Automatic extraction method and system of power tower components based on three-dimensional point cloud model

Abstract

A method for automatically extracting power tower components based on a three-dimensional point cloud model. The method includes: obtaining original laser point cloud data and performing data preprocessing; dividing the point cloud data into voxel grids, randomly sampling the point cloud in each voxel grid, and extracting the point cloud features in each voxel grid using a multi-scale feature fusion algorithm; extracting the deep features of the point cloud data in the voxel grid using a 3D convolution model, splicing the deep features with the manual features input by the user to form a comprehensive feature vector, and inputting the comprehensive feature vector into a classification network to classify the voxel grid; clustering the classified voxel grids to form a point cloud area of ​​the tower, performing three-dimensional reconstruction based on the clustering results and the point cloud data, and extracting the three-dimensional model of the tower. The solution of the present invention improves the accuracy and efficiency of extracting tower components in complex environments.

Description

Technical Field

[0001] The invention belongs to the field of remote sensing recognition, and in particular relates to an automatic extraction method for power tower components based on a three-dimensional point cloud model. Background Art

[0002] High-precision, high-resolution drone-mounted lidar technology has been widely used in power inspection operations, including the extraction of pole tower components. It takes less time and provides more reliable data, can better achieve spatial positioning and high-precision measurement, and can make more accurate judgments and characterizations of faults and defects, achieving multi-dimensional three-dimensional inspections.

[0003] Traditional tower automatic extraction methods can be summarized as follows:

[0004] (1) 3D region growing search method: By giving an initial cube, region growing is performed on the basis of the cube to extract the coordinates of the tower. This method relies on local features in point cloud data and identifies and extracts the tower structure through search and growth operations.

[0005] (2) Multi-scale supervised classification: Multi-scale information and supervised learning techniques are used to classify point cloud data to distinguish different parts of the tower. By analyzing point cloud data at different scales, the various parts of the tower can be identified, thereby achieving fine extraction.

[0006] (3) Edge detection algorithm: Use edge detection algorithms (such as the Sobel operator) to extract the outlines of poles and components. By designing operators with specific dimensions and directions, the edges of pole components can be identified in a complex background environment.

[0007] (4) Deep learning convolutional neural network: Use deep learning technology, especially convolutional neural network (CNN), to identify feature areas of visible light images. By training the network to identify different parts of the tower, such as the tower body, tower head, cross arm, etc., the tower parts can be extracted.

[0008] (5) Infrared image temperature information combination: Combining the temperature information of infrared images with the contours extracted by edge detection algorithms. Infrared images can provide additional information, such as temperature distribution, to distinguish tower components from the background.

[0009] (6) Oriented Bounding Box (OBB) fitting: The point cloud data of the tower body is extracted by OBB fitting. First, an initial OBB fitting is performed to identify the tower edges, and then a precise OBB fitting is performed on the tower body with the tower edges identified, thereby realizing the extraction of the point cloud data of the tower body.

[0010] (7) Circle fitting: For the point cloud data of the top part of the tower, the point cloud data of the top part of the tower is horizontally sliced, and then circle fitting is performed on the point cloud data in each slice. Finally, the point cloud data in each slice is merged to obtain the point cloud data of the top part of the tower.

[0011] In general, the tower component extraction method in the prior art has the following defects:

[0012] (1) In terms of transmission line classification and extraction, traditional methods usually rely on manual editing. Key points need to be selected based on the local geometric properties of point cloud data. They are highly dependent on operators and are easily affected by factors such as point cloud data noise, density changes, and angle changes. As a result, the extraction of key points of transmission towers of different types, densities, and angles is not accurate, which in turn affects the accuracy and precision of subsequent work.

[0013] (2) Supervised learning-based methods require a large amount of labeled data for training, which not only increases the workload, but also may be difficult to obtain sufficient training samples for new types of towers or towers in different environments, limiting the scope of applicability of the method.

[0014] (3) In complex terrain environments, such as mountainous areas, the extraction of tower point clouds is easily interfered by high-altitude objects (such as trees and buildings), resulting in a decrease in the accuracy of tower identification. Lack of sufficient terrain adaptability may make it impossible to effectively process and extract tower point clouds for variable terrain conditions, such as undulating mountains and hills.

[0015] (4) In terms of tower positioning accuracy, existing technologies may not be able to meet higher accuracy requirements, especially at the decimeter or even centimeter level, which limits the use of tower point cloud extraction in high-precision application scenarios.

[0016] (5) For large-scale transmission line inspections, existing technologies may be insufficient in processing efficiency and unable to quickly extract tower point clouds from massive point cloud data, thus affecting the efficiency of power inspections.

[0017] (6) After extracting the tower point cloud, it may not be able to provide sufficient auxiliary information for subsequent processing, such as scene classification and line suspension point extraction, which limits the further application of point cloud data in fields such as power inspection. Summary of the invention

[0018] In order to address the deficiencies in the prior art, the present invention provides an automatic extraction method for power tower components based on a three-dimensional point cloud model, so as to solve the problems that the traditional tower component point cloud extraction algorithm is sensitive to ground undulations, and the classification accuracy of the extracted tower component point cloud is low and it is difficult to remove ground points.

[0019] In order to solve the above technical problems, the present invention adopts the following technical solutions.

[0020] The present invention first discloses a method for automatically extracting power tower components based on a three-dimensional point cloud model, the method comprising the following steps:

[0021] Step 1: Obtain the original laser point cloud data and perform data preprocessing;

[0022] Step 2: Divide the point cloud data into voxel grids, randomly sample the point clouds in each voxel grid, and use the multi-scale feature fusion algorithm to extract the point cloud features in each voxel grid;

[0023] Step 3: Use the 3D convolution model to extract deep features of the point cloud data in the voxel grid, concatenate the deep features with the manual features input by the user to form a comprehensive feature vector, and input the comprehensive feature vector into the classification network to classify the voxel grid;

[0024] Step 4: Cluster the classified voxel grids to form a point cloud area of ​​the tower, perform three-dimensional reconstruction based on the clustering results and point cloud data, and extract the three-dimensional model of the tower.

[0025] The present invention further includes the following preferred embodiments:

[0026] The data preprocessing further comprises:

[0027] Step 1.1: Use a diffusion autoencoder with a 4-tensor structure to filter the original laser point cloud data, remove noise information, and retain the tower structure information:

[0028] ,

[0029] ,

[0030] in is the original input, , is the topological coefficient of the point cloud, W is the multi-layer weight of the autoencoder, b and c are weight biases, and lk is the output of the autoencoder after noise removal;

[0031] Step 1.2: Divide the point cloud data space into grid units, and thin out the point cloud in each grid unit according to the preset density standard;

[0032] Step 1.3: Apply the convolution filtering algorithm according to the characteristics of the point cloud data to filter the point cloud data and remove error points and noise.

[0033] The randomly sampling the point cloud in each voxel grid further includes:

[0034] Determine the size of the voxel grid according to the voltage level and the resolution of the point cloud data, use a three-dimensional space partitioning algorithm to distribute the point cloud data to the corresponding voxel grid, and create an index for each voxel grid, wherein the three-dimensional space partitioning algorithm is a three-dimensional grid or an octree algorithm;

[0035] The sampling rate is adaptively determined according to the density and quantity of the point cloud in the voxel grid, and the corresponding speed-up grid is randomly sampled based on the sampling rate.

[0036] The extracting deep features of the point cloud data in the voxel grid by using the 3D convolution model further includes:

[0037] The spatial distribution within each voxel grid is analyzed using a 3D convolutional neural network to capture the spatial arrangement and geometric shape of the point cloud within the voxel grid, and multi-level feature representation is learned by stacking multiple convolutional layers;

[0038] The 3D convolution calculation process of the 3D convolution model is:

[0039] .

[0040] The classification network includes one or more fully connected layers, the last layer adopts a softmax activation function, and a regularization algorithm is used to reduce overfitting.

[0041] The three-dimensional reconstruction based on the clustering result and the point cloud data to extract the three-dimensional model of the tower further includes:

[0042] Clustering the classified voxel grids, removing isolated voxel grids, and forming a continuous tower point cloud area;

[0043] Combining the clustering results with the spatial information of the point cloud data, a 3D reconstruction algorithm is used to extract the 3D model of the tower.

[0044] The reconstructed 3D model is post-processed, and the noise on the model surface is removed using a smoothing algorithm, holes in the model are identified and filled, and the boundaries of the 3D model are corrected.

[0045] The present invention also discloses a system for automatically extracting power tower components based on a three-dimensional point cloud model using the above-mentioned method for automatically extracting power tower components based on a three-dimensional point cloud model, comprising:

[0046] Point cloud data acquisition module, used to acquire original laser point cloud data and perform data preprocessing;

[0047] The voxel grid division module is used to divide the point cloud data into voxel grids, randomly sample the point cloud in each voxel grid, and extract the point cloud features in each voxel grid using a multi-scale feature fusion algorithm;

[0048] A stacked convolution classification module is used to extract deep features of point cloud data in a voxel grid using a 3D convolution model, concatenate the deep features with manual features input by a user to form a comprehensive feature vector, and input the comprehensive feature vector into a classification network to classify the voxel grid;

[0049] The clustering reconstruction module is used to cluster the classified voxel grids to form a point cloud area of ​​the tower, perform three-dimensional reconstruction based on the clustering results and point cloud data, and extract the three-dimensional model of the tower.

[0050] Accordingly, the present application also discloses a terminal, including a processor and a storage medium;

[0051] The storage medium is used to store instructions;

[0052] The processor is used to operate according to the instructions to execute the steps of the aforementioned method for automatically extracting power pole tower components based on a three-dimensional point cloud model.

[0053] Correspondingly, the present application also discloses a computer-readable storage medium on which a computer program is stored. When the program is executed by a processor, the steps of the aforementioned method for automatically extracting power tower components based on a three-dimensional point cloud model are implemented.

[0054] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides an automatic extraction method for power tower components based on a three-dimensional point cloud model, which improves the accuracy and efficiency of tower component extraction, reduces the need for manual intervention, can adapt to the extraction of tower components in complex environments, has strong robustness, and the generated three-dimensional model of tower components is of high accuracy, providing strong support for subsequent design, construction and maintenance work. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a flow chart of the method for automatically extracting power tower components based on a three-dimensional point cloud model in the present invention. DETAILED DESCRIPTION

[0056] In order to make the purpose, technical solution and advantages of the present invention more clear, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0057] The embodiments described in this application are only some embodiments of the present invention, not all embodiments. Based on the spirit of the present invention, other embodiments obtained by ordinary technicians in this field without creative work are all within the protection scope of the present invention.

[0058] In view of the shortcomings of the prior art, the present invention proposes a pole tower component extraction method based on LiDAR laser point cloud data. Through data preprocessing, feature extraction and voxel grid division, deep learning model learning classification and extraction, clustering and three-dimensional reconstruction and other steps, the point cloud of each component of the pole tower is automatically and accurately extracted, reducing manual intervention, and solving the problems of traditional algorithms being sensitive to ground undulations, low classification accuracy, and difficulty in eliminating ground points. In environments with large terrain undulations and interference from tall vegetation, the method still has good robustness, applicability and accuracy.

[0059] See also Figure 1 As shown, the method for automatically extracting power tower components based on a three-dimensional point cloud model disclosed in the present invention comprises the following steps:

[0060] Step 1: Obtain raw laser point cloud data and perform data preprocessing.

[0061] The step 1 further comprises:

[0062] Step 1.1: Perform a preliminary screening of the LiDAR raw laser point cloud data, use a diffusion autoencoder with a 4-tensor structure to remove noise information and extract effective features of the point cloud data. The autoencoder is an unsupervised learning model that can automatically learn an effective representation of the data, thereby helping to identify and remove noise points while retaining key tower structure information.

[0063] In a specific embodiment, firstly, according to the statistical characteristics of the point cloud data, the reflectivity histogram is used to input the point cloud data into the diffusion autoencoder with a 4-tensor structure proposed in the present invention to eliminate noise or ground point cloud:

[0064]

[0065]

[0066] in is the original input, , is the topological coefficient of the point cloud, W is the multi-layer weight of the autoencoder, b and c are weight biases, and lk is the output of the autoencoder after noise removal;

[0067] Step 1.2: Use the spatial grid thinning method to perform regular segmented thinning on the original data, effectively retaining the point cloud density of areas such as conductor towers and insulators, while controlling the point cloud data and density of surrounding objects to further highlight the data quality of key areas.

[0068] In a specific embodiment, the point cloud data space is first divided into regular grid units, and then the point cloud in each grid unit is thinned according to a preset density standard to retain the point cloud of the key structure.

[0069] Step 1.3: Select and apply the convolution filtering algorithm according to the characteristics of the point cloud data, filter the point cloud data, further remove error points and noise, and optimize the data quality.

[0070] The convolution filter calculation process is as follows:

[0071] ,

[0072] in, is a local continuous sample centered at point P (x, y, z); W is the third-order tolerance matrix of the filter, , , is the filtering tolerance in three dimensions of the point cloud;

[0073] Step 2: Divide the point cloud data into voxel grids, randomly sample the point clouds in each voxel grid, and use the multi-scale feature fusion algorithm to extract the point cloud features in each voxel grid.

[0074] The step 2 further comprises:

[0075] Step 2.1: Divide the point cloud data into voxel grids of equal size, each voxel grid represents a small spatial unit, which is convenient for subsequent feature extraction and analysis.

[0076] In a specific embodiment, the size of the voxel grid is determined according to the voltage level, the resolution of the point cloud data and the analysis requirements. The size of the voxel grid should be small enough to capture the fine structure of the tower while ensuring that important details are not lost. A three-dimensional space partitioning algorithm, such as a three-dimensional grid or an octree, is used to allocate the point cloud data to the corresponding voxel grid. An index is established for each voxel grid for fast access and processing.

[0077] Step 2.2: Randomly sample the point cloud in each voxel grid, dynamically adjust the sampling number based on the number of point clouds in the voxel grid, and retain the representative point cloud data in each voxel grid.

[0078] In a specific embodiment, the sampling rate is adaptively determined according to the density and number of point clouds in the voxel grid. For areas with dense point clouds, the sampling rate can be higher; for areas with sparse point clouds, the sampling rate can be lower. A random method is used to select point clouds in the voxel grid for sampling to avoid exploiting systematic biases. It is ensured that the sampled point cloud data can represent the distribution characteristics of the original data.

[0079] Step 2.3: Apply a multi-scale feature fusion algorithm based on voxel grid division to extract richer features by combining voxel grid data of different resolutions. In addition, use a graph neural network (GNN) to process point cloud data to capture the complex relationship between points and provide deeper connection information for subsequent deep learning models.

[0080] Step 3: Use the 3D convolutional model to extract the deep features of the point cloud data in the voxel grid, concatenate the deep features with the manual features input by the user to form a comprehensive feature vector, and input the comprehensive feature vector into the classification network to classify the voxel grid.

[0081] The step 3 further comprises:

[0082] Step 3.1: Use the convolutional layers of a 3D convolutional neural network (CNN) to perform a deep analysis of the spatial distribution within each voxel grid. The 3D convolutional layer can capture the spatial arrangement and geometry of the point cloud within the voxel grid, which contains key information for identifying the tower structure. By inventing a multi-convolutional layer stacking technique, the model is able to learn multi-level feature representations from coarse to fine. The initial layers capture basic shape information such as edges and corners, while the deep layers recognize more complex patterns, such as specific structural parts of the tower.

[0083] The present invention uses different types of convolution kernels as stacking elements, including spherical convolution kernels, to better adapt to the irregularity of point cloud data. Atrous convolution increases the receptive field while keeping the number of parameters unchanged. Batch normalization is used to accelerate the training process and improve the stability of the model. The key formulas of multiple convolution kernels are extracted and integrated to form an independent multi-core stacked convolution algorithm.

[0084] Step 3.2: In addition to the deep features extracted by the convolutional layer, the hand-crafted geometric features such as relative height difference and point density are fused, which are obtained by calculating the spatial distribution and statistical properties of the point cloud within the voxel grid. A feature fusion layer is designed, which concatenates the output of the convolutional layer with the hand-crafted features to form a comprehensive feature vector of the point cloud. The comprehensive feature vector contains rich spatial and geometric information, providing a solid foundation for classification.

[0085] The formula for constructing the comprehensive feature vector of the point cloud is:

[0086] ,

[0087] Among them, x, y, z are the point cloud coordinates, K is the number of voxelized target features, I is the reflection intensity value of the point cloud, k is the number of point clouds of a single voxel, and W, H, D are the angle conversion weights of the point cloud in three dimensions;

[0088] The present invention uses graph convolutional networks (GCN) to capture local connection information in point cloud data. The spatial attention mechanism is used to highlight key geometric features. The adaptive feature fusion algorithm is used to dynamically adjust the fusion weight according to the importance of the feature. The improved graph convolutional network algorithm is:

[0089] ,

[0090] Among them, different neighbor nodes are assigned different weight calculation methods as shown above. is the activation function f(x) = max(0, x), i and j are the two nodes for which link information needs to be calculated, N(i) is the initialization set of point positions, W is the dynamically updated adjacency weight, is the weight calibration factor of the neighboring points, and H is the dynamic graph matrix;

[0091] Step 3.3: Map the comprehensive feature vector to the classification decision space. The core of the classification network is one or more fully connected layers. The last layer of the network uses a softmax activation function, which can not only perform binary classification, but also be extended to multi-classification problems, providing convenience for possible future expansion. In order to improve the generalization ability of the model, the present invention uses regularization algorithms in the network, such as Dropout and L2 regularization, to reduce the risk of overfitting.

[0092] During the network training process, a cross entropy loss function is used to quantify the difference between the probability distribution predicted by the model and the true label. Through the back propagation algorithm, the present invention updates the weights of the network to minimize the loss. The present invention selects the Adam optimizer, which combines the advantages of momentum and RMSProp, automatically adjusts the learning rate during the training process, and accelerates the convergence speed.

[0093] The present invention adopts weighted cross entropy loss function to deal with the problem of class imbalance and uses focal loss (FocalLoss) to increase the model's learning of difficult-to-classify samples.

[0094] ,

[0095] y is the true label value (positive class value is 1, negative class value is 0), and y is the predicted probability value ( ∈(0,1)). It represents the difference between the true sample label and the predicted probability.

[0096] In a further preferred embodiment, the AdamW optimizer is used, and weight decay is added on the basis of the Adam optimizer to further improve the generalization ability of the model.

[0097] ,

[0098] ,

[0099] in, yes The first moment of time, yes The second moment of time, is the time gradient, yes The weight parameter at the moment, is the learning rate, , is the weight decay rate, is the attenuation coefficient, is a constant to avoid zero denominator.

[0100] In the face of class imbalance, undersampling and oversampling algorithms are used. Undersampling randomly selects representative samples from the majority class to reduce its influence in training. Oversampling is used to increase the number of minority class samples to improve the model's recognition ability for these classes. The present invention uses the synthetic minority oversampling algorithm (SMOTE) to increase the number of minority class samples while maintaining data diversity. Focal Loss is used to reduce the focus on easy-to-classify samples and increase the learning of difficult-to-classify samples.

[0101] Use class-balanced data augmentation algorithms, such as Mixup and CutMix, to improve the model's ability to identify minority class samples. Use the Online Hard Example Mining strategy to dynamically select difficult-to-classify samples for training to improve the robustness of the model.

[0102] Step 4: Cluster the classified voxel grids to form a point cloud area of ​​the tower, perform three-dimensional reconstruction based on the clustering results and point cloud data, and extract the three-dimensional model of the tower.

[0103] The step 4 further comprises:

[0104] Step 4.1: Apply the clustering algorithm K-means to cluster the classified voxels, remove isolated voxels, and form a continuous tower point cloud area.

[0105] Step 4.2: Combine the clustering results with the spatial information of the point cloud data and use a 3D reconstruction algorithm, such as multi-view stereo matching or surface fitting, to finely extract the 3D model of the tower.

[0106] Step 4.3: Post-process the reconstructed 3D model, including using various smoothing algorithms to remove noise on the model surface, identify and fill holes in the model to ensure the closure of the model, correct the boundaries of the 3D model, and improve the geometric accuracy and integrity of the model.

[0107] The beneficial effect of the present invention is that, compared with the prior art, the present invention provides an automatic extraction method for power tower components based on a three-dimensional point cloud model, which improves the accuracy and efficiency of tower component extraction, reduces the need for manual intervention, can adapt to the extraction of tower components in complex environments, has strong robustness, and the generated three-dimensional model of tower components is of high accuracy, providing strong support for subsequent design, construction and maintenance work.

[0108] The present invention may be a system, a method and / or a computer program product. The present invention also discloses a system for automatically extracting power pole tower components based on a three-dimensional point cloud model based on the aforementioned method for automatically extracting power pole tower components based on a three-dimensional point cloud model, comprising:

[0109] Point cloud data acquisition module, used to acquire original laser point cloud data and perform data preprocessing;

[0110] The voxel grid division module is used to divide the point cloud data into voxel grids, randomly sample the point cloud in each voxel grid, and extract the point cloud features in each voxel grid using a multi-scale feature fusion algorithm;

[0111] A stacked convolution classification module is used to extract deep features of point cloud data in a voxel grid using a 3D convolution model, concatenate the deep features with manual features input by a user to form a comprehensive feature vector, and input the comprehensive feature vector into a classification network to classify the voxel grid;

[0112] The clustering reconstruction module is used to cluster the classified voxel grids to form a point cloud area of ​​the tower, perform three-dimensional reconstruction based on the clustering results and point cloud data, and extract the three-dimensional model of the tower.

[0113] Based on the spirit of the present invention, those skilled in the art can easily think that a computer program product can be obtained based on the aforementioned method for automatically extracting power pole tower components based on a three-dimensional point cloud model. The computer program product may include a computer-readable storage medium, which carries computer-readable program instructions for enabling a processor to implement various aspects of the present disclosure. That is, the present application also includes a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the steps of the aforementioned method for automatically extracting power pole tower components based on a three-dimensional point cloud model.

[0114] Computer readable storage medium can be a tangible device that can keep and store the instructions used by the instruction execution device. Computer readable storage medium can be, for example, - but not limited to - electrical storage device, magnetic storage device, optical storage device, electromagnetic storage device, semiconductor storage device or any suitable combination of the above. More specific examples (non-exhaustive list) of computer readable storage medium include: portable computer disk, hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disk read-only memory (CD-ROM), digital versatile disk (DVD), memory stick, floppy disk, mechanical encoding device, such as punch card or groove protrusion structure with instructions stored thereon and any suitable combination of the above. Computer readable storage medium used here is not interpreted as instantaneous signal itself, such as radio wave or other free propagating electromagnetic wave, electromagnetic wave propagated by waveguide or other transmission medium (for example, light pulse by optical fiber cable) or electrical signal transmitted by wire.

[0115] The computer-readable program instructions described herein can be downloaded from a computer-readable storage medium to each computing / processing device, or downloaded to an external computer or external storage device via a network, such as the Internet, a local area network, a wide area network, and / or a wireless network. The network can include copper transmission cables, optical fiber transmissions, wireless transmissions, routers, firewalls, switches, gateway computers, and / or edge servers. The network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards the computer-readable program instructions for storage in the computer-readable storage medium in each computing / processing device.

[0116] The computer program instructions for performing the operations of the present disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as "C" language or similar programming languages. The computer-readable program instructions may be executed entirely on the user's computer, partially on the user's computer, as a separate software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the case of a remote computer, the remote computer may be connected to the user's computer through any type of network including a local area network (LAN) or a wide area network (WAN), or may be connected to an external computer (e.g., using an Internet service provider to connect through the Internet). In some embodiments, by using the state information of the computer-readable program instructions to personalize an electronic circuit, such as a programmable logic circuit, a field programmable gate array (FPGA) or a programmable logic array (PLA), the electronic circuit may execute the computer-readable program instructions, thereby implementing various aspects of the present disclosure.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, ordinary technicians in the relevant field should understand that the specific implementation methods of the present invention can still be modified or replaced by equivalents, and any modifications or equivalent replacements that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for automatically extracting power tower components based on a three-dimensional point cloud model. It is characterized in that The following steps are involved: Step 1: Obtain the original laser point cloud data and perform data preprocessing; Step 2: Divide the point cloud data into voxel grids, randomly sample the point clouds in each voxel grid, and use the multi-scale feature fusion algorithm to extract the point cloud features in each voxel grid; Step 3: Use the 3D convolution model to extract deep features of the point cloud data in the voxel grid, concatenate the deep features with the manual features input by the user to form a comprehensive feature vector, and input the comprehensive feature vector into the classification network to classify the voxel grid; Step 4: Cluster the classified voxel grids to form a point cloud area of ​​the tower, perform 3D reconstruction based on the clustering results and point cloud data, and extract the 3D model of the tower; The three-dimensional reconstruction based on the clustering result and the point cloud data to extract the three-dimensional model of the tower further includes: Clustering the classified voxel grids, removing isolated voxel grids, and forming a continuous tower point cloud area; Combining the clustering results with the spatial information of the point cloud data, a 3D reconstruction algorithm is used to extract the 3D model of the tower. The reconstructed 3D model is post-processed, and the noise on the model surface is removed using a smoothing algorithm, holes in the model are identified and filled, and the boundaries of the 3D model are corrected.

2. The method for automatically extracting power tower components based on a three-dimensional point cloud model according to claim 1, It is characterized in that The data preprocessing further comprises: Step 1.1: Using a diffusion autoencoder with a 4-tensor structure to filter the original laser point cloud data, remove noise information, and retain the tower structure information; in is the original input, , is the topological coefficient of the point cloud, W is the multi-layer weight of the autoencoder, b and c are weight biases, and lk is the output of the autoencoder after noise removal; Step 1.2: Divide the point cloud data space into grid units, and thin out the point cloud in each grid unit according to the preset density standard; Step 1.3: Apply the convolution filtering algorithm according to the characteristics of the point cloud data to filter the point cloud data and remove error points and noise.

3. The automatic extraction method of power tower components based on a three-dimensional point cloud model according to claim 2, It is characterized in that The randomly sampling the point cloud in each voxel grid further includes: Determine the size of the voxel grid according to the voltage level and the resolution of the point cloud data, use a three-dimensional space partitioning algorithm to distribute the point cloud data to the corresponding voxel grid, and create an index for each voxel grid, wherein the three-dimensional space partitioning algorithm is a three-dimensional grid or an octree algorithm; The sampling rate is adaptively determined according to the density and quantity of the point cloud in the voxel grid, and the corresponding speed-up grid is randomly sampled based on the sampling rate.

4. The method for automatically extracting power tower components based on a three-dimensional point cloud model according to claim 3, It is characterized in that The classification network includes one or more fully connected layers, the last layer adopts a softmax activation function, and a regularization algorithm is used to reduce overfitting.

5. An automatic extraction system for power tower components based on 3D point cloud model, It is characterized in that include: Point cloud data acquisition module, used to acquire original laser point cloud data and perform data preprocessing; The voxel grid division module is used to divide the point cloud data into voxel grids, randomly sample the point cloud in each voxel grid, and extract the point cloud features in each voxel grid using a multi-scale feature fusion algorithm; A stacked convolution classification module is used to extract deep features of point cloud data in a voxel grid using a 3D convolution model, concatenate the deep features with manual features input by a user to form a comprehensive feature vector, and input the comprehensive feature vector into a classification network to classify the voxel grid; The clustering reconstruction module is used to cluster the classified voxel grids to form a point cloud area of ​​the tower, perform three-dimensional reconstruction based on the clustering results and point cloud data, and extract the three-dimensional model of the tower; The cluster reconstruction module is further used for: Clustering the classified voxel grids, removing isolated voxel grids, and forming a continuous tower point cloud area; Combining the clustering results with the spatial information of the point cloud data, a 3D reconstruction algorithm is used to extract the 3D model of the tower. The reconstructed 3D model is post-processed, and the noise on the model surface is removed using a smoothing algorithm, holes in the model are identified and filled, and the boundaries of the 3D model are corrected.

6. The automatic extraction system for power tower components based on three-dimensional point cloud model according to claim 5, It is characterized in that The point cloud data acquisition module is further used for: The original laser point cloud data is screened using a diffusion autoencoder with a 4-tensor structure to remove noise information and retain the tower structure information: in is the original input, , is the topological coefficient of the point cloud, W is the multi-layer weight of the autoencoder, b and c are weight biases, and lk is the output of the autoencoder after noise removal; The point cloud data space is divided into grid units, and the point cloud in each grid unit is thinned according to the preset density standard; According to the characteristics of point cloud data, the convolution filtering algorithm is applied to filter the point cloud data to remove error points and noise.

7. The automatic extraction system for power tower components based on three-dimensional point cloud model according to claim 6, It is characterized in that The voxel grid partitioning module is further used for: Determine the size of the voxel grid according to the voltage level and the resolution of the point cloud data, use a three-dimensional space partitioning algorithm to distribute the point cloud data to the corresponding voxel grid, and create an index for each voxel grid, wherein the three-dimensional space partitioning algorithm is a three-dimensional grid or an octree algorithm; The sampling rate is adaptively determined according to the density and quantity of the point cloud in the voxel grid, and the corresponding speed-up grid is randomly sampled based on the sampling rate.

8. The automatic extraction system for power tower components based on three-dimensional point cloud model according to claim 7, It is characterized in that The classification network includes one or more fully connected layers, the last layer adopts a softmax activation function, and a regularization algorithm is used to reduce overfitting.

9. A computer-readable storage medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, the steps of the method for automatically extracting power tower components based on a three-dimensional point cloud model as described in any one of claims 1 to 4 are implemented.

Images