Substation three-dimensional point cloud registration method, system, device, medium and product

Abstract

The invention relates to the technical field of transformer substations, and discloses a transformer substation three-dimensional point cloud registration method, system, device, medium and product, and the method comprises the steps: collecting global point cloud data of a transformer substation, and carrying out the local feature extraction and down-sampling of a source point cloud and a target point cloud in the global point cloud data through employing a kernel point convolution network; and based on multi-head cross attention, according to the point cloud features corresponding to the candidate overlapping point cloud pairs, training point cloud data in the multiple groups of candidate overlapping point cloud pairs through a PointNet neural network to obtain an overlapping prediction network model, and determining an overlapping point cloud region through the overlapping prediction network model, thereby accurately extracting an overlapping region. And point cloud registration is carried out on the source point cloud and the target point cloud in the overlapped point cloud region based on a global registration algorithm, so that the point cloud registration precision of the overlapped region is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of substations, and particularly to a three-dimensional point cloud registration method, system, device, medium and product for substations. Background Art

[0002] With the advancement of the digital transformation of the power industry, three-dimensional point cloud modeling of substations has gradually become one of the core technologies for equipment detection and maintenance. The popularization of sensors such as three-dimensional laser scanning devices and drones has made it more convenient to obtain digital data of substation equipment. However, the equipment in substations is complex and diverse, and the environmental occlusion is relatively significant. The collected point cloud data often exhibits the characteristics of partial overlap or low overlap, resulting in difficulties in accurately registering the point cloud by existing methods.

[0003] Traditional point cloud registration algorithms require a high point cloud overlap rate to ensure the accuracy of matching point pairs. For scenarios such as substations with significant occlusion and complex equipment, the point clouds often show low overlap or missing local features. Existing point cloud registration algorithms are prone to falling into local optima, resulting in a decrease in registration accuracy. Although deep learning methods have improved, most of them focus on global feature extraction, lack of special design for partially overlapping regions, and are lacking in the recognition and in-depth mining of overlapping regions, which makes it difficult to guarantee accuracy when dealing with small overlapping regions. Summary of the Invention

[0004] In view of this, the present invention provides a three-dimensional point cloud registration method, system, device, medium and product for substations, which solves the technical problem of poor point cloud registration accuracy of existing point cloud registration algorithms.

[0005] The first aspect of the present invention provides a three-dimensional point cloud registration method for substations, including:

[0006] Collecting the global point cloud data of the substation and preprocessing the global point cloud data; wherein, the global point cloud data includes multiple groups of point cloud pairs, and each group of point cloud pairs includes a source point cloud and a target point cloud;

[0007] Performing local feature extraction and downsampling on the source point cloud and the target point cloud respectively through a kernel point convolution network to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs;

[0008] Based on multi-head cross-attention, training the PointNet neural network with the point cloud data in multiple groups of candidate overlapping point cloud pairs according to the point cloud features corresponding to the candidate overlapping point cloud pairs to obtain an overlap prediction network model;

[0009] Performing overlap prediction on the source point cloud and the target point cloud according to the overlap prediction network model, and determining the overlapping point cloud region according to the overlap prediction result;

[0010] Based on the global registration algorithm, perform point cloud registration on the source point cloud and the target point cloud within the overlapping point cloud region to obtain the point cloud registration result of the overlapping point cloud region.

[0011] Preferably, the preprocessing includes point cloud denoising, downsampling, and normalization.

[0012] Preferably, the local feature extraction and downsampling of the source point cloud and the target point cloud are respectively performed through a kernel point convolution network to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs, including:

[0013] Perform convolution operations on the source point cloud and the target point cloud respectively to obtain the convolution features corresponding to the source point cloud and the target point cloud respectively;

[0014] Determine the kernel points within the spherical neighborhoods corresponding to the source point cloud and the target point cloud respectively through voxel downsampling;

[0015] Determine the correlation coefficients according to the correlation functions between the kernel points corresponding to the source point cloud and the target point cloud respectively and multiple arbitrary points within the spherical neighborhoods;

[0016] Filter out the kernel points with correlation coefficients greater than a preset correlation coefficient threshold and the points corresponding to the relevant kernel points according to the magnitudes of the correlation coefficients;

[0017] Determine multiple groups of the candidate overlapping point cloud pairs according to the filtered kernel points and the points corresponding to the relevant kernel points;

[0018] Determine the point cloud features corresponding to multiple groups of the candidate overlapping point cloud pairs according to the convolution features corresponding to the filtered kernel points and the points corresponding to the relevant kernel points respectively.

[0019] Preferably, the PointNet neural network includes a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward neural network; the candidate overlapping point cloud pairs include candidate overlapping source point clouds and candidate overlapping target point clouds;

[0020] Based on the multi-head cross-attention, train the PointNet neural network through the point cloud data in multiple groups of the candidate overlapping point cloud pairs according to the point cloud features corresponding to the candidate overlapping point cloud pairs to obtain an overlapping prediction network model, including:

[0021] Input the point cloud features corresponding to the candidate overlapping source point clouds and the candidate overlapping target point clouds respectively into the multi-head self-attention layer, and perform feature enhancement on the point cloud features corresponding to the candidate overlapping source point clouds and the candidate overlapping target point clouds respectively based on the multi-head self-attention, and output the first enhanced point cloud features;

[0022] Input the first enhanced point cloud feature into the multi-head cross-attention layer, and enhance the feature of the fused point cloud based on the multi-head cross-attention to output a second enhanced point cloud feature;

[0023] Input the second enhanced point cloud feature into the feed-forward neural network, and enhance the feature of the first enhanced point cloud based on the non-linear activation function to obtain a third enhanced point cloud feature;

[0024] Predict the coincidence rate of the candidate overlapping source point cloud and the candidate overlapping target point cloud through a multi-layer perceptron according to the third enhanced point cloud feature;

[0025] Judge whether the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide according to the coincidence rate prediction result;

[0026] Generate an overlapping mask in the case of judging that the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide;

[0027] Perform supervised learning training through the multi-layer perceptron according to the point cloud feature and the overlapping mask corresponding to the point cloud feature to obtain the overlapping prediction network model.

[0028] Preferably, the overlapping prediction of the global point cloud data according to the overlapping prediction network model and determining the overlapping point cloud region according to the overlapping prediction result include:

[0029] Extract the point cloud features corresponding to the source point cloud and the target point cloud in each point cloud pair of the global point cloud data respectively;

[0030] For each point cloud pair, input the point cloud features corresponding to the source point cloud and the target point cloud into the overlapping prediction network model respectively, and output the overlapping prediction results of the source point cloud and the target point cloud;

[0031] Determine multiple groups of overlapping point cloud pairs according to the overlapping prediction results, and integrate multiple groups of the overlapping point cloud pairs into the overlapping point cloud region.

[0032] Preferably, the point cloud registration of the source point cloud and the target point cloud in the overlapping point cloud region based on the global registration algorithm to obtain the point cloud registration result of the overlapping point cloud region includes:

[0033] Extract multiple groups of overlapping point cloud pairs in the overlapping point cloud region;

[0034] For each group of the overlapping point cloud pairs, perform feature extraction on the overlapping source point cloud and the overlapping target point cloud in the overlapping point cloud pair through a dynamic graph convolutional neural network to obtain the local features corresponding to the overlapping source point cloud and the overlapping target point cloud respectively;

[0035] Perform high-dimensional mapping on the local features corresponding to the overlapping source point cloud and the overlapping target point cloud through a multi-layer perceptron, and determine the similarity matrix of the local features corresponding to the overlapping source point cloud and the overlapping target point cloud after high-dimensional mapping through the Softmax function; wherein, each element in the similarity matrix represents the overlapping probability of the overlapping source point cloud and the overlapping target point cloud.

[0036] Determine each of the overlapping source point clouds and the overlapping target point cloud that matches the overlapping source point cloud according to the similarity matrix, and construct multiple point cloud registration pairs.

[0037] For each of the point cloud registration pairs, determine the pose transformation matrix from the overlapping source point cloud to the overlapping target point cloud in the point cloud registration pair through the least squares method.

[0038] Perform pose transformation on the overlapping source point cloud in the point cloud registration pair according to the pose transformation matrix to obtain the point cloud registration result of the overlapping source point cloud and the overlapping target point cloud.

[0039] In a second aspect, the present invention also provides a three-dimensional point cloud registration system for a substation, including:

[0040] A point cloud acquisition module, configured to acquire the global point cloud data of the substation and preprocess the global point cloud data; wherein, the global point cloud data includes multiple groups of point cloud pairs, and each group of point cloud pairs includes a source point cloud and a target point cloud.

[0041] A point cloud feature extraction module, configured to respectively perform local feature extraction and downsampling on the source point cloud and the target point cloud through a kernel point convolution network to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs.

[0042] An overlapping prediction construction module, configured to, based on multi-head cross-attention, train a PointNet neural network through the point cloud data in multiple groups of the candidate overlapping point cloud pairs according to the point cloud features corresponding to the candidate overlapping point cloud pairs to obtain an overlapping prediction network model.

[0043] An overlapping area determination module, configured to perform overlapping prediction on the source point cloud and the target point cloud according to the overlapping prediction network model, and determine the overlapping point cloud area according to the overlapping prediction result.

[0044] A point cloud registration module, configured to perform point cloud registration on the source point cloud and the target point cloud in the overlapping point cloud area based on a global registration algorithm to obtain the point cloud registration result of the overlapping point cloud area.

[0045] In a third aspect, the present invention further provides an electronic device, which includes a memory and a processor. A computer program is stored in the memory. When the computer program is executed by the processor, the processor is caused to execute the steps of the substation three-dimensional point cloud registration method as described in the first aspect.

[0046] In a fourth aspect, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed, the steps of the substation three-dimensional point cloud registration method as described in the first aspect are implemented.

[0047] In a fifth aspect, the present invention further provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer is caused to execute the steps of the substation three-dimensional point cloud registration method as described in the first aspect.

[0048] As can be seen from the above technical solutions, the present invention collects the global point cloud data of the substation, uses a kernel point convolution network to respectively perform local feature extraction and downsampling on the source point cloud and the target point cloud in the global point cloud data, and based on multi-head cross-attention, according to the candidate overlapping point clouds, the corresponding point cloud features are trained on the point cloud data in multiple groups of candidate overlapping point clouds through a PointNet neural network to obtain an overlapping prediction network model, and the overlapping point cloud region is determined through the overlapping prediction network model, so as to accurately extract the overlapping region, and the source point cloud and the target point cloud in the overlapping point cloud region are registered through a global registration algorithm, thereby improving the point cloud registration accuracy of the overlapping region. Description of the Drawings

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0050] Figure 1 It is an application environment of a substation three-dimensional point cloud registration method provided by an embodiment of the present invention;

[0051] Figure 2 It is a flowchart of a substation three-dimensional point cloud registration method provided by an embodiment of the present invention;

[0052] Figure 3 It is a control block diagram of a substation three-dimensional point cloud registration method provided by an embodiment of the present invention;

[0053] Figure 4Schematic diagram of the training process of the overlap prediction network model provided by the embodiment of the present invention;

[0054] Figure 5 Schematic diagram of the point cloud registration process of the overlapping point cloud region provided by the embodiment of the present invention;

[0055] Figure 6 Schematic diagram of the structure of a three-dimensional point cloud registration system for a substation provided by the embodiment of the present invention;

[0056] Figure 7 Schematic diagram of the structure of an electronic device provided by the embodiment of the present invention. Detailed implementation manners

[0057] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0058] The three-dimensional point cloud registration method for a substation provided by the embodiments of the present application can be applied to the application environment as shown in Figure 1 Among them, the terminal 101 communicates with the server 102 through the network. The data storage system can store the data that the server 102 needs to process. The data storage system can be integrated on the server 102, or placed in the cloud or other network servers. The terminal 101 or the server 102 collects the global point cloud data of the substation and preprocesses the global point cloud data; among them, the global point cloud data includes multiple groups of point cloud pairs, and each group of point cloud pairs includes a source point cloud and a target point cloud; the kernel point convolution network is used to perform local feature extraction and downsampling on the source point cloud and the target point cloud respectively to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs; based on the multi-head cross-attention, according to the point cloud features corresponding to the candidate overlapping point cloud pairs, the PointNet neural network is used to train the point cloud data in the multiple groups of candidate overlapping point cloud pairs to obtain an overlap prediction network model; according to the overlap prediction network model, the source point cloud and the target point cloud are subjected to overlap prediction, and the overlapping point cloud region is determined according to the overlap prediction result; based on the global registration algorithm, the source point cloud and the target point cloud in the overlapping point cloud region are subjected to point cloud registration to obtain the point cloud registration result of the overlapping point cloud region.

[0059] The terminal 101 can be, but is not limited to, various personal computers, laptop computers, smart phones, tablet computers, etc.

[0060] The server 102 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0061] As Figure 2 shown, an embodiment of this application provides a three-dimensional point cloud registration method for a substation. Taking the method applied to Figure 1 the terminal 101 or the server 102 in

[0062] Step S1: Collect the global point cloud data of the substation and preprocess the global point cloud data. Among them, the global point cloud data includes multiple groups of point cloud pairs, and each group of point cloud pairs includes a source point cloud and a target point cloud.

[0063] Among them, the global point cloud data of the substation refers to the set of all point cloud data obtained by three-dimensional scanning technology within the overall range of the substation.

[0064] Exemplarily, the global point cloud data of the substation is collected by sensors (such as laser scanners, cameras, or drones, etc.). The collection process needs to cover multiple perspectives to ensure comprehensive coverage of key equipment and scenarios. Each group of point cloud pairs is respectively labeled as a source point cloud and a target point cloud, and is stored in a standard point cloud file format (such as PLY or PCD), which can not only ensure the integrity of the data, but also facilitate subsequent processing and analysis. Among them, the source point cloud is the initial point cloud data before registration, and the target point cloud is the reference point cloud data expected to be registered and aligned with the source point cloud.

[0065] The preprocessing steps include point cloud denoising, downsampling, and normalization processing. Among them, point cloud denoising is used to remove the noise points generated during the scanning process. First, statistical filtering analyzes the density based on the neighborhood of each point and eliminates the points that are significantly inconsistent with the surrounding points. Then, radius filtering clears the isolated points that are too far away according to the distance relationship between the point and its neighborhood.

[0066] The downsampling process is used to reduce the density of the point cloud data and reduce the computational complexity. Specifically, the voxel grid filtering method is used to uniformly downsample the point cloud, reduce the data volume, and at the same time retain the overall geometric structure.

[0067] The normalization process is used to convert the point cloud data to a unified coordinate scale for subsequent feature extraction and registration processing. Specifically, position and scale normalization operations are performed on the point cloud data to center the point cloud at the origin of the coordinate system and normalize its range to a unit cube.

[0068] Step S2: Respectively perform local feature extraction and downsampling on the source point cloud and the target point cloud through a kernel point convolution network to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs.

[0069] Among them, through the designed kernel point convolution network architecture, the network can efficiently process large-scale point cloud data and extract discriminative local features. During the feature extraction process, the network first learns the local geometric structure and neighborhood information of each point in the point cloud, and then integrates this information into a compact feature vector through convolution operations. The downsampling step further reduces the number of points in the point cloud while retaining key feature points, which helps reduce the computational burden of subsequent processing and improve the accuracy of registration. Multiple groups of candidate overlapping point cloud pairs are generated based on preliminary feature matching, and they may contain overlapping regions, which will be further verified and refined in subsequent steps. Each group of candidate overlapping point cloud pairs is accompanied by its corresponding point cloud features, which provide key information for the subsequent registration process.

[0070] Step S3: Based on multi-head cross-attention, according to the point cloud features corresponding to the candidate overlapping point cloud pairs, train the PointNet neural network through the point cloud data in multiple groups of candidate overlapping point cloud pairs to obtain an overlapping prediction network model.

[0071] Among them, input the point cloud data in the candidate overlapping point cloud pairs into the PointNet neural network. This network uses multi-head self-attention layers and multi-head cross-attention layers for in-depth learning and feature fusion of point cloud features. The multi-head self-attention layer can capture the mutual relationships between different parts within the point cloud and enhance the expressive ability of point cloud features; while the multi-head cross-attention layer further promotes the feature interaction between the source point cloud and the target point cloud and improves the accuracy of feature matching. The fused features are further processed and enhanced through a feed-forward neural network, and finally enhanced point cloud features are obtained. These features are used to train the overlapping prediction network model, which can learn and predict the overlapping degree between candidate overlapping point cloud pairs, providing a key basis for subsequent point cloud registration. During the training process, by continuously adjusting the network parameters, the prediction performance of the model is optimized to ensure its accurate judgment of overlapping point cloud pairs.

[0072] Step S4: Perform overlapping prediction on the source point cloud and the target point cloud according to the overlapping prediction network model, and determine the overlapping point cloud region according to the overlapping prediction result.

[0073] Among them, extract the prediction result of the overlapping prediction network model, and this prediction result contains the overlapping region information between the source point cloud and the target point cloud.

[0074] Step S5: Based on the global registration algorithm, perform point cloud registration on the source point cloud and the target point cloud within the overlapping point cloud region to obtain the point cloud registration result of the overlapping point cloud region.

[0075] Among them, a global registration algorithm is used to finely register the source point cloud and the target point cloud in the overlapping point cloud region. In each iteration, according to the current best estimated transformation, the source point cloud is transformed into the coordinate system of the target point cloud, and then the nearest point pairs between the two point clouds are calculated, and the transformation parameters are adjusted based on these point pairs to minimize the distance error between them. Repeat this process until a predetermined number of iterations or error convergence criteria are reached, so as to obtain the accurate registration result of the overlapping point cloud region. The registration result can be used for applications such as 3D reconstruction, fault detection, and status monitoring of substation equipment, providing strong support for the operation and maintenance management of substations.

[0076] It should be noted that in the embodiments of the present application, by collecting the global point cloud data of the substation, the kernel point convolution network is used to respectively extract local features and downsample the source point cloud and the target point cloud in the global point cloud data. Based on the multi-head cross-attention, according to the candidate overlapping point cloud pairs, the corresponding point cloud features are trained on the point cloud data in multiple groups of candidate overlapping point cloud pairs through the PointNet neural network to obtain an overlapping prediction network model, and the overlapping point cloud region is determined through the overlapping prediction network model, so as to accurately extract the overlapping region. Based on the global registration algorithm, the source point cloud and the target point cloud in the overlapping point cloud region are point cloud registered, thereby improving the point cloud registration accuracy of the overlapping region.

[0077] As Figure 3 shown, the following further describes in detail the specific implementation manner of the substation three-dimensional point cloud registration method provided by the embodiments of the present application.

[0078] In some embodiments, the kernel point convolution network is used to respectively extract local features and downsample the source point cloud and the target point cloud, obtaining multiple groups of candidate overlapping point cloud pairs and the corresponding point cloud features of the candidate overlapping point cloud pairs, including:

[0079] Step S201: Perform convolution operations on the source point cloud and the target point cloud respectively to obtain the convolution features corresponding to the source point cloud and the target point cloud respectively;

[0080] Step S202: Determine the kernel points within the spherical neighborhoods corresponding to the source point cloud and the target point cloud respectively through voxel downsampling;

[0081] Step S203: Determine the correlation coefficients according to the correlation functions between the kernel points corresponding to the source point cloud and the target point cloud respectively and multiple arbitrary points within the spherical neighborhoods;

[0082] Step S204: Screen out the kernel points and the points corresponding to the relevant kernel points whose correlation coefficients are greater than a preset correlation coefficient threshold according to the magnitudes of the correlation coefficients;

[0083] Step S205: Determine multiple groups of candidate overlapping point cloud pairs according to the screened kernel points and the points corresponding to the relevant kernel points;

[0084] Step S206: Determine point cloud features corresponding to multiple groups of candidate overlapping point cloud pairs, based on the convolution features corresponding to the selected core points and the points corresponding to the core points.

[0085] Among them, the kernel point convolution KPConv network model is used for feature extraction of the source point cloud and the target point cloud, and a series of ResNet-like blocks and strided convolutions are used to convert each point cloud pair into a set of reduced key points and features related to the key points.

[0086] For example, taking the source point cloud as an example, the kernel point convolution can be expressed by the general definition of point convolution, which is usually composed of the feature set of the source point cloud With the convolution kernel function g at the center point The general point convolution at is defined as:

[0087] (1)

[0088] In the formula, are points from the source point cloud, is the local neighborhood N of the center point x x The point in is the eigenvalue corresponding to point xi, from the point cloud Its own characteristics, is a local neighborhood of the center point x, usually a spherical neighborhood with the center point x as the center and a radius of r. The points in the neighborhood are used to participate in the convolution calculation related to the center point. , M is the total number of points in the point cloud, , D is the eigenvalue of the point cloud.

[0089] Among them, a local neighborhood of the center point x That is:

[0090] (2)

[0091] Where r is the radius.

[0092] Any point within the radius r The kernel function on It can be defined as:

[0093]

[0094] In the formula, are the kernel points generated by voxel downsampling, For local neighborhood Any point in the network, W is the parameter learned by the network. h is and The correlation function between them is defined as follows: the closer they are, the higher the correlation.

[0095] (3)

[0096] In the formula, is the correlation function, whose value is the correlation coefficient, and σ is a hyperparameter that determines how many neighboring points within the point cloud each point can perceive when calculating the convolution. The same applies to the target point cloud.

[0097] By comparing the magnitudes of the correlation coefficients, the kernel points whose correlation coefficients exceed a preset threshold and their corresponding correlation points are selected. Based on these selected kernel points and their corresponding points, multiple groups of candidate overlapping point cloud pairs are determined. These candidate overlapping point cloud pairs include the source point cloud and the target point cloud that the selected kernel points and their corresponding points belong to respectively and may overlap with each other. Further, according to the convolution features of the selected kernel points and their corresponding points respectively, the point cloud features corresponding to these candidate overlapping point cloud pairs are determined., that is, a point cloud composed of a set of key points is obtained and and its corresponding features and , where M' and N' are the number of points in the point cloud after downsampling, is the dimension after mapping in the high-dimensional feature space.

[0098] To identify the common overlapping region between the candidate overlapping source point cloud and the candidate overlapping target point cloud, it is necessary not only to establish the relationship between points within a single point cloud, but also to establish the relationship between the two point clouds, so as to identify their common part according to the feature relationship. Therefore, first, the multi-head self-attention mechanism is used to further strengthen and aggregate the context relationship between points in a single point cloud, and the multi-head cross-attention mechanism is needed to establish the connection between the two point clouds, so as to achieve information transmission.

[0099] In some embodiments, the PointNet neural network includes a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward neural network; the candidate overlapping point cloud pairs include a candidate overlapping source point cloud and a candidate overlapping target point cloud;

[0100] As Figure 4 shown, based on the multi-head cross-attention, according to the point cloud features corresponding to the candidate overlapping point cloud pairs, the PointNet neural network is trained with the point cloud data in multiple groups of candidate overlapping point cloud pairs to obtain an overlapping prediction network model, including:

[0101] Step S301: Input the point cloud features corresponding to the candidate overlapping source point cloud and the candidate overlapping target point cloud into the multi-head self-attention layer, and enhance the point cloud features corresponding to the candidate overlapping source point cloud and the candidate overlapping target point cloud based on the multi-head self-attention, and output the first enhanced point cloud feature.

[0102] Among them, the multi-head self-attention layer allows the model to concurrently focus on information at different positions in different representation subspaces of the input sequence, thereby capturing the dependency relationships between the candidate overlapping source point cloud and the candidate overlapping target point cloud.

[0103] The multi-head attention operation in each sub-layer is defined as follows:

[0104] (4)

[0105] In the formula, is the query vector, is the key vector, is the value vector, Headi represents the i-th attention head, the multi-head self-attention consists of multiple parallel attention heads, each head independently calculates the attention score and weights and sums the value vectors to capture information in different aspects, (i = 1, …, H), represents concatenation in the channel dimension, H represents the number of selected heads, which is set to 8 here, is the learnable weight matrix, which performs a linear transformation on the outputs (after concatenation) of all attention heads, allowing the model to adjust and fuse features according to the task requirements to obtain the final multi-head self-attention output.

[0106] Among them, (5)

[0107] In the formula, , , are all matrices learned by the network.

[0108] Each attention head adopts single-head dot-product attention:

[0109] (6)

[0110] In the formula, is the attention score, , d is the feature dimension, is the scaling factor. For the self-attention mechanism, where q, k, and v are all , while q in the cross-attention mechanism is , k, v are . Similarly, the same processing needs to be done on the point cloud features extracted from the target point cloud.

[0111] The multi-head self-attention will pass the input through H different single-head dot-product attention simultaneously to obtain H different attention outputs, concatenate them, and then pass through a linear transformation matrix to obtain the final multi-head self-attention output, which is the first enhanced point cloud feature.

[0112] Step S302: Input the first enhanced point cloud feature into the multi-head cross-attention layer, and perform feature enhancement on the fused point cloud feature based on the multi-head cross-attention to output the second enhanced point cloud feature.

[0113] Among them, the multi-head cross-attention layer is used to capture the dependencies between two different sequences. For example, in point cloud processing, it may be to interact the local features and the global features. By calculating the attention scores, applying the softmax function to obtain the attention weights, performing weighted summation, and finally performing multi-head concatenation and linear transformation, the multi-head cross-attention output is obtained, which is the second enhanced point cloud feature.

[0114] Step S303: Input the second enhanced point cloud feature into the feed-forward neural network, and perform feature enhancement on the first enhanced point cloud feature based on the non-linear activation function to obtain the third enhanced point cloud feature.

[0115] Among them, the feed-forward neural network usually consists of two linear layers and a non-linear activation function, and is used to perform further feature transformation and information processing on the output of the multi-head attention. In the entire feed-forward neural network, residual connections and layer normalization operations are usually added to improve the training stability and performance of the model. For example, after the multi-head self-attention layer, the multi-head cross-attention layer, and the feed-forward neural network layer, residual connections (adding the input and the output) and layer normalization operations are performed.

[0116] The feed-forward neural network is defined as follows:

[0117] (7)

[0118] In the formula, is the input feature of the l-th layer, usually a vector or a matrix, and its dimension depends on the output of the previous layer or the feature dimension of the data itself; , are both weight matrices, , are both bias parameters, is the activation function, is the output of the feed-forward neural network.

[0119] Finally, through residual connection and layer normalization, the residual connection ensures more efficient gradient flow during backpropagation by directly passing the input to the subsequent layer, thus alleviating the problem of vanishing gradients. Normalization reduces the range of data variation (i.e., reduces the variance of the data), making the input distribution of each layer more stable. This helps the neural network avoid problems such as exploding or vanishing gradients during training, thereby accelerating the network training process.

[0120] After N cycles, the final output of the attention module is and Then, by superimposing the initial features on the output features to highlight their significance, the final output is and Specifically as follows:

[0121] (8)

[0122] (9)

[0123] In the formula, and are the original features of the source point cloud and the target point cloud respectively.

[0124] Step S304: Predict the coincidence rate of the candidate overlapping source point cloud and the candidate overlapping target point cloud based on the third enhanced point cloud feature through a multi-layer perceptron.

[0125] Among them, two-layer multi-layer perceptrons (MLPs) are used to predict whether each point is an overlapping point and generate an overlap mask. Specifically, based on the third enhanced point cloud feature of each point, the points in each candidate overlapping point cloud pair are classified through a two-layer MLP network to determine whether they belong to the overlapping region.

[0126] The first layer of the MLP network takes the third enhanced point cloud feature as the input, performs feature transformation through linear transformation and the non-linear activation function ReLU, and outputs intermediate features.

[0127] The second layer takes the intermediate features as the input, performs linear transformation again, and outputs the probability that each point belongs to the overlapping region. According to this probability, a threshold is set. When the probability exceeds the threshold, it is considered that the point belongs to the overlapping region, and the corresponding overlap mask is generated.

[0128] The overlap mask is a matrix with the same dimension as the point cloud data. The element value in the matrix is 1 indicating that the point at that position belongs to the overlapping region, and 0 indicating that it does not belong to the overlapping region. Through the overlap mask, the overlapping region in the source point cloud and the target point cloud can be accurately extracted.

[0129] Specifically, the operations of the two-layer multi-layer perceptrons (MLPs) are as follows:

[0130] (10)

[0131] (11)

[0132] In the formula, and are both the outputs of the multi-layer perceptron, is the rectified linear unit function, is a common activation function, and are the feature quantities of the source point cloud and the target point cloud respectively, and are both weight matrices, and are both bias vectors.

[0133] The result is normalized through the sigmoid activation function and mapped between 0 and 1, representing the probability of overlapping points.

[0134] Step S305: According to the coincidence rate prediction result, determine whether the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide.

[0135] Among them, after obtaining the overlapping point probability, by setting a threshold, the points with an overlapping probability higher than the threshold are regarded as overlapping points, so as to determine the precise overlapping area. That is, the overlapping area and . M'' and N'' are the numbers of points in the newly extracted point clouds respectively. Through the coincidence rate prediction, the overlapping area in the three-dimensional point cloud of the substation can be automatically and accurately identified, providing a reliable basis for subsequent point cloud registration.

[0136] Step S306: When it is determined that the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide, generate an overlapping mask.

[0137] Among them, according to the determined overlapping area, using the preset mask generation rule, generate the corresponding overlapping mask. The overlapping mask is a binary representation used to clearly distinguish the overlapping area and the non-overlapping area. Among them, the part with a mask value of 1 represents the overlapping area, while the part with a mask value of 0 represents the non-overlapping area. By generating the overlapping mask, the overlapping situation between the source point cloud and the target point cloud can be more intuitively displayed, providing strong support for subsequent point cloud registration.

[0138] Step S307: Perform supervised learning training through the multi-layer perceptron according to the point cloud features and the overlapping mask corresponding to the point cloud features to obtain an overlapping prediction network model.

[0139] The point cloud features of the candidate overlapping source point cloud and the candidate overlapping target point cloud, along with their corresponding overlapping masks, are used as the input data for supervised learning and trained through a multi-layer perceptron. During the training process, the overlapping mask is used as label information to guide the multi-layer perceptron in learning how to accurately predict the overlapping region from the point cloud features. By continuously adjusting the network parameters, the difference between the prediction result and the actual overlapping mask is minimized, thereby obtaining an optimized overlapping prediction network model.

[0140] Among them, for the overlapping region evaluation task, since its essence is a binary classification problem, that is, to determine whether each point belongs to the overlapping region, cross-entropy loss is used for supervised learning. The loss function for the overlapping prediction network model can be set as:

[0141]

[0142]

[0143]

[0144] In the formula, and are both loss values, is the true overlapping mask of the i-th sample belonging to class p, is the predicted overlapping mask of the i-th sample belonging to class p, is the number of samples, is the true overlapping mask of the i-th sample belonging to class Q, is the predicted overlapping mask of the i-th sample belonging to class Q, is the loss sum value.

[0145] In some embodiments, according to the overlapping prediction network model, overlapping prediction is performed on the global point cloud data, and the overlapping point cloud region is determined according to the overlapping prediction result, including:

[0146] Step S401: Extract the point cloud features corresponding to the source point cloud and the target point cloud in each point cloud pair of the global point cloud data;

[0147] Step S402: For each point cloud pair, input the point cloud features corresponding to the source point cloud and the target point cloud into the overlapping prediction network model, and output the overlapping prediction results of the source point cloud and the target point cloud;

[0148] Step S403: Determine multiple groups of overlapping point cloud pairs according to the overlapping prediction result, and integrate the multiple groups of overlapping point cloud pairs into an overlapping point cloud region.

[0149] Among them, the overlapping prediction network model is integrated into the main network to guide the identification of the overlapping region, thereby achieving more accurate overlapping region prediction.

[0150] As Figure 5 shown, in some embodiments, based on the global registration algorithm, point cloud registration is performed on the source point cloud and the target point cloud in the overlapping point cloud region to obtain the point cloud registration result of the overlapping point cloud region, including:

[0151] Step S501: Extract multiple groups of overlapping point cloud pairs in the overlapping point cloud region.

[0152] Among them, each group of overlapping point cloud pairs contains the corresponding overlapping parts in the source point cloud and the target point cloud, that is, it contains the overlapping source point cloud and the overlapping target point cloud.

[0153] Step S502: For each group of overlapping point cloud pairs, use the dynamic graph convolutional neural network to extract the features of the overlapping source point cloud and the overlapping target point cloud in the overlapping point cloud pair, and obtain the local features corresponding to the overlapping source point cloud and the overlapping target point cloud respectively.

[0154] Among them, the dynamic graph convolutional neural network (Dynamic Graph Convolutional Neural Network, DGCNN) is used to extract the features of the point cloud data in the overlapping point cloud pair. DGCNN can dynamically construct a graph structure according to the local geometric structure of the point cloud data, and extract local features through graph convolution operations. During the feature extraction process, DGCNN will gradually aggregate the information of neighboring points, so that each point contains the local feature information of its surrounding points. Through the feature extraction of DGCNN, the local features corresponding to the overlapping source point cloud and the overlapping target point cloud can be obtained, and these local features will be used in the subsequent point cloud registration process.

[0155] Specifically, given the input point cloud , a directed graph is constructed for each point in the point cloud and its surrounding neighborhood representing the local structure of the point cloud, where and are the vertices and edges respectively. Among them, the point in is used as the vertex, is the neighboring point of in the graph structure, and then the edge feature is defined as For the neighborhood set and the point , the local feature of

[0156]

[0157] is iteratively updated as: where Denote the feature vector of point $i$ after the $g$-th iteration (or after the $g$-th layer of network processing); Denote the feature vector of point $j$ after the $g$-th iteration (or after the $g$-th layer of network processing); is the set of neighborhood points of point $i$, that is, it contains a series of points adjacent to point $i$. In scenarios such as point cloud processing, neighborhood points are often determined by distance metrics (such as Euclidean distance), etc. $\max(\cdot)$ represents element-wise max pooling which aggregates features, and $\text{cat}[\cdot,\cdot]$ represents concatenation. represents a convolution operation; The local feature of and the overall local feature . For the point cloud the feature embedding is the same, and the extracted local feature of and the overall local feature .

[0158] Step S503: Perform high-dimensional mapping on the local features corresponding to the overlapping source point cloud and the overlapping target point cloud respectively through a multi-layer perceptron, and determine the similarity matrix of the local features corresponding to the overlapping source point cloud and the overlapping target point cloud after high-dimensional mapping through the Softmax function; where each element in the similarity matrix represents the overlapping probability of the overlapping source point cloud and the overlapping target point cloud.

[0159] Among them, the local features of the overlapping source point cloud and the overlapping target point cloud are respectively used as inputs, and high-dimensional mapping is performed through a multi-layer perceptron to capture more complex feature relationships. The feature vectors after high-dimensional mapping have higher dimensions and can express richer information. Subsequently, the Softmax function is used to normalize the feature vectors after high-dimensional mapping to obtain the similarity matrix. Each element in the similarity matrix represents the similarity or overlapping probability between a point in the overlapping source point cloud and a point in the overlapping target point cloud. This similarity matrix provides key information for subsequent point cloud registration. By comparing the element values in the similarity matrix, the most likely matching point in the target point cloud for each point in the source point cloud is found, thereby determining the corresponding point pairs. This can be achieved by selecting the column index with the maximum probability in each row of the similarity matrix.

[0160] Among them, softmax is used to calculate the similarity matrix, which is defined as follows:

[0161]

[0162] In the formula, , are respectively the local features corresponding to the overlapping source point cloud and the overlapping target point cloud after high-dimensional mapping.

[0163] Step S504: Determine each overlapping source point cloud and the overlapping target point cloud that matches the overlapping source point cloud according to the similarity matrix, and construct multiple point cloud registration pairs.

[0164] Step S505: For each point cloud registration pair, determine the pose transformation matrix from the overlapping source point cloud to the overlapping target point cloud in the point cloud registration pair by the least squares method.

[0165] Among them, by calculating the optimal rigid body transformation between the source point cloud and the target point cloud, the overlapping parts can be aligned as much as possible. The least squares method is used to minimize the sum of the squares of the errors to find the best function matching of the data. Among them, the least squares method is used to solve the pose transformation matrix from the source point cloud to the target point cloud, including the rotation matrix and the translation vector, so that the overlapping part between the transformed source point cloud and the target point cloud has the minimum error.

[0166] Among them, the weighted least squares method is used to describe the point cloud registration problem:

[0167]

[0168] In the formula, is a point vector in the source point cloud in the overlapping area, is the vector in the target point cloud that matches the source point cloud in the overlapping area, is the weight of the contribution of the corresponding relationship in the rigid motion calculation, Rotation matrix, used to rotate the source point vector to change its direction to make it more match the direction of the target point cloud. Usually, it is a 3×3 orthogonal matrix, and t is the translation vector with a dimension of 3×1, which is used to translate the rotated source point vector so that its spatial position is closer to the target point vector.

[0169] is the weight coefficient, taking the maximum value of each row of the similarity matrix. Assigning different weights to the point pairs can further weaken the negative impact of incorrect correspondences. is the square of the two-norm of the vector. Given the corresponding relationship with the corresponding weight, the optimal rotation matrix R and translation vector t are solved by SVD.

[0170] Step S506: Perform pose transformation on the overlapping source point cloud in the point cloud registration pair according to the pose transformation matrix to obtain the point cloud registration result of the overlapping source point cloud and the overlapping target point cloud.

[0171] Among them, the source point cloud within the overlapping point cloud region is transformed using the obtained pose transformation matrix, including rotation and translation operations, so that the transformed source point cloud can be accurately aligned with the target point cloud in the overlapping part. Then, the transformed source point cloud and the target point cloud are fused. Through weighted averaging or other fusion strategies, the point cloud data in the overlapping part is merged into a unified point cloud representation, thereby obtaining the registered point cloud data.

[0172] Among them, for the loss function of the global registration algorithm, the loss between the predicted transformation matrix and the ground truth transformation matrix can be measured, that is:

[0173]

[0174] In the formula, is the loss value, is the identity matrix, g is the value in the true transformation matrix, and T is the matrix transpose.

[0175] It can be understood that the embodiments of the present application realize the accurate registration of the three-dimensional point cloud of the substation through steps such as extracting point cloud features, constructing candidate overlapping regions, predicting the coincidence rate, generating overlapping masks, training the overlapping prediction network model, and performing point cloud registration based on the global registration algorithm.

[0176] Based on the same inventive concept, the embodiments of the present application also provide a three-dimensional point cloud registration system for a substation for implementing the above-mentioned three-dimensional point cloud registration method for a substation.

[0177] The implementation solution provided by this system to solve the problem is similar to the implementation solution described in the above method. Therefore, the specific limitations in one or more embodiments of the three-dimensional point cloud registration system for a substation provided below can refer to the limitations on the three-dimensional point cloud registration method for a substation in the above text, and will not be repeated here.

[0178] As Figure 6 shown, the embodiments of the present application provide a three-dimensional point cloud registration system for a substation, including:

[0179] A point cloud acquisition module 100, configured to acquire the global point cloud data of the substation and preprocess the global point cloud data; among them, the global point cloud data includes multiple groups of point cloud pairs, and each group of point cloud pairs includes a source point cloud and a target point cloud;

[0180] A point cloud feature extraction module 200, configured to respectively perform local feature extraction and downsampling on the source point cloud and the target point cloud through a kernel point convolution network to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs;

[0181] The overlapping prediction construction module 300 is used to train the PointNet neural network with the point cloud data pairs in multiple groups of candidate overlapping point cloud pairs based on the point cloud features corresponding to the candidate overlapping point cloud pairs through multi-head cross-attention, so as to obtain an overlapping prediction network model;

[0182] The overlapping area determination module 400 is used to perform overlapping prediction on the source point cloud and the target point cloud according to the overlapping prediction network model, and determine the overlapping point cloud area according to the overlapping prediction result;

[0183] The point cloud registration module 500 is used to perform point cloud registration on the source point cloud and the target point cloud in the overlapping point cloud area based on the global registration algorithm to obtain the point cloud registration result of the overlapping point cloud area.

[0184] In some embodiments, the preprocessing includes point cloud denoising processing, downsampling processing, and normalization processing.

[0185] In some embodiments, the point cloud feature extraction module 200 is used for:

[0186] Performing convolution operations on the source point cloud and the target point cloud respectively to obtain the convolution features corresponding to the source point cloud and the target point cloud respectively;

[0187] Determining the kernel points within the spherical neighborhoods corresponding to the source point cloud and the target point cloud respectively through voxel downsampling;

[0188] Determining the correlation coefficients according to the correlation functions between the kernel points corresponding to the source point cloud and the target point cloud respectively and multiple arbitrary points within the spherical neighborhoods;

[0189] According to the magnitudes of the correlation coefficients, screening out the kernel points and the points corresponding to the relevant kernel points whose correlation coefficients are greater than a preset correlation coefficient threshold;

[0190] Determining multiple groups of candidate overlapping point cloud pairs according to the screened kernel points and the points corresponding to the relevant kernel points;

[0191] Determining the point cloud features corresponding to multiple groups of candidate overlapping point cloud pairs according to the convolution features corresponding to the screened kernel points and the points corresponding to the relevant kernel points respectively.

[0192] In some embodiments, the PointNet neural network includes a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward neural network; the candidate overlapping point cloud pairs include candidate overlapping source point clouds and candidate overlapping target point clouds;

[0193] The overlapping prediction construction module 300 is used for:

[0194] Input the point cloud features corresponding to the candidate overlapping source point cloud and the candidate overlapping target point cloud into the multi-head self-attention layer, and enhance the point cloud features corresponding to the candidate overlapping source point cloud and the candidate overlapping target point cloud respectively based on the multi-head self-attention, and output the first enhanced point cloud features;

[0195] Input the first enhanced point cloud features into the multi-head cross-attention layer, and enhance the fused point cloud features based on the multi-head cross-attention, and output the second enhanced point cloud features;

[0196] Input the second enhanced point cloud features into the feed-forward neural network, and enhance the first enhanced point cloud features based on the non-linear activation function to obtain the third enhanced point cloud features;

[0197] Predict the coincidence rate of the candidate overlapping source point cloud and the candidate overlapping target point cloud through a multi-layer perceptron according to the third enhanced point cloud features;

[0198] Judge whether the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide according to the coincidence rate prediction result;

[0199] Generate an overlapping mask in the case of judging that the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide;

[0200] Perform supervised learning training through a multi-layer perceptron according to the point cloud features and the overlapping mask corresponding to the point cloud features to obtain an overlapping prediction network model.

[0201] In some embodiments, the overlapping region determination module 400 is used for:

[0202] Extract the point cloud features corresponding to the source point cloud and the target point cloud respectively in each point cloud pair of the global point cloud data;

[0203] For each point cloud pair, input the point cloud features corresponding to the source point cloud and the target point cloud into the overlapping prediction network model respectively, and output the overlapping prediction results of the source point cloud and the target point cloud;

[0204] Determine multiple groups of overlapping point cloud pairs according to the overlapping prediction results, and integrate the multiple groups of overlapping point cloud pairs into an overlapping point cloud region.

[0205] In some embodiments, the point cloud registration module 500 is used for:

[0206] Extract multiple groups of overlapping point cloud pairs in the overlapping point cloud region;

[0207] For each group of overlapping point cloud pairs, perform feature extraction on the overlapping source point cloud and the overlapping target point cloud in the overlapping point cloud pair through a dynamic graph convolutional neural network to obtain the local features corresponding to the overlapping source point cloud and the overlapping target point cloud respectively;

[0208] The local features corresponding to the overlapping source point cloud and the overlapping target point cloud are subjected to high-dimensional mapping through a multi-layer perceptron, and a similarity matrix of the local features corresponding to the overlapping source point cloud and the overlapping target point cloud after high-dimensional mapping is determined through a Softmax function; wherein, each element in the similarity matrix represents the overlapping probability of the overlapping source point cloud and the overlapping target point cloud.

[0209] Each overlapping source point cloud and the overlapping target point cloud matched with the overlapping source point cloud are determined according to the similarity matrix, and a plurality of point cloud registration pairs are constructed.

[0210] For each point cloud registration pair, a pose transformation matrix from the overlapping source point cloud to the overlapping target point cloud in the point cloud registration pair is determined by the least squares method.

[0211] The overlapping source point cloud in the point cloud registration pair is subjected to pose transformation according to the pose transformation matrix to obtain the point cloud registration result of the overlapping source point cloud and the overlapping target point cloud.

[0212] As Figure 7 shown, an embodiment of the present application provides an electronic device. The electronic device 10 includes a memory 20 and a processor 30. A computer program is stored in the memory 20. When the computer program is executed by the processor 30, the processor 30 is caused to execute the steps of the substation three-dimensional point cloud registration method in the above embodiment.

[0213] An embodiment of the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed, the steps of the substation three-dimensional point cloud registration method in the above embodiment are implemented.

[0214] An embodiment of the present application provides a computer program product. The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer is caused to execute the steps of the substation three-dimensional point cloud registration method in the above embodiment.

[0215] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described system, electronic device, computer storage medium, and computer program product can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0216] It should be noted that the terms "first", "second", etc. in the description, claims and above-mentioned drawings of the present invention are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that the data used in this way can be interchanged under appropriate circumstances so that the embodiments of the present invention described here can be implemented in an order other than those illustrated or described here. In addition, the terms "comprising" and "having" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device comprising a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0217] It should be understood that although the steps in the flowcharts involved in the above-described embodiments are shown in sequence according to the indication of the arrows, these steps do not necessarily have to be executed in the order indicated by the arrows. Unless there is a clear indication in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, at least some of the steps in the flowcharts involved in the above-described embodiments may include multiple steps or multiple stages. These steps or stages do not necessarily have to be executed at the same time, but can be executed at different times. The execution order of these steps or stages does not necessarily have to be sequential, but can be executed alternately or in turn with at least some of the steps or stages in other steps or other steps.

[0218] In several embodiments provided by the present invention, it should be understood that the disclosed system, electronic device, computer storage medium, computer program product and method can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed coupling or direct coupling or communication connection between each other can be through some interfaces, and the indirect coupling or communication connection of the device or unit can be in an electrical, mechanical or other form.

[0219] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0220] In addition, in each embodiment of the present invention, each functional unit can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit. The above integrated unit can be implemented in the form of hardware or in the form of a software functional unit.

[0221] If the above integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present invention. The foregoing storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (English full name: Read-Only Memory, English abbreviation: ROM), random access memories (English full name: Random Access Memory, English abbreviation: RAM), magnetic disks, or optical discs.

[0222] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of each embodiment of the present invention.

Claims

1. A three-dimensional point cloud registration method for a substation, characterized in that, Including: Collecting the global point cloud data of a substation and preprocessing the global point cloud data; wherein, the global point cloud data includes multiple groups of point cloud pairs, and each group of point cloud pairs includes a source point cloud and a target point cloud; Separately performing local feature extraction and downsampling on the source point cloud and the target point cloud through a kernel point convolution network to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs; Based on multi-head cross-attention, training a PointNet neural network through the point cloud data in multiple groups of the candidate overlapping point cloud pairs according to the point cloud features corresponding to the candidate overlapping point cloud pairs to obtain an overlapping prediction network model; Performing overlapping prediction on the source point cloud and the target point cloud according to the overlapping prediction network model, and determining an overlapping point cloud region according to the overlapping prediction result; Performing point cloud registration on the source point cloud and the target point cloud in the overlapping point cloud region based on a global registration algorithm to obtain a point cloud registration result of the overlapping point cloud region.

2. The three-dimensional point cloud registration method for a substation according to claim 1, wherein The preprocessing includes point cloud denoising processing, downsampling processing, and normalization processing.

3. The three-dimensional point cloud registration method for a substation according to claim 1, wherein The separately performing local feature extraction and downsampling on the source point cloud and the target point cloud through a kernel point convolution network to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs includes: Separately performing convolution operations on the source point cloud and the target point cloud to obtain the convolution features corresponding to the source point cloud and the target point cloud respectively; Determining kernel points within the spherical neighborhoods corresponding to the source point cloud and the target point cloud respectively through voxel downsampling; Determining a correlation coefficient according to the correlation function between the kernel points corresponding to the source point cloud and the target point cloud respectively and multiple arbitrary points within the spherical neighborhood; Filtering out the kernel points with the correlation coefficient greater than a preset correlation coefficient threshold and the points corresponding to the relevant kernel points; Determining multiple groups of the candidate overlapping point cloud pairs according to the filtered kernel points and the points corresponding to the relevant kernel points; Determining the point cloud features corresponding to multiple groups of the candidate overlapping point cloud pairs according to the convolution features corresponding to the filtered kernel points and the points corresponding to the relevant kernel points respectively.

4. The three-dimensional point cloud registration method for a substation according to claim 1, wherein The PointNet neural network includes a multi-head self-attention layer, a multi-head cross-attention layer, and a feed-forward neural network; the candidate overlapping point cloud pairs include a candidate overlapping source point cloud and a candidate overlapping target point cloud; The training the PointNet neural network through the point cloud data in multiple groups of the candidate overlapping point cloud pairs based on multi-head cross-attention according to the point cloud features corresponding to the candidate overlapping point cloud pairs to obtain an overlapping prediction network model includes: Inputting the point cloud features corresponding to the candidate overlapping source point cloud and the candidate overlapping target point cloud respectively into the multi-head self-attention layer, and enhancing the features of the point cloud features corresponding to the candidate overlapping source point cloud and the candidate overlapping target point cloud respectively based on multi-head self-attention, and outputting a first enhanced point cloud feature; Inputting the first enhanced point cloud feature into the multi-head cross-attention layer, and enhancing the features of the fused point cloud feature based on multi-head cross-attention, and outputting a second enhanced point cloud feature; Input the second enhanced point cloud feature into the feed-forward neural network, and enhance the feature of the first enhanced point cloud feature based on a non-linear activation function to obtain a third enhanced point cloud feature; Predict the coincidence rate of the candidate overlapping source point cloud and the candidate overlapping target point cloud through a multi-layer perceptron according to the third enhanced point cloud feature; Judge whether the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide according to the coincidence rate prediction result; Generate an overlapping mask in the case of judging that the candidate overlapping source point cloud and the candidate overlapping target point cloud coincide; Perform supervised learning training through the multi-layer perceptron according to the point cloud feature and the overlapping mask corresponding to the point cloud feature to obtain the overlapping prediction network model.

5. The 3D point cloud registration method for a substation according to claim 1, wherein The overlapping prediction of the global point cloud data according to the overlapping prediction network model, and determining the overlapping point cloud region according to the overlapping prediction result, includes: Extract the point cloud features corresponding to the source point cloud and the target point cloud in each point cloud pair of the global point cloud data; For each point cloud pair, input the point cloud features corresponding to the source point cloud and the target point cloud into the overlapping prediction network model respectively, and output the overlapping prediction results of the source point cloud and the target point cloud; Determine multiple groups of overlapping point cloud pairs according to the overlapping prediction result, and integrate multiple groups of the overlapping point cloud pairs into the overlapping point cloud region.

6. The 3D point cloud registration method for a substation according to claim 1 or 5, characterized in that The point cloud registration of the source point cloud and the target point cloud in the overlapping point cloud region based on the global registration algorithm to obtain the point cloud registration result of the overlapping point cloud region, includes: Extract multiple groups of overlapping point cloud pairs in the overlapping point cloud region; For each group of the overlapping point cloud pairs, extract features of the overlapping source point cloud and the overlapping target point cloud in the overlapping point cloud pair through a dynamic graph convolutional neural network to obtain the local features corresponding to the overlapping source point cloud and the overlapping target point cloud respectively; Perform high-dimensional mapping on the local features corresponding to the overlapping source point cloud and the overlapping target point cloud respectively through a multi-layer perceptron, and determine the similarity matrix of the local features corresponding to the overlapping source point cloud and the overlapping target point cloud after high-dimensional mapping through a Softmax function; wherein, each element in the similarity matrix represents the overlapping probability of the overlapping source point cloud and the overlapping target point cloud; Determine each overlapping source point cloud and the overlapping target point cloud matched with the overlapping source point cloud according to the similarity matrix, and construct multiple point cloud registration pairs; For each point cloud registration pair, determine the pose transformation matrix from the overlapping source point cloud to the overlapping target point cloud in the point cloud registration pair through the least squares method; Perform pose transformation on the overlapping source point cloud in the point cloud registration pair according to the pose transformation matrix to obtain the point cloud registration result of the overlapping source point cloud and the overlapping target point cloud.

7. A three-dimensional point cloud registration system for a substation, characterized in that, Includes: A point cloud acquisition module for acquiring the global point cloud data of the substation and preprocessing the global point cloud data; wherein, the global point cloud data includes multiple groups of point cloud pairs, and each group of the point cloud pairs includes a source point cloud and a target point cloud; A point cloud feature extraction module, which is used to perform local feature extraction and downsampling on the source point cloud and the target point cloud respectively through a kernel point convolution network, to obtain multiple groups of candidate overlapping point cloud pairs and the point cloud features corresponding to the candidate overlapping point cloud pairs; An overlapping prediction construction module, which is used to train a PointNet neural network based on the point cloud data in multiple groups of the candidate overlapping point cloud pairs according to the point cloud features corresponding to the candidate overlapping point cloud pairs by using multi-head cross-attention, to obtain an overlapping prediction network model; An overlapping region determination module, which is used to perform overlapping prediction on the source point cloud and the target point cloud according to the overlapping prediction network model, and determine the overlapping point cloud region according to the overlapping prediction result; A point cloud registration module, which is used to perform point cloud registration on the source point cloud and the target point cloud in the overlapping point cloud region based on a global registration algorithm, to obtain the point cloud registration result of the overlapping point cloud region.

8. An electronic device, characterized in that, The electronic device includes a memory and a processor. When the computer program stored in the memory is executed by the processor, the processor is caused to execute the steps of the substation three-dimensional point cloud registration method according to any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed, the steps of the substation three-dimensional point cloud registration method according to any one of claims 1-6 are implemented.

10. A computer program product, characterized in that, The computer program product includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions. When the program instructions are executed by a computer, the computer is caused to execute the steps of the substation three-dimensional point cloud registration method according to any one of claims 1-6.

Images