Cable working well three-dimensional model construction method and system based on point cloud data

Abstract

The invention provides a cable working well three-dimensional model construction method and system based on point cloud data, and belongs to the field of underground cable working wells, and the method comprises the steps: collecting original point cloud data and auxiliary observation data of an underground power pipeline; performing three-dimensional tunnel point cloud multi-class target segmentation on the original point cloud data by adopting a working well point cloud semantic segmentation model; wherein the work well point cloud semantic segmentation model adopts a trained PointNet + + codec based on self-attention, and the PointNet + + codec based on self-attention comprises a PointNet + + codec backbone network, a lightweight self-attention module, a multi-scale feature pyramid network and a graph optimization module; and reconstructing the three-dimensional model of each component based on the segmentation result of the multiple types of targets, and splicing the three-dimensional models of each component according to the position information to generate the three-dimensional model of the cable working well. According to the invention, accurate reconstruction of the three-dimensional model of the cable working well is realized.

Description

Technical Field

[0001] This invention belongs to the field of underground cable manholes, and particularly relates to a method and system for constructing a three-dimensional model of a cable manhole based on point cloud data. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of digitalization of above-ground power grid facilities, a vast network of power grid information resources has been formed, covering power plants, transmission and distribution lines, substations, dedicated lines, low-voltage distribution areas, and their interconnected network topologies. Underground cables are a crucial component of this network resource, contributing to the comprehensive coverage of electricity information collection systems. However, due to their underground location, complex environment, technological limitations, difficulties in data collection, susceptibility to cable damage, and numerous safety hazards, the quality of underground cable information is currently inconsistent. Given the characteristics of underground cable management—numerous elements, large data volume, dispersed data, and multiple connected systems—non-contact measurement to provide comprehensive, detailed, accurate, traceable, and high-quality 3D data modeling of underground cables is an effective way to achieve cable data monitoring and management.

[0004] However, existing technologies have the following drawbacks: First, the severe obstruction, poor lighting conditions, and limited operating space in underground spaces make existing image-based information acquisition methods highly susceptible to the effects of light and dust, resulting in poor imaging quality, inaccurate data acquisition, and incomplete reconstruction results. Second, cable manholes contain numerous equipment structures with significant differences in size and shape, such as slender structures like cable ducts, making geometric accuracy and topological relationships prone to errors during reconstruction. Third, the sheer volume of data required for 3D reconstruction of the acquired data creates computational pressure, placing extremely high demands on hardware and making it unsuitable for deployment with portable equipment on construction sites. Furthermore, the acquired data includes numerous irrelevant objects and non-target structures, causing the reconstructed model to contain a large number of non-target structures, interfering with and distorting the geometry of the cable itself, and reducing model accuracy. Summary of the Invention

[0005] To overcome the shortcomings of the existing technologies, this invention proposes a method and system for constructing a 3D model of a cable well based on point cloud data. The method collects point cloud data based on environmental issues such as dim lighting, and uses a semantic segmentation model to capture local geometric features and better understand the overall features of the scene's global context to obtain segmentation maps of each component. The method of reconstructing each component separately and then integrating them reduces the amount of reconstruction computation while improving reconstruction accuracy.

[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions: In a first aspect, this invention discloses a method for constructing a three-dimensional model of a cable manhole based on point cloud data, comprising: The original point cloud data and auxiliary observation data of underground power lines were collected. The original point cloud data consisted of three-dimensional laser scanning data of underground cable manholes and cable tunnels collected on site, while the auxiliary observation data consisted of location information collected on site for the positioning and orientation of underground cable point clouds. A manhole point cloud semantic segmentation model is used to segment the original point cloud data into three-dimensional tunnel point clouds for multi-class targets. The manhole point cloud semantic segmentation model uses a pre-trained PointNet++ codec based on self-attention. The PointNet++ codec based on self-attention includes: a PointNet++ codec backbone network, a lightweight self-attention module, a multi-scale feature pyramid network, and a graph optimization module. Based on the results of the multi-target segmentation, the three-dimensional models of each component are reconstructed, and the three-dimensional models of each component are spliced ​​together according to the location information to generate a three-dimensional model of the cable well.

[0007] Secondly, this invention discloses a three-dimensional model construction system for cable manholes based on point cloud data, comprising: The data acquisition module is used to collect raw point cloud data and auxiliary observation data of underground power lines. The raw point cloud data is the three-dimensional laser scanning data of underground cable manholes and cable tunnels collected on site, and the auxiliary observation data is the location information collected on site for the positioning and orientation of underground cable point clouds. The semantic segmentation module is used to perform multi-class target segmentation of the original point cloud data in three-dimensional tunnel point cloud using the manhole point cloud semantic segmentation model; wherein, the manhole point cloud semantic segmentation model adopts a trained self-attention-based PointNet++ encoder and decoder, which includes: PointNet++ encoder and decoder backbone network, lightweight self-attention module, multi-scale feature pyramid network and graph optimization module. The 3D reconstruction module is used to reconstruct the 3D models of each component based on the results of the multi-target segmentation, and to stitch the 3D models of each component together according to the location information to generate a 3D model of the cable well.

[0008] Thirdly, the present invention discloses an electronic device, including a memory and a processor, and computer instructions stored in the memory and running on the processor, wherein the computer instructions, when run by the processor, complete the steps of the above-mentioned method for constructing a 3D model of a cable well based on point cloud data.

[0009] Fourthly, the present invention discloses a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described method for constructing a 3D model of a cable well based on point cloud data.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention provides a method for constructing a three-dimensional model of a cable manhole. The entire construction process is divided into two main steps: segmentation and reconstruction. First, a high-precision segmentation model is used to distinguish various components within the cable manhole. Then, each component is reconstructed in a targeted manner, thereby improving the accuracy of model reconstruction.

[0011] The segmentation model proposed in this invention introduces a self-attention mechanism and multi-scale feature fusion. The self-attention mechanism allows the model to dynamically focus on more relevant points in the global context when calculating features, rather than being limited to points within a fixed physical radius, thus better understanding long-distance dependencies and complex structures. Furthermore, a multi-scale feature pyramid network structure is added at different levels of the encoder, not only passing features to the decoder but also explicitly constructing the multi-scale feature pyramid and performing predictions or feature enhancements independently at each scale. The combination of these two approaches improves both the detailed delineation of each component and the extraction of complete semantic features from long-distance, large-scale components with special shapes such as slender ones.

[0012] Furthermore, reconstructing each segmented component significantly reduces the computational load and improves reconstruction accuracy. Moreover, since reconstruction is based on the segmented results, it eliminates the influence of non-target structures, further enhancing reconstruction precision.

[0013] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0014] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0015] Figure 1 This is a flowchart of the method for constructing a 3D model of a cable well based on point cloud data, as described in Embodiment 1 of the present invention.

[0016] Figure 2 This is a schematic diagram of the path-layout modeling process described in Embodiment 1 of the present invention. Detailed Implementation

[0017] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0018] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations of the present invention.

[0019] Where there is no conflict, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0020] Example 1 In one or more embodiments, a method for constructing a 3D model of a cable manhole based on point cloud data is disclosed, such as... Figure 1 As shown, it includes the following steps: Step S1: Collect raw point cloud data and auxiliary observation data of underground power lines. The raw point cloud data is three-dimensional laser scanning data of underground cable manholes and cable tunnels collected on site, and the auxiliary observation data is location information collected on site for the positioning and orientation of underground cable point clouds.

[0021] Step S2: Use the manhole point cloud semantic segmentation model to segment the three-dimensional laser scanning data into three-dimensional tunnel point cloud multi-targets; multi-targets include walls, pipes, pipe supports, and small components (other components with complex and irregular structures, such as cross-connection boxes, transformer boxes, pipe clamps, light tubes, and stairs, etc.).

[0022] The point cloud semantic segmentation model for manholes uses a pre-trained PointNet++ codec based on self-attention. The PointNet++ codec based on self-attention includes: a PointNet++ codec backbone network, a lightweight self-attention module, a multi-scale feature pyramid network, and a graph optimization module.

[0023] Specifically, the input layer first feeds the raw point cloud data into the encoder network, which consists of three convolutional blocks. Each convolutional block includes interconnected convolutional layers, a Rectified Linear Unit (ReLU) activation function, a batch normalization layer, and a max pooling layer. A lightweight self-attention module is then connected after each convolutional block. Based on the local geometric features extracted by the PointNet++ backbone network, this module applies a self-attention mechanism to the point set of the current convolutional block, allowing each point to interact with features from all points in the same layer.

[0024] The encoder network has a specific structure consisting of a first convolutional block, a first self-attention module, a second convolutional block, a second self-attention module, a third convolutional block, and a third self-attention block connected in sequence. The first self-attention module outputs the first-level encoded features, the second self-attention module outputs the second-level encoded features, and the third self-attention module outputs the third-level encoded features. The input and output dimensions of the self-attention module remain unchanged, but the global context awareness of the features is enhanced before they are passed to the next level or to the multi-scale feature pyramid network.

[0025] In this embodiment, convolutional layers use a set of learnable kernels (filters) to perform convolution operations, transforming the input data into feature maps. Specifically, convolution operations are performed between the input and the kernels, where each kernel slides across the input by one step, producing the spatial output of the convolutional layer. The output size of the convolutional layer depends on the input size, padding, kernel size, and stride. Padding is done using the same method, i.e., maintaining the size of the kernel convolutional window by traversing all inputs with zero padding. After the convolutional layers, the good performance of ReLU is used to activate the elements of the feature maps, followed by batch normalization to simplify computation, speed up the training process, and improve the robustness of the model. Subsequently, max-pooling is used as the downsampling core to reduce the size of the feature maps. Max-pooling extracts the maximum value in the pooling window based on the input size, padding size, kernel size, and stride, and slides the output across the input. The advantage of max-pooling is that it can reduce the computational load and prevent overfitting.

[0026] The decoder network consists of three deconvolutional layers (first, second, and third deconvolutional layers), which perform upsampling operations to amplify the input and extract high-resolution features. A deconvolutional layer is considered a transposed deconvolutional layer that transforms a coarse input into a dense output based on the input size, padding, and stride.

[0027] The multi-scale feature pyramid network is connected between the encoder network and the decoder network. It receives and fuses feature maps from different levels (i.e. different scales) of the encoder. It upsamples high-level semantic features through a top-down path and fuses them with shallow features of the same scale from the encoder through lateral connections to generate a set of enhanced multi-scale features.

[0028] Specifically, the multi-scale feature pyramid network receives the third encoded feature as the top-level feature P3. The top-level feature is upsampled so that its feature map size is the same as the second encoded feature size. The upsampled top-level feature is then element-wise added to the second encoded feature. The resulting feature is then processed by a convolutional block to smooth the fusion, generating the enhanced mid-level feature P2. The convolutional block includes a convolutional layer, a rectified linear unit (ReLU) activation function, and batch normalization. The upsampled mid-level feature is then element-wise added to the first encoded feature. The resulting feature is then processed by a convolutional block to generate the enhanced bottom-level feature P1.

[0029] Furthermore, a multi-scale feature pyramid network is skip-connected to the decoder. Top-level features are input to the decoder and upsampled through a first deconvolutional layer to obtain the first decoded feature. The first decoded feature is concatenated with mid-level features and upsampled through a second deconvolutional layer to obtain the second decoded feature. At this point, the first decoded feature provides high-level semantic guidance, providing semantic information about the category of the entire object, while the mid-level features provide corresponding geometric details. The second decoded feature is concatenated with bottom-level features and upsampled through a second deconvolutional layer to restore the number of points in the original point cloud. Here, a richer semantic context and the highest-resolution shallow geometric features are fused to finely restore details such as object edges and small objects. Finally, a fully connected layer with shared weights outputs the classification score for each point, completing semantic segmentation.

[0030] The first encoded feature has the highest resolution, containing rich geometric details such as edges and corners, but weak semantic information. The second encoded feature balances detail and semantics. The third encoded feature has the richest semantic information, having undergone multiple abstractions, and possesses the strongest semantic information, capable of distinguishing long-distance, large-area structures such as cables and supports, but suffers from significant detail loss. This embodiment obtains a set of enhanced multi-scale features (P1, P2, P3) through a multi-scale feature pyramid network, preserving high-resolution details from the shallow layers while injecting strong semantic information from the deeper layers. During decoding, when directly using the encoder's original features for skip connections, the shallow features have weak semantics. The multi-scale feature pyramid network acts as a semantic enhancer, injecting high-level semantics into each scale, enabling the enhanced features to possess both detail and semantics, better matching the decoder features, greatly reducing the difficulty of interpolation during decoder upsampling, and making the recovered details more accurate.

[0031] This embodiment collects point cloud data of underground cable manholes in a certain area. The point cloud scene of the cable manhole mainly includes the ground, roof, ladder, wall, cable, support, etc. The CloudCompare point cloud processing software is used to label the category attributes of the cable manhole point cloud and build a training dataset for training the above segmentation model.

[0032] Step S3: Based on the results of the multi-target segmentation, the three-dimensional model of the cable well is generated by splicing the data according to the location information.

[0033] Step S3-1: Rebuild the walls, pipes, pipe supports, and small components.

[0034] I. Extraction and reconstruction of wall boundary features.

[0035] The segmented wall point cloud data is further divided into four wall surfaces for boundary feature extraction. Specifically, before boundary feature extraction, the topological relationships between the point clouds need to be organized. R is used. The `-tree` dynamic spatial index structure organizes the topological relationships between point clouds. It uses 3D R... The `-tree` function dynamically clusters and divides the point cloud data of the utility tunnel, enabling fast and accurate querying of local surface reference point sets. The data complexity of each layer increases with the depth of the layer.

[0036] Using R The -tree dynamic spatial index structure reorganizes the wall point cloud data. It also uses the KNN nearest neighbor algorithm to search for the k nearest neighbors of each point in the point cloud. Each point and its neighbors form a local reference point set. Then, a micro-tangent plane is fitted using the least squares method. The reference point set is then projected onto the micro-tangent plane, and the maximum angle between the lines connecting each projected point and its neighbors is determined. This is used to identify the boundary features of the wall point cloud. The specific process is as follows: Fitting the micro-tangent plane of the local surface reference point set. Let the matrix equation be Ac=0, specifically as shown in formula (1), and its plane equation be F(x,y,z)=c1x+c2x+c3x+c4=0: (1) Solve the equation Ac=0, and use the eigenvector estimation method to estimate matrix A. T A undergoes singular value decomposition, i.e.: (2) In the formula, U and V are orthogonal matrices; A T The eigenvectors of A are w1, w2, w3, w4; when A T When A has a minimum eigenvalue, its corresponding eigenvector is a solution to the matrix equation Ac=0.

[0037] Calculation of the angle between adjacent points. Let the coordinates of a point in the original point set be (x...). i ,y i ,z i If ), then the point projected onto the plane is (x). i ’ , yi ’ ,z i ’ In the process of calculating the angle between vectors, P is used. i ’ (i=0,1,...,k) as the starting point, and N j Using '(j=0,1,..., k-1) as the endpoint, calculate vector P. i 'N j 'Calculate the cross product v with the normal vector of the microtangent plane, and then solve for the angles between other vectors and Pi'N'j and v.' , ,when hour, After traversing all points in the point set, sort them according to the principle of descending order. If the calculation results are sorted, then the angle between adjacent vectors... As shown in the following formula: (3) Set an angle threshold When the angle between adjacent vectors If the value exceeds this threshold, the point can be identified as a boundary feature point. Boundary curves are fitted using the extracted boundary feature points, and then each boundary line forms a complete wall surface, creating a solid 3D model of the wall.

[0038] II. Reconstruct the pipeline model.

[0039] From the perspective of 3D modeling, the pipes in the underground utility tunnel are three-dimensional homogeneous tubular objects formed by scanning a two-dimensional cross-section along a certain path. Therefore, the pipes have the characteristics of being long and narrow and being parameterizable. Based on this characteristic, rapid pipe modeling can be achieved.

[0040] This embodiment uses a path-based lofting modeling method, such as... Figure 2 The pipeline model is constructed by obtaining the pipeline's centerline and cross-sectional curve shape information. A segmented centerline fitting method is used to extract the pipeline centerline. This method involves first segmenting the pipe point cloud data, extracting the center point of each segment in 3D, and finally connecting these center points to form the pipeline's centerline. Based on the extracted centerline and the pipeline diameter, a pipeline can be fitted. Specifically: (1) Centerline extraction.

[0041] This embodiment uses a segmented centerline fitting method to extract the centerline of the entire pipeline. Let P be a set of segmented points. i (i=1,2,3,...,n), the maximum value in each coordinate axis direction of the point set is represented by x.max y max , z max The minimum value is represented by x. min y min , z min The formulas for calculating the length, width, and height of a circumscribed cuboid are as follows: (4) In the formula, l, w, h The length, width, and height of the circumscribed cuboid are given. The segmentation direction and spacing of the pipe point cloud are determined using the length, width, and height of the circumscribed cuboid. Finally, the center point of each region is calculated, and the centerline of the pipe is obtained by connecting the center points with a smooth curve. The formula for calculating the coordinates of the center point of a region is as follows: (5) In the formula, n represents the number of point clouds in each segmented region, (x i , y i ,z i ) represents the coordinates of the i-th point cloud in the current segmented region.

[0042] (2) Input diameter parameters.

[0043] After obtaining the centerline of the pipe, the external cross-section needs to be fitted using the pipe diameter parameters. The external cross-section is actually a regular circular surface. Since the centerlines are not on the same plane, translation and rotation operations are required based on the position of each center point on the centerline to scan out a 3D pipe. The rotation methods include rotation along the X-axis, rotation along the Y-axis, and rotation along the Z-axis, with rotation matrices R0, R1, and R2, respectively. X ,R Y ,R Z As shown in the following formula: (6) (7) (8) In the formula, The rotation angle is used. After translation and rotation, it is ensured that the tangent direction of each centerline segment is consistent with the normal direction of the cross-section. Then, the final 3D model of the pipe is obtained by scanning according to the diameter parameters corresponding to each centerline segment using the path lofting modeling method.

[0044] III. Clustering and Reconstruction of Pipe Supports

[0045] The pipe supports in the scene will be reconstructed based on the semantic segmentation results. The support reconstruction mainly consists of two steps: First, the support point cloud data corresponding to the pipe support labels after semantic segmentation is extracted, and the pipe supports in the underground utility tunnel scene are divided and clustered one by one; Second, the support model is automatically drawn using the trilateral surface reconstruction method based on the clustered support point cloud data.

[0046] (1) Clustering of pipe supports.

[0047] The pipe supports in underground utility tunnels are independent entities, all with the same shape but varying locations. Creating a separate model for each support would be a massive undertaking. Therefore, this embodiment employs Euclidean clustering to unify the point cloud of the supports. Euclidean clustering measures the positional relationship between two points based on Euclidean distance, which is a straight-line distance in Euclidean space. In three-dimensional space, its formula is: (9) Based on the preset minimum number of cluster points and distance threshold, scaffold clustering is completed for each point.

[0048] (2) Reconstruction of pipe supports.

[0049] After clustering the point clouds of each support structure, the support structure model is quickly generated using the trilateral surface reconstruction method. However, due to the varying quantity and quality of point clouds for each support structure, performing situational modeling for each structure results in inconsistent quality of the 3D model output. Therefore, to ensure modeling quality, upsampling is performed on the original point clouds. Upsampling effectively addresses the sparsity and irregularity issues of the original point clouds, significantly improving the creation of the 3D model.

[0050] IV. Reconstruction of small component models.

[0051] For other components with complex and irregular structures, such as interconnection boxes, transformer boxes, pipe clamps, light tubes, and stairs, the models fitted using automated modeling methods are unlikely to produce high-quality, accurate 3D models. Therefore, it is necessary to construct 3D models of each small component according to the actual dimensions using professional 3D modeling software based on the dimensional data and images collected on-site. Tools such as AutoCAD, 3ds Max, PointCloud, Geomagic, ZBrush, Cyclone, and Rhino are required.

[0052] Step S3-2: Integrate the reconstructed walls, pipes, pipe supports, and other components into a complete 3D model. Specifically, align the spatial coordinates and map them to the actual point cloud position information. Then, import the aligned components into a unified 3D scene or model file to obtain a complete 3D model.

[0053] Example 2 In one or more embodiments, a 3D model construction system for cable manholes based on point cloud data is disclosed, specifically including: The data acquisition module is used to collect raw point cloud data and auxiliary observation data of underground power lines. The raw point cloud data is the three-dimensional laser scanning data of underground cable manholes and cable tunnels collected on site, and the auxiliary observation data is the location information collected on site for the positioning and orientation of underground cable point clouds. The semantic segmentation module is used to perform multi-class target segmentation of the original point cloud data in three-dimensional tunnel point cloud using the manhole point cloud semantic segmentation model; wherein, the manhole point cloud semantic segmentation model adopts a trained self-attention-based PointNet++ encoder and decoder, which includes: PointNet++ encoder and decoder backbone network, lightweight self-attention module, multi-scale feature pyramid network and graph optimization module. The 3D reconstruction module is used to reconstruct the 3D models of each component based on the results of the multi-target segmentation, and to stitch the 3D models of each component together according to the location information to generate a 3D model of the cable well.

[0054] Example 3 This embodiment provides an electronic device, including a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the computer instructions are executed by the processor, they complete the steps of the above-described method for constructing a 3D model of a cable well based on point cloud data.

[0055] Example 4 This embodiment provides a computer-readable storage medium for storing computer instructions, which, when executed by a processor, complete the steps of the above-described method for constructing a 3D model of a cable well based on point cloud data.

[0056] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0057] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0058] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0059] The descriptions of each embodiment in the above embodiments have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0060] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a 3D model of a cable manhole based on point cloud data, characterized in that, include: The original point cloud data and auxiliary observation data of underground power lines were collected. The original point cloud data consisted of three-dimensional laser scanning data of underground cable manholes and cable tunnels collected on site, while the auxiliary observation data consisted of location information collected on site for the positioning and orientation of underground cable point clouds. A manhole point cloud semantic segmentation model is used to segment the original point cloud data into three-dimensional tunnel point clouds for multi-class targets. The manhole point cloud semantic segmentation model uses a pre-trained PointNet++ codec based on self-attention. The PointNet++ codec based on self-attention includes: a PointNet++ codec backbone network, a lightweight self-attention module, a multi-scale feature pyramid network, and a graph optimization module. Based on the results of the multi-target segmentation, the three-dimensional models of each component are reconstructed, and the three-dimensional models of each component are spliced ​​together according to the location information to generate a three-dimensional model of the cable well.

2. The method for constructing a 3D model of a cable manhole based on point cloud data as described in claim 1, characterized in that, The PointNet++ codec backbone network includes the encoder network and the decoder network; The raw point cloud data is input into the encoder network, which includes three convolutional blocks. Each convolutional block includes a concatenated convolutional layer, a rectified linear unit activation function, a batch normalization layer, and a max pooling layer. A lightweight self-attention module is connected after each convolutional block. Specifically, the encoder network includes a first convolutional block, a first self-attention module, a second convolutional block, a second self-attention module, a third convolutional block, and a third self-attention block connected in sequence; the first self-attention module outputs first-level encoded features, the second self-attention module outputs second-level encoded features, and the third self-attention module outputs third-level encoded features.

3. The method for constructing a 3D model of a cable manhole based on point cloud data as described in claim 2, characterized in that, The decoder network includes three deconvolutional layers: a first deconvolutional layer, a second deconvolutional layer, and a third deconvolutional layer. The deconvolutional layers perform upsampling operations to amplify the input and extract high-resolution features.

4. The method for constructing a 3D model of a cable manhole based on point cloud data as described in claim 1, characterized in that, The multi-scale feature pyramid network is connected between the encoder network and the decoder network. It receives and fuses feature maps from different levels of the encoder, upsamples high-level semantic features through a top-down path, and fuses them with shallow features of the same scale from the encoder through lateral connections to generate a set of enhanced multi-scale features.

5. The method for constructing a 3D model of a cable manhole based on point cloud data as described in claim 4, characterized in that, The multi-scale feature pyramid network receives the third encoded feature as the top-level feature; The top-level features are upsampled so that their feature map size is the same as that of the second encoded features. The upsampled top-level features are then added element-wise to the second encoded features. The added features are then processed by convolutional blocks to generate enhanced mid-level features. The upsampled mid-level features are then added element-wise to the first encoded features. The added features are then processed by convolutional blocks to generate enhanced bottom-level features.

6. The method for constructing a 3D model of a cable manhole based on point cloud data as described in claim 1, characterized in that, The multi-scale feature pyramid network is connected to the decoder in a skip connection. The top-level features are input to the decoder and upsampled through the first deconvolution layer to obtain the first decoded features. The first decoded features are concatenated with the middle-level features and upsampled through the second deconvolution layer to obtain the second decoded features. The second decoded features are concatenated with the underlying features and upsampled through the second deconvolution layer to restore the number of points in the original point cloud; finally, the classification score of each point is output through a fully connected layer with shared weights to complete semantic segmentation.

7. The method for constructing a 3D model of a cable manhole based on point cloud data as described in claim 1, characterized in that, Using R -The tree dynamic spatial index structure and KNN nearest neighbor algorithm are used to extract the boundary features of the wall, and the boundary curve is fitted by the extracted boundary feature points to reconstruct the three-dimensional model of the wall; The pipe centerline is extracted using a segmented centerline fitting method. Based on the position of each center point on the centerline, translation and rotation operations are performed to scan and obtain a three-dimensional pipe model. Pipeline support reconstruction includes extracting the support point cloud data corresponding to the semantically segmented pipeline support labels, clustering the pipeline supports in the underground utility tunnel scene, and drawing the support model based on the clustered support point cloud data using the tri-domain surface reconstruction method.

8. A system for constructing a 3D model of a cable manhole based on point cloud data, characterized in that, include: The data acquisition module is used to collect raw point cloud data and auxiliary observation data of underground power lines. The raw point cloud data is the three-dimensional laser scanning data of underground cable manholes and cable tunnels collected on site, and the auxiliary observation data is the location information collected on site for the positioning and orientation of underground cable point clouds. The semantic segmentation module is used to perform multi-class target segmentation of the original point cloud data in three-dimensional tunnel point cloud using the manhole point cloud semantic segmentation model; wherein, the manhole point cloud semantic segmentation model adopts a trained self-attention-based PointNet++ encoder and decoder, which includes: PointNet++ encoder and decoder backbone network, lightweight self-attention module, multi-scale feature pyramid network and graph optimization module. The 3D reconstruction module is used to reconstruct the 3D models of each component based on the results of the multi-target segmentation, and to stitch the 3D models of each component together according to the location information to generate a 3D model of the cable well.

9. An electronic device, characterized in that, The method includes a memory and a processor, as well as computer instructions stored in the memory and running on the processor. When the processor executes the computer instructions, it completes the method for constructing a 3D model of a cable well based on point cloud data as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, Used to store computer instructions, which, when executed by a processor, complete the method for constructing a 3D model of a cable well based on point cloud data as described in any one of claims 1-7.

Images