomous driving, augmented reality, and three-dimensional scene reconstruction has been booming, and depth sensor technology has also developed vigorously, enabling people to easily collect a large amount of point cloud data. The industrial and academic communities have gradually shifted their attention from two-dimensional images to the processing of three-dimensional point clouds.

[0003] For traditional substation point cloud segmentation, it is necessary to manually design feature descriptors, and send the extracted geometric features into classifiers such as SVM, random forest, and naive Bayes for recognition and segmentation, or perform segmentation through clustering methods such as region growing. The design of feature descriptors is cumbersome and complex, with poor generalization and low recognition accuracy. Inspired by the deep learning of 2D images, multi-view based methods have been proposed by researchers. By projecting the point cloud onto images from multiple perspectives, compared with the traditional method based on manually designed features, good results have been achieved. However, the projection method causes a large amount of key geometric information to be lost, thus affecting the segmentation effect of the point cloud. The proposal of PointNet in 2017 enables the direct input of point clouds into the network, and the 3D point cloud field has officially entered the era of deep learning. So far, the relatively mainstream algorithms are divided into the following three categories. One category is voxel-based methods such as Sparse-ConvNet and MinkowskiNet, which convert point clouds into voxels and can directly use the mature 2D pipeline. The disadvantage is that when the point cloud scene is too large, the computational cost of 3D convolution is very high. The second category is graph-based methods represented by KCNet and DGCNN. Point clouds, which are non-Euclidean data, are naturally suitable for graph structures, and feature aggregation is performed through the nearest neighbor graph. This method can learn better local information, but how to handle complex large-scale point cloud data still needs to be solved. The third category is point-based methods represented by KPConv, PointTransformer, and RandLANet, which directly process the original point cloud data with high computational efficiency. Currently, 3D modeling mainly relies on manual work, with a long modeling cycle and a large amount of human resources consumed. Summary of the Invention

[0004] Objective of the Invention: To overcome the deficiencies in the prior art, the present invention provides a method for segmenting point clouds of substation equipment based on an improved RandLA-Net. By constructing a basic model library, the present invention performs semantic segmentation on the point clouds obtained by scanning a substation, and then matches the segmentation results with the models constructed in the model library, thereby automatically constructing a complete 3D model of the substation. Since the data collected in the substation scenario is large-scale dense point clouds, the computationally efficient RandLA-Net is selected as the benchmark model and corresponding improvements are made according to the characteristics of the substation dataset and task requirements to achieve 3D visualization of the data, presenting the overall structure and equipment distribution of the substation in three dimensions, making the operation and maintenance of substation equipment more efficient and fast, and effectively reducing the negative safety impacts and economic losses of the power system.

[0005] Technical Solution: In a first aspect, the present invention provides a method for segmenting point clouds of substation equipment based on an improved RandLA-Net, which is characterized by comprising:

[0006] Receiving the scanned parts and images of the substation respectively, and obtaining point cloud data and image data from the scanned parts and images respectively;

[0007] Making a training dataset according to the obtained point cloud data and image data, and performing MIX3D data augmentation on the point cloud data to obtain the augmented point cloud data; wherein the training dataset is divided into a training set, a validation set and a test set;

[0008] Importing the augmented point cloud data and image data into a point cloud segmentation dual-stream network to obtain point cloud features and image color texture features, wherein the point cloud segmentation dual-stream network comprises: a point cloud branch network and an image branch network;

[0009] Substituting the point cloud data and image data in the training dataset into the camera imaging model for calculation to obtain the mapping relationship between the image data and the point cloud data;

[0010] Fusing and splicing the point cloud features and the image color texture features according to the mapping relationship between the image data and the point cloud data to obtain a feature vector;

[0011] Importing the feature vector into a segmentation head network to output the point cloud segmentation result;

[0012] Substituting the point cloud segmentation result into a loss function to calculate the cross-entropy loss, and performing backpropagation derivation on the cross-entropy loss to obtain training parameters;

[0013] Correcting the error weights in the segmentation network according to the training parameters;

[0014] Based on the corrected segmentation network, inputting the point cloud data to obtain the segmentation result of the point clouds of substation equipment.

[0015] In a further embodiment, the point cloud data in the training dataset is labeled, and the training dataset is divided into a training set, a validation set, and a test set according to 7:1.5:1.5;

[0016] Among them, the training set is the data sample for training the network;

[0017] The validation set is used to provide validation information for hyperparameter tuning;

[0018] The test set is the data sample for testing the network error.

[0019] In a further embodiment, the method for performing MIX3D data augmentation on the point cloud data to obtain the augmented point cloud data includes:

[0020] For the point cloud data in each sample data in the training set, data augmentation is sequentially performed using algorithms of random subsampling, elastic distortion, random brightness and contrast enhancement, and color jitter, so as to obtain the augmented point cloud data for each sample;

[0021] All the sample data with augmented point cloud data in the training set are mixed pairwise, and the annotations of the point cloud data are spliced to obtain the augmented point cloud data.

[0022] In a further embodiment, the point cloud branch network is the RandLA-Net network that improves the downsampling method of the encoder;

[0023] The image branch network consists of a downsampled backbone network and a transposed convolution for upsampling.

[0024] In a further embodiment, the improved downsampling method of the RandLA-Net network is the farthest point sampling method, including:

[0025] Randomly select a single point cloud A from the initial point cloud dataset i into the sampling dataset, and calculate the distances from the remaining point clouds to the point cloud A in the sampling dataset with the single point cloud A i as the query point; i The distance;

[0026] According to the distances from the remaining point cloud data to the point cloud data A in the sampling dataset i , select another point cloud Aj with a distance greater than others and put it into the sampling dataset B. At this time, B = {Ai, Aj};

[0027] Determine the query point according to the remaining point clouds in the point cloud dataset A, continue to calculate the distances from each remaining point cloud in the point cloud dataset A to the query point, and select the point cloud with a distance greater than others and input it into the sampling dataset B,

[0028] Determine the query point in a loop, and select the sampling data according to the determined query point until the number of data in the sampling data set B is equal to the target data M;

[0029] Among them, determining the query point includes: calculating the distance from any remaining point cloud in the point cloud data set A to all point clouds in the sampling data set B, and selecting the point cloud with a distance less than others in the sampling data set B as the query point;

[0030] A is the point cloud data set, and in the initial state, A = {A1, A2,..., AN}, where N is the total amount of point cloud data in A in the initial state;

[0031] B is the sampling data set, and it is empty in the initial state;

[0032] M is the set target data volume.

[0033] In a further embodiment, the image branch network uses an encoder-decoder architecture;

[0034] The encoder is designed as follows: change the stride of the 3×3 convolutional layer and the skip connection in the first residual structure of the residual network block 4 from 2 to 1, and replace the 3×3 ordinary convolutional layer with a dilated convolutional layer. Blocks 5 / 6 and block 4 adopt the same design with different dilation rates to obtain a dilated convolutional layer residual network;

[0035] Calculate and design the dilation rate of all convolutional kernels in the dilated convolutional layer residual network according to the calculation formula of the upper limit distance of non-zero elements to obtain the optimized value of the dilation rate of each convolutional kernel in the dilated convolutional layer residual network;

[0036] Among them, the decoder uses 3 transposed convolutional layers for upsampling decoding;

[0037] The calculation formula of the upper limit distance of non-zero elements is:

[0038] M i = max[M o+1 - 2r i , M i+1 - 2(M i+1 - r i ), r i , M 2 ≤ K

[0039] In the formula, M i represents the maximum distance of non-zero elements in the i-th layer, and M n = r n, , r iLet \(r_i\) denote the \(i\)-th dilation rate, and \(K\) denote the size unit of the convolutional kernel. Among them, to avoid the problem of dilated convolution, for \(N\) dilation rates \([r_1,...r_i,...,r_n]\) with \(N\) convolutional kernels of size \(K\times K\), the constraint condition \(M2\leq K\) is set.

[0040] The conversion formula for the shape of the transposed convolution is:

[0041] \(n' = sn + k - 2p - s\)

[0042] In the formula, \(n\) is the height or width of the input, \(n'\) is the height or width of the output, \(k\) is the size of the convolutional kernel, \(p\) is the padding, and \(s\) is the stride.

[0043] In a further embodiment, the expression of the camera imaging model is:

[0044] \(z*I\) i \(= K*[R|t]*P\) i

[0045] In the formula, \(K\) is the internal parameter of the camera, \(R\) is the rotation matrix, \(t\) is the translation variable, and \(P\) i is the coordinate of the point cloud data, and \(I\) i is the image pixel coordinate.

[0046] In a further embodiment, the segmentation head network is composed of three fully connected layers and a dropout function layer.

[0047] In a further embodiment, the expression of the loss function is:

[0048]

[0049] In the formula, \(N\) is the number of samples, \(C\) is the number of categories, \(\hat{y}\) ij is the predicted value, and \(y\)

[0050] Beneficial effects: Compared with the prior art, the present invention has the following advantages:

[0051] (1) Collect data in multiple formats to make a dataset, and introduce MIX3D for data augmentation, effectively balancing the global semantic information and local structure information of the substation, preventing overfitting, and thus improving the model training effect; finally, use the trained network model to perform semantic segmentation on the point cloud data scanned from the substation, and then match the segmentation result with the model constructed in the model library, so as to automatically construct the complete three-dimensional model data of the substation.

[0052] (2) Adopt farthest point sampling to cover all points in space as much as possible and ensure uniform sampling of the samples.

[0053] (3) Introduce an image branch network to perform feature-level fusion of images and point clouds through coordinate mapping relationships, enhancing the segmentation accuracy of the model. Description of the Drawings

[0054] Figure 1 is the training flowchart of the point cloud segmentation network described in the present invention;

[0055] Figure 2 is the structure diagram of the point cloud segmentation network described in the present invention;

[0056] Figure 3 is the backbone network structure diagram of the image branch network of the present invention. Detailed Embodiments

[0057] To more fully understand the technical content of the present invention, the technical solutions of the present invention will be further introduced and described below in conjunction with specific embodiments, but not limited thereto.

[0058] The point cloud segmentation method for substation equipment provided by the present invention is improved on the basis of the RandLA-Net network, replacing the downsampling method in the encoder, adding an image branch network to the original feature extraction module. And data augmentation is performed on the dataset to improve the segmentation performance of the model.

[0059] Step 1: Scan and photograph the substation, receive the scanned copy and the photographed substation image, and obtain point cloud data and image data;

[0060] Step 2: Make a training dataset according to the obtained point cloud data and image data; the point cloud data in the training dataset is labeled, and the image data does not need to be labeled. In this embodiment, the image size is 2048*2048; perform MIX3D data augmentation on the dataset;

[0061] The specific method of the above data augmentation is as follows:

[0062] Perform data augmentation on the point cloud data in each sample data in the training set in turn by using algorithms of random sub-sampling, elastic distortion, enhancement of random brightness and contrast, and color jitter, so as to obtain enhanced point cloud data for each sample;

[0063] Mix all the sample data with enhanced point cloud data in the training set pairwise, and splice the labels of the point cloud data to obtain enhanced point cloud data.

[0064] The method of pairwise mixing includes: First, subtract the centroid from two scenes so that the center of each scene is located at the origin of the coordinates; to ensure that the two scenes can overlap with each other in the next stage; randomly flip the point cloud along the horizontal direction, and randomly rotate the scene along the vertical axis and another axis by uniform sampling in the range [-π / 64, π / 64]; since the order of the point cloud data in the scene is not changed during the mixing process, directly take the union of the annotation values and save them for calculating the parameters of the loss function.

[0065] Step 3: Import the enhanced point cloud data and image data into the point cloud segmentation two-stream network to obtain point cloud features and image color texture features, where the point cloud segmentation two-stream network includes: a point cloud branch network and an image branch network; as Figure 2 shown, the point cloud branch is the RandLA-Net network that replaces the encoder downsampling method, and the downsampling method used is FPS. The specific steps are as follows:

[0066] The original point cloud of the substation scene A = {A1, A2,..., AN}, there are N points in A, the sampled point set is B, and B is empty in the initial state. Select M points from the point set A and add them to the point set B.

[0067] S301: Randomly select a point Ai from A and add it to B, B = {Ai}

[0068] S302: Calculate the distances from the remaining N - 1 points in A to the point Ai in the point set B, select the point Aj with the maximum distance, and add it to the point set B, B = {Ai, Aj}.

[0069] The distance from a point in the point set A to the set B is the distance from this point to the point with the minimum distance among all points in the set B. Calculate the distances from the remaining points in the point set A to the set B in turn, and select the point with the maximum distance and add it to the set B.

[0070] S304: Loop the third step until the number of points in the set B is equal to M

[0071] The design of the image branch network is as follows:

[0072] The image branch network uses an encoder-decoder architecture;

[0073] The encoder is designed as follows: Change the stride of the 3×3 convolutional layer and the skip connection in the first residual structure of the residual network block 4 from 2 to 1, and replace the 3×3 ordinary convolutional layer with a dilated convolutional layer. Blocks 5 / 6 and block 4 adopt the same design with different dilation rates to obtain a dilated convolutional layer residual network;

[0074] Calculate and design the dilation rate of all convolutional kernels in the dilated convolutional layer residual network according to the calculation formula of the upper limit distance of non-zero elements, and obtain the optimized value of the dilation rate of each convolutional kernel in the dilated convolutional layer residual network;

[0075] Among them, the decoder uses 3 transposed convolutional layers for upsampling decoding;

[0076] The calculation formula of the upper limit distance of non-zero elements is:

[0077] M i = max[M i+1 - 2r i , M i+1 - 2(M i+1 - r i ), r i , M 2 ≤K

[0078] In the formula, M i represents the maximum distance of non-zero elements in the i-th layer, and M n = r n , r i is the i-th dilation rate, K is the size element of the convolutional kernel, and N convolutional kernels are of size K*K; among them, in order to avoid the gridding effect problem of dilated convolution, for N dilation rates [r1,...ri,...,rn] of N convolutional kernels of size K*K, the constraint condition M2≤K is set.

[0079] The conversion formula for the shape of the transposed convolution is:

[0080] n′ = sn + k - 2p - s

[0081] In the formula, n is the height or width of the input, n’ is the height or width of the output, k is the size of the convolutional kernel, p is the padding, and s is the stride.

[0082] In this embodiment, the image branch network draws on the design of the Deeplabv3 network. For example, Figure 3 Blocks 1 / 2 / 3 in it are the layer structures in the original ResNet network. Change the stride of the 3×3 convolutional layer and the skip connection in the first residual structure in the residual network block 4 to 1, and change the 3×3 ordinary convolutional layer to a dilated convolutional layer. Blocks 5 / 6 and block 4 adopt the same design, and the dilation rates are 2, 4, and 8. Obtain the dilated convolutional layer residual network. Step 4: Substitute the point cloud data and image data in the training dataset into the camera imaging model for calculation to obtain the mapping relationship between the image data and the point cloud data;

[0083] According to the principle of the camera imaging model, the above mapping relationship can be known, and the specific representation formula is:

[0084] z*I i= K * [R|t] * P i

[0085] Among them, K is the internal parameter of the camera, R is the rotation matrix, t is the translation variable, Pi is the coordinate of the point cloud data, and Ii is the coordinate of the image pixel.

[0086] Step 5: According to the mapping relationship between the image data and the point cloud data, fuse and splice the point cloud features and the image color texture features to obtain a feature vector;

[0087] In this embodiment, the number of channels of the feature map output by the point cloud branch decoder is 8, and the number of channels of the feature map output by the image branch is 14. We splice the feature maps through the above corresponding relationship to obtain a feature map with a dimension of 32.

[0088] Step 6: Import the feature vector into the segmentation head network to output the point cloud segmentation result;

[0089] The segmentation head network consists of three fully connected layers and one dropout layer, with dimensions (64 -> 32 -> 14) respectively.

[0090] Step 7: Substitute the point cloud segmentation result into the loss function to calculate the cross-entropy loss, and use backpropagation derivation for the cross-entropy loss to obtain the training parameters; where the loss function is:

[0091]

[0092] In the formula, N is the number of samples, y ij is the predicted value, is the labeled value.

[0093] The present invention collects data in multiple formats to make a dataset, and introduces MIX3D for data augmentation, effectively balancing the global semantic information and local structure information of the substation, preventing overfitting, and thus improving the model training effect; uses farthest point sampling to cover all points in space as much as possible to ensure uniform sampling of the samples; introduces an image branch network to perform feature-level fusion of the image and the point cloud through the coordinate mapping relationship to enhance the segmentation accuracy of the model; finally, uses the trained network model to perform semantic segmentation on the point cloud data scanned from the substation, and then matches the segmentation result with the model constructed in the model library, thereby automatically constructing the complete three-dimensional model data of the substation.

[0094] Embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0095] Embodiments of the present application may be provided as a method, a system, or a computer program product. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0096] The present application is described with reference to the flowcharts and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and combinations of flows and / or blocks in the flowcharts and / or block diagrams, can be realized by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, such that the instructions executed by the processor of the computer or other programmable data processing device generate means for implementing the specified functions in one Figure 1 flow or multiple flows and / or blocks Figure 1 block or multiple blocks.

[0097] These computer program instructions can also be stored in a computer-readable memory that can direct a computer or other programmable data processing device to work in a specific manner, such that the instructions stored in the computer-readable memory produce a manufactured article including instruction means that implement the specified functions in one Figure 1 flow or multiple flows and / or blocks Figure 1 block or multiple blocks.

[0098] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to produce a computer-implemented process, and thus the instructions executed on the computer or other programmable device provide steps for implementing the specified functions in one Figure 1 flow or multiple flows and / or blocks Figure 1 block or multiple blocks.

[0099] The above are only the preferred embodiments of the present invention. Without departing from the technical principle of the present invention, several improvements and modifications can be made, and these improvements and modifications should also be regarded as the protection scope of the present invention.

Claims

1. A method for substation equipment point cloud segmentation based on an improved RandLA-Net, characterized in that, it includes: Receiving the scanned parts and images of the substation respectively, and obtaining point cloud data and image data from the scanned parts and images respectively; Making a training data set according to the obtained point cloud data and image data, and performing MIX3D data augmentation on the point cloud data to obtain the augmented point cloud data; wherein the training data set is divided into a training set, a validation set and a test set; Importing the augmented point cloud data and image data into a point cloud segmentation two-stream network to obtain point cloud features and image color texture features, wherein the point cloud segmentation two-stream network includes: a point cloud branch network and an image branch network; Calculating the mapping relationship between the image data and the point cloud data by substituting the point cloud data and the image data in the training data set into the camera imaging model; Fusing and splicing the point cloud features and the image color texture features according to the mapping relationship between the image data and the point cloud data to obtain a feature vector; Importing the feature vector into the segmentation head network to output the point cloud segmentation result; Substituting the point cloud segmentation result into the loss function to calculate the cross-entropy loss, and using backpropagation derivation for the cross-entropy loss to obtain training parameters; According to the training parameters, correcting the error weights in the segmentation network; Based on the corrected segmentation network, inputting the point cloud data to obtain the segmentation result of the substation equipment point cloud; The point cloud branch network is a RandLA-Net network that improves the encoder downsampling method; The image branch network is composed of a downsampled backbone network and a transposed convolution for upsampling; The image branch network uses an encoder-decoder architecture; The encoder is designed as follows: changing the stride of the 3×3 convolutional layer and the skip connection in the first residual structure of the residual network block 4 from 2 to 1, and changing the 3×3 ordinary convolutional layer to a dilated convolutional layer. Blocks 5 / 6 and block 4 adopt the same design and use different dilation rates to obtain a dilated convolutional layer residual network; Calculating and designing the dilation rate of all convolutional kernels in the dilated convolutional layer residual network according to the calculation formula of the upper limit distance of non-zero elements to obtain the optimized value of the dilation rate of each convolutional kernel in the dilated convolutional layer residual network; Among them, the decoder uses 3 transposed convolutional layers for upsampling decoding; The calculation formula of the upper limit distance of non-zero elements is: M i = max[M i+1 - 2r i , M i+1 - 2(M i+1 - r i ), r i , M 2 ≤ K Where M i represents the maximum distance of non-zero elements in the i-th layer, and M n = r n , r i represents the i-th dilation rate, and K represents the size unit of the convolution kernel; among them, in order to avoid the problem of dilated convolution, for N dilation rates [r1,...ri,...,rn] with N convolution kernels of size K*K, the constraint condition M2 ≤ K is set; The conversion formula for the shape of the transposed convolution is: n′=sn+k-2p-s In the formula, n is the height or width of the input, n’ is the height or width of the output, k is the convolutional kernel size, p is the padding, and s is the stride.

2. The method for substation equipment point cloud segmentation based on an improved RandLA-Net according to claim 1, characterized in that, the point cloud data in the training data set is labeled, and the training data set is divided into a training set, a validation set and a test set according to 7:1.5:1.5; wherein, the training set is the data sample for training the network; the validation set is used to provide validation information for hyperparameter tuning; the test set is the data sample for testing the network error.

3. The method for substation equipment point cloud segmentation based on an improved RandLA-Net according to claim 1, characterized in that, The method for performing MIX3D data augmentation on point cloud data to obtain the augmented point cloud data includes: Successively performing data augmentation on the point cloud data in each sample data in the training set by using algorithms of random subsampling, elastic distortion, random brightness and contrast enhancement, and color jitter, so as to obtain the augmented point cloud data for each sample; Mixing all the sample data with augmented point cloud data in the training set pairwise, and splicing and merging the annotations of the point cloud data to obtain the augmented point cloud data.

4. The method for segmenting substation equipment point cloud based on the improved RandLA-Net according to claim 1, characterized in that, The improved downsampling method of the RandLA-Net network is the farthest point sampling method, including: Randomly select a single point cloud A from the initial point cloud dataset i to the sampling dataset, and use the single point cloud A i as the query point to calculate the distances from the remaining point clouds to the point cloud A in the sampling dataset i ; According to the distance from the remaining point cloud data to the point cloud data A in the sampling data set i select the point cloud Aj with a distance greater than others again and put it into the sampling data set B. At this time, B = {Ai, Aj}; Determining a query point according to the remaining point cloud in the point cloud dataset A, and continuing to calculate the distance from each remaining point cloud in the point cloud dataset A to the query point, and selecting the point cloud with a distance greater than others to input into the sampling dataset B; Determining the query point in a loop, and selecting sampling data according to the determined query point until the number of data in the sampling dataset B is equal to the target data M; Among them, determining the query point includes: calculating the distance from any remaining point cloud in the point cloud dataset A to all the point clouds in the sampling dataset B, and selecting the point cloud with a distance less than others in the sampling dataset B as the query point; A is the point cloud dataset, and in the initial state, A = {A1, A2,..., AN}, and N is the total amount of all point cloud data in A in the initial state; B is the sampling dataset, and in the initial state, it is empty; M is the set target data volume.

5. The method for segmenting substation equipment point cloud based on the improved RandLA-Net according to claim 1, characterized in that, The expression of the camera imaging model is: z*I i = K*[R|t]*P i where K is the camera internal parameter, R is the rotation matrix, t is the translation variable, and P i is the point cloud data coordinate, and I i is the image pixel coordinate.

6. The method for segmenting substation equipment point cloud based on the improved RandLA-Net according to claim 1, characterized in that, The segmentation head network is composed of three fully connected layers and a dropout function layer.

7. The method for segmenting substation equipment point cloud based on the improved RandLA-Net according to claim 1, characterized in that, The expression of the loss function is: where N is the number of samples, C is the number of categories, and y ij is the predicted value, and is the labeled value.

Images