Elevation perception multi-resolution network processing method for power transmission corridor point cloud classification

Abstract

The embodiment of the invention discloses an elevation perception multi-resolution network processing method for power transmission corridor point cloud classification, and the method comprises the steps: carrying out the preprocessing of LiDAR point cloud data of a power transmission corridor, and obtaining the preprocessed point cloud data; inputting the preprocessed point cloud data into a feature extraction module, and mapping point cloud coordinates to a high-dimensional feature space through a feature embedding layer to obtain initial features; respectively inputting the initial features into an elevation embedding module and a multi-resolution module for feature enhancement and feature extraction to obtain elevation enhancement features and multi-resolution features; fusing the elevation enhancement feature and the multi-resolution feature to obtain a fused feature; performing a classification task and / or a segmentation task based on the fused features, and outputting a classification result and / or a segmentation result; and the point cloud classification precision and the complex scene robustness can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of point cloud classification, and in particular to an elevation-aware multi-resolution network processing method for point cloud classification in power transmission corridors. Background Art

[0002] As an important form of spatial data, 3D point cloud data is widely used in fields such as automation, autonomous driving, augmented reality (AR), geographic information systems (GIS), and power inspection. Classifying 3D point clouds is a key task in 3D point cloud processing. Point cloud classification can identify different types of objects or features within point cloud data. For example, in a transmission corridor scenario, it can distinguish between power lines, power towers, and vegetation, which is crucial for the maintenance and management of power facilities.

[0003] However, current methods have many limitations when processing transmission corridor point cloud data. Multi-view projection methods destroy the three-dimensional topological structure during the projection process, resulting in the loss of continuity features of power lines. Voxelization methods have difficulty effectively handling the complex terrain of transmission corridors due to their high computational complexity and discretization noise issues for slender targets (such as power lines). While direct point cloud processing methods avoid the shortcomings of projection and voxelization, they fail to fully utilize the vertical distribution characteristics of power equipment and are prone to over-smoothing effects in sparse areas (such as power lines). In addition, current methods lack classification accuracy and robustness when dealing with scenes with uneven sample distribution (low proportion of key target point clouds such as power lines) and dynamic density changes (from sparse to dense).

[0004] Therefore, a new method is urgently needed to address the shortcomings of three-dimensional point cloud data processing in transmission corridor scenarios and improve the accuracy and robustness of point cloud classification. Summary of the Invention

[0005] The main purpose of the present invention is to provide an elevation-aware multi-resolution network processing method for transmission corridor point cloud classification, which selectively enhances the input data and improves the accuracy of point cloud classification.

[0006] To achieve the above objectives, the present application provides, in a first aspect, an elevation-aware multi-resolution network processing method for transmission corridor point cloud classification, the method comprising:

[0007] Preprocess the LiDAR point cloud data of the transmission corridor to obtain preprocessed point cloud data;

[0008] Input the preprocessed point cloud data into the feature extraction module, and map the point cloud coordinates to the high-dimensional feature space through the feature embedding layer to obtain initial features;

[0009] Inputting the initial features into the elevation embedding module and the multi-resolution module for feature enhancement and feature extraction respectively, to obtain elevation enhancement features and multi-resolution features;

[0010] Fusing the elevation enhancement feature and the multi-resolution feature to obtain a fused feature;

[0011] A classification task and / or a segmentation task is performed based on the fused features, and a classification result and / or a segmentation result is output.

[0012] A second aspect of the present application provides an elevation-aware multi-resolution network processing system for transmission corridor point cloud classification, comprising:

[0013] A data preprocessing unit, used to preprocess the LiDAR point cloud data of the transmission corridor to obtain preprocessed point cloud data;

[0014] A feature extraction module is used to input the preprocessed point cloud data into a feature embedding layer, map the point cloud coordinates to a high-dimensional feature space, and obtain initial features;

[0015] An elevation embedding module, configured to perform feature enhancement on the initial features to obtain elevation enhanced features;

[0016] A multi-resolution module, used to extract features from the initial features to obtain multi-resolution features;

[0017] A feature fusion unit, configured to fuse the elevation enhancement feature and the multi-resolution feature to obtain a fused feature;

[0018] A classification unit is used to perform a classification task and / or a segmentation task based on the fused features and output a classification result and / or a segmentation result.

[0019] A third aspect of the present application provides an electronic device, comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the first aspect and any possible implementation thereof.

[0020] To achieve the above-mentioned objectives, the fourth aspect of the present application provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the various steps in the method described in the first aspect.

[0021] The present application provides an elevation-aware multi-resolution network processing method for transmission corridor point cloud classification, which preprocesses the LiDAR point cloud data of the transmission corridor to obtain preprocessed point cloud data; inputs the preprocessed point cloud data into a feature extraction module, and maps the point cloud coordinates to a high-dimensional feature space through a feature embedding layer to obtain initial features; inputs the initial features into an elevation embedding module and a multi-resolution module for feature enhancement and feature extraction, respectively, to obtain elevation-enhanced features and multi-resolution features; fuses the elevation-enhanced features and the multi-resolution features to obtain fused features; performs classification tasks and / or segmentation tasks based on the fused features, and outputs classification results and / or segmentation results; this method can improve the point cloud classification accuracy and the robustness of complex scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0023] in:

[0024] Figure 1 A flowchart of an elevation-aware multi-resolution network processing method for transmission corridor point cloud classification provided by an embodiment of the present application;

[0025] Figure 2 A schematic diagram of a network structure provided in an embodiment of the present application;

[0026] Figure 3 A schematic structural diagram of an elevation embedding module provided in an embodiment of the present application;

[0027] Figure 4 A schematic diagram of the structure of a multi-resolution module provided in an embodiment of the present application;

[0028] Figure 5 A schematic diagram of the structure of an elevation-aware multi-resolution network processing system for transmission corridor point cloud classification provided in an embodiment of the present application;

[0029] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the present invention, the following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0031] The terms "first," "second," and the like in the specification and claims of this application and the accompanying drawings are used to distinguish between different objects, not to describe a particular order. Furthermore, the terms "including," "having," and any variations thereof, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or apparatus comprising a series of steps or elements is not limited to the listed steps or elements but may optionally include steps or elements not listed, or may optionally include other steps or elements inherent to the process, method, product, or apparatus.

[0032] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0033] The laser radar (Light Detection and Ranging, LiDAR) involved in the embodiments of this application is an active remote sensing technology that obtains the three-dimensional spatial coordinates (XYZ) and reflection intensity information of the target by emitting laser pulses and receiving reflected signals, thereby generating high-precision point cloud data.

[0034] The elevation embedding module (EE module) involved in the embodiments of the present application explicitly models the vertical spatial distribution characteristics of power equipment by calculating local elevation differences and fusing original coordinates.

[0035] The multi-resolution module (EE module) involved in the embodiments of the present application adopts dual-branch progressive downsampling and cross-resolution feature fusion to adapt to the density heterogeneity of the transmission corridor point cloud.

[0036] The embodiments of the present application are described below in conjunction with the drawings in the embodiments of the present application.

[0037] See also Figure 1, which is a flow chart of an elevation-aware multi-resolution network processing method for transmission corridor point cloud classification provided by an embodiment of the present application, such as Figure 1 As shown, the method includes:

[0038] 101. Preprocess the LiDAR point cloud data of the transmission corridor to obtain preprocessed point cloud data.

[0039] The execution subject of the method in the embodiment of the present application can be an elevation-aware multi-resolution network processing system for transmission corridor point cloud classification, which can be implemented through a terminal device, such as a computer, in practical applications.

[0040] In the embodiment of the present application, point cloud data can be obtained by LiDAR scanning. Specifically, the three-dimensional point cloud of the transmission corridor generated by LiDAR scanning can be expressed as Each point contains spatial coordinates (x,y,z).

[0041] The LiDAR point cloud data of the transmission corridor can be preprocessed, including filtering out outliers and noise points, and normalizing the point cloud data. The purpose of preprocessing is to improve data quality and reduce errors in subsequent processing.

[0042] 102. Input the preprocessed point cloud data into the feature extraction module, and map the point cloud coordinates to the high-dimensional feature space through the feature embedding layer to obtain the initial features.

[0043] The preprocessed point cloud data is input into the feature extraction module, which first maps the point cloud coordinates to a high-dimensional feature space through a feature embedding layer to obtain initial features. This step converts the low-dimensional coordinates into semantic features that can be parsed by the neural network.

[0044] Specifically, the original coordinates can be mapped to a high-dimensional feature space through MLP (Multi-layer Perceptron), and the low-dimensional coordinates can be converted into semantic features that can be parsed by the neural network:

[0045]

[0046] In an optional embodiment, the feature extraction module includes an attention feature extractor for performing attention feature extraction on the initial features to enhance the correlation of local features.

[0047] Specifically, in the embodiment of the present application, self-canceling attention can be used to enhance the relevance of local features:

[0048]

[0049] in The query, key, and value matrices after linear transformation are used to model the local geometric relationships of the point cloud (such as the connection structure between power lines and tower materials).

[0050] 103. The above initial features are respectively input into the elevation embedding module and the multi-resolution module for feature enhancement and feature extraction to obtain elevation enhancement features and multi-resolution features.

[0051] The extracted features can be input into the elevation embedding module for processing to obtain elevation enhancement features. This mainly includes calculating the local elevation difference and splicing it with the original coordinates, and then performing feature expansion and compression through the LBR layer (Linear BatchNorm ReLU) to generate elevation enhancement features.

[0052] In an optional embodiment, the elevation embedding module includes:

[0053] The Z-Sampling layer is used to search for the nearest neighbor of each point, calculate the local elevation difference, and concatenate the local elevation difference with the original coordinates to obtain the concatenated features;

[0054] At least one feature processing layer (such as an LBR layer) is used to process the above-mentioned splicing features and output the above-mentioned elevation enhancement features.

[0055] Based on the above description, the following steps are mainly performed in the elevation embedding module:

[0056] a. Calculation of local elevation difference:

[0057] For each point p i , search its k=16 nearest neighbors and calculate:

[0058] Δz i =z i -min(z NN(k) )

[0059] b. Feature splicing and expansion:

[0060] Concatenate the original coordinates with the elevation difference and input them into the LBR layer (Linear→BatchNorm→ReLU):

[0061] F EE =LBR2([LBR1([x i ,y i ,z i ,Δz i ]),[x i ,y i ,z i ]])

[0062] c. Output: Generate 32-dimensional elevation enhancement features

[0063] The processing of the multi-resolution module can be parallel to the processing of the elevation embedding module.

[0064] In the multi-resolution module, the initial features can be downsampled through a dual-branch downsampling structure (including the main sampling path and the offset sampling path) to obtain features of different resolutions. Then, the feature fusion structure upsamples the features of different resolutions to the original number of points and performs cross-resolution splicing and fusion to obtain multi-resolution features.

[0065] In an optional embodiment, the multi-resolution module includes:

[0066] A dual-branch downsampling structure, including a main sampling path and an offset sampling path, is used to downsample the initial features to obtain features of different resolutions;

[0067] The feature fusion structure is used to upsample the features of different resolutions to the original number of points, and perform cross-resolution splicing and fusion to obtain the multi-resolution features.

[0068] For example, two-branch downsampling includes:

[0069] Branch 1 (main sampling):

[0070] The number of points is gradually reduced through uniform sampling (N→N / 2→N / 4→N / 8), and Conv1d is used at each level to expand the feature dimension (32→64→128→256).

[0071] Branch 2 (offset sampling): Sampling starts from the second point and performs the same downsampling operation to ensure spatial coverage integrity.

[0072] Feature fusion:

[0073] Upsample the features of each branch to the original number of points N;

[0074] Perform cross-resolution stitching:

[0075]

[0076] Feature fusion in the multi-resolution module refers to the process of fusing features obtained by sampling at different resolutions within the module. The purpose of this process is to capture and integrate feature information at different scales to adapt to the density heterogeneity of point cloud data.

[0077] 104. Fuse the elevation enhancement feature and the multi-resolution feature to obtain a fused feature.

[0078] In the later stages of the network, the elevation enhanced features extracted by the elevation embedding module and the multi-resolution features extracted by the multi-resolution module can be fused. The purpose of this process is to combine vertical spatial features and multi-scale spatial features to improve the performance of classification and segmentation tasks.

[0079] 105. Perform a classification task and / or a segmentation task based on the above fused features, and output a classification result and / or a segmentation result.

[0080] In an optional embodiment, performing a classification task based on the fused features and outputting a classification result includes:

[0081] Perform a global maximum pooling operation on the above fused features to obtain global features, and output the scene type label through the fully connected layer.

[0082] In an optional embodiment, performing a segmentation task based on the fused features and outputting a segmentation result includes:

[0083] The above fused features are spliced ​​point by point and then spliced ​​with the global features to obtain enhanced features;

[0084] Based on the enhanced features above, the category probability of each point is output through point-by-point classification operation.

[0085] Specifically, for example, feature fusion and classification result output can be divided into the following two categories:

[0086] 1. Split tasks:

[0087] Splicing EE and MR features:

[0088]

[0089] The class probability is predicted point by point through 3-layer LBR (160→128→64→32).

[0090] 2. Classification task:

[0091] Global max pooling:

[0092]

[0093] The fully connected layer (128→64→32→6) outputs the scene type label.

[0094] The technical solutions of the embodiments of this application focus on:

[0095] 1. Elevation feature enhancement:

[0096] By explicitly modeling local elevation differences (such as the elevation difference between the wire suspension height and the surrounding vegetation) through the EE module, the problem of insufficient utilization of vertical spatial features in existing methods is solved, and the classification accuracy of power lines and power towers is significantly improved.

[0097] 2. Multi-scale feature fusion:

[0098] The dual-branch multi-resolution sampling strategy of the MR module effectively captures the detailed structure of dense areas (tower materials) and the long-range correlation of sparse areas (power lines), solving the classification problem of large differences in point cloud density in transmission corridors.

[0099] In order to more clearly illustrate the method in the embodiment of the present application, the network structure in the embodiment of the present application is introduced below.

[0100] Figure 2 A schematic diagram of a network structure provided in an embodiment of the present application. Figure 2 As shown in Figure 2, an elevation-aware multi-resolution network architecture for LiDAR point cloud classification in transmission corridors is presented, as follows:

[0101] (1) Feature Embedding Layer (Input Embedding):

[0102] The input point cloud data P∈RN×3 first passes through the input embedding layer to map the original XYZ coordinates into a 128-dimensional feature space.

[0103] (2) Attention Feature Extractor (Attention):

[0104] A self-canceling attention mechanism is used to enhance the correlation of local features, which is crucial for modeling local geometric relationships in point clouds (such as the connection structure between power lines and towers).

[0105] (3) Dual modules in parallel

[0106] Elevation Embedding: Enhances vertical spatial features.

[0107] Multiple Resolutions: Extracts local features across scales.

[0108] (4) Multi-task output head:

[0109] Segmentation task: output point-by-point category probabilities

[0110] Classification task: output scene type label

[0111] This network architecture effectively addresses the challenges in transmission corridor point cloud classification by combining elevation perception and multi-resolution feature extraction, improving classification accuracy and model robustness.

[0112] Furthermore, the above-mentioned elevation embedding module is explained.

[0113] Figure 3 This is a structural diagram of an elevation embedding module provided in an embodiment of the present application. Figure 3 As shown, the elevation embedding module includes:

[0114] Input layer, Z-Sampling layer, feature concatenation layer (Concat), LBR layer, output layer;

[0115] exist Figure 3 In

[15] , B×N×3 represents the dimension of the input point cloud, B×N×2n represents the dimension after processing by the Z-Sampling layer, and B×N×512 and B×N×256 represent the dimensions after processing by the two LBR layers, respectively. The final output elevation enhancement feature F is used in the subsequent multi-task output head, including segmentation and classification tasks.

[0116] This structure enhances the model's perception of the vertical spatial distribution of power equipment by explicitly modeling local elevation differences, thereby improving the accuracy and robustness of classification.

[0117] Based on the above structure, the cascade structure of Z-Sampling and LBR layer is further explained:

[0118] (1) Z-Sampling layer operation:

[0119] Search for n nearest neighbors for each point;

[0120] Calculate the local elevation difference: Δz_i = z_i-min(z_NN);

[0121] Feature splicing: F_in=[x_i,y_i,z_i,Δz_i];

[0122] (2) Feature processing flow:

[0123] First LBR layer: linear layer expands feature dimensions;

[0124] Feature splicing: Splice the extended features with the original coordinates twice;

[0125] Second LBR layer: The linear layer compresses and outputs the final elevation features.

[0126] The elevation embedding module explicitly models the vertical spatial distribution characteristics of power equipment by calculating local elevation differences and fusing them with the original coordinates. This enhances the model's utilization of vertical spatial features and significantly improves the classification accuracy of power lines and towers. This design enables the network to better adapt to the characteristics of transmission corridor point cloud data, improving classification performance.

[0127] Figure 4 This is a schematic diagram of the structure of a multi-resolution module provided in an embodiment of the present application. Figure 4 As shown, the multi-resolution module mainly includes:

[0128] Dual-branch downsampling structure:

[0129] Branch 1 (main sampling path):

[0130] (1) Uniform downsampling to N / 2 points

[0131] (2) Feature expansion through convolutional layers

[0132] (3) Repeat downsampling to N / 4 and N / 8 points

[0133] Branch 2 (offset sampling path):

[0134] (1) Start sampling from the second point to ensure spatial coverage integrity

[0135] (2) Perform the same three-level downsampling as branch 1

[0136] Feature fusion:

[0137] (1) Upsample the features of each layer to the original number of points N

[0138] (2) Cross-resolution splicing followed by compression through convolutional layers

[0139] Specifically, the input point cloud data P is processed by batch normalization and convolution (BNC), and the output shape is B×N×C, where C is the number of feature channels.

[0140] Branch 1 (main sampling path): Starting from the input layer, it passes through a series of convolutional layers. After each convolution, the number of feature channels doubles and the number of points is halved (B×N / 2×2C, B×N / 4×4C, B×N / 8×8C).

[0141] Branch 2 (offset sampling path): starts sampling from the second point of the input layer and performs the same downsampling operation as branch 1 to ensure the integrity of spatial coverage.

[0142] The feature maps of the two branches at each downsampling level are concatenated (Concat), and then feature fusion is performed through the convolution layer, and the output shape is B×N / 8×2C.

[0143] The fused feature map is restored to the original number of points N through upsampling operation, and the number of feature channels is halved (B×N×2C).

[0144] Finally, the upsampled feature map is processed again by the convolutional layer to output the final multi-resolution feature P′ with a shape of B×N×C.

[0145] Based on the description of the aforementioned method embodiment, the present application also provides an elevation-aware multi-resolution network processing system for transmission corridor point cloud classification.

[0146] Figure 5 A schematic diagram of the structure of an elevation-aware multi-resolution network processing system for transmission corridor point cloud classification provided in an embodiment of the present application is shown in FIG. Figure 5 As shown, the system 500 includes:

[0147] A data preprocessing unit 510 is used to preprocess the LiDAR point cloud data of the transmission corridor to obtain preprocessed point cloud data;

[0148] The feature extraction module 520 is used to input the pre-processed point cloud data into the feature embedding layer, map the point cloud coordinates into a high-dimensional feature space, and obtain initial features;

[0149] The elevation embedding module 530 is used to enhance the initial features to obtain elevation enhanced features;

[0150] A multi-resolution module 540 is used to extract features from the initial features to obtain multi-resolution features;

[0151] A feature fusion unit 550 is used to fuse the elevation enhancement feature and the multi-resolution feature to obtain a fused feature;

[0152] The classification unit 560 is configured to perform a classification task and / or a segmentation task based on the fused features, and output a classification result and / or a segmentation result.

[0153] Understandably, Figure 5 The relevant contents of each module in the above method embodiment have been described in detail, and the details can be referred to the contents of the method embodiment; Figure 5 The risk factor identification device 500 provided for disaster inspection image data can be performed as follows: Figure 1 Any steps in the illustrated embodiment will not be described in detail here.

[0154] Current point cloud classification methods can be divided into three major technical routes. Their application limitations in transmission corridor scenarios are as follows:

[0155] 1. Multi-view projection method

[0156] Representative algorithms: MVCNN (Su et al., 2015), SnapNet (Boulch et al., 2018), MV3D (Chen et al., 2017)

[0157] This type of algorithm may have the following problems:

[0158] The projection process destroys the three-dimensional topological structure and loses the continuity features of the power lines; the multi-view feature fusion uses a simple maximum pooling (MVCNN), which cannot model the spatial symmetry features of the power tower; the back projection produces spatial dislocation and positioning errors of the conductor suspension points.

[0159] 2. Voxelization method

[0160] Representative algorithms: VoxNet (Maturana et al., 2015), O-CNN (Wang et al., 2017), PointGrid (Le et al., 2018)

[0161] This type of algorithm may have the following problems: the voxel resolution is cubically proportional to the computational effort, and the GPU memory usage is large; slender targets such as power lines generate discretization noise after voxelization (the signal-to-noise ratio decreases); and the indexing efficiency decreases in complex terrain.

[0162] 3. Point cloud direct processing method

[0163] Feature extraction type: PointNet++ (layered sampling), DGCNN (dynamic graph convolution), PointConv (density normalized convolution)

[0164] Attention mechanism type: PCT (local Transformer), Point-BERT (pre-trained global encoding)

[0165] This type of algorithm may have the following problems: it has not been specifically improved for the vertical distribution characteristics of power equipment; existing local aggregation methods (such as EdgeConv) produce excessive smoothing effects in sparse power line areas; and the pre-trained model (Point-BERT) has a domain adaptation gap in the transmission corridor scenario.

[0166] In general, existing methods have the following problems:

[0167] Insufficient geometric feature preservation: Existing methods lose elevation information during the projection / voxelization process.

[0168] Poor dynamic density adaptability: the point cloud density of the transmission corridor varies from 0.1 to 200 points / m 2 , the existing sampling strategy cannot take both into account.

[0169] Existing methods typically use the original Z coordinate directly as input features, without explicitly modeling the vertical distribution characteristics of power equipment (such as differences in conductor suspension heights and regularity in tower structure heights). The method in the embodiments of this application proposes an elevation embedding module (EE module). By calculating local elevation differences (Δz_i = z_i - minimum elevation of adjacent points) and dynamically weighted fusion, the model's ability to perceive the vertical spatial distribution of power equipment is enhanced, solving the classification error problem caused by insufficient utilization of elevation information in existing methods.

[0170] Traditional methods (such as PointNet++'s single density sampling and DGCNN's fixed radius neighborhood) are difficult to adapt to the extreme density differences of transmission corridor point clouds (tower material dense areas > 100 points / m2). 2 vs conductor sparse area <1 point / m 2 This application designs a multi-resolution module (MR module) that adopts a dual-branch progressive downsampling strategy: the main branch retains details in dense areas through uniform sampling, and the offset branch ensures the coverage integrity of sparse areas. Finally, adaptive extraction of multi-scale spatial features is achieved through cross-resolution feature fusion.

[0171] The elevation-aware multi-resolution network processing method for transmission corridor point cloud classification in the embodiment of the present application has the following technical effects:

[0172] 1. Improved classification accuracy:

[0173] Experiments show that the F1-score of this application is significantly better than existing methods (such as PointNet++ and DGCNN) in the power line and power tower classification task.

[0174] 2. Robustness in complex scenarios:

[0175] Through multi-resolution feature fusion, the classification consistency of the model in dense and sparse mixed areas (such as the junction of tower materials and wires) is improved.

[0176] 3. Engineering practicality:

[0177] The network design takes into account feature extraction efficiency and supports real-time processing of point cloud data of long-distance transmission corridors.

[0178] In one embodiment of the present application, an electronic device is also provided. Figure 6 , Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application. Figure 6 As shown, the electronic device 600 includes a processor 601 and a memory 602. The memory 602 stores a computer program. When the computer program is executed by the processor 601, the following operations are performed: Figure 1Any step in the method embodiment shown; the electronic device 600 can be used for model training and / or application of a data augmentation network and a point cloud classification network. The electronic device 600 may also include input / output devices, etc. In a specific embodiment, the electronic device may be a terminal device, etc.

[0179] In one embodiment, a computer-readable storage medium is further provided. The computer-readable storage medium stores a computer program. When the computer program is executed by the processor 601, the processor 601 executes any step in the above method embodiment.

[0180] Those skilled in the art will appreciate that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program. The program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0181] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0182] The above-described embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. An elevation-aware multi-resolution network processing method for transmission corridor point cloud classification, characterized in that: The method comprises: Preprocess the LiDAR point cloud data of the transmission corridor to obtain preprocessed point cloud data; Input the preprocessed point cloud data into the feature extraction module, and map the point cloud coordinates to the high-dimensional feature space through the feature embedding layer to obtain initial features; Inputting the initial features into the elevation embedding module and the multi-resolution module for feature enhancement and feature extraction respectively, to obtain elevation enhancement features and multi-resolution features; Fusing the elevation enhancement feature and the multi-resolution feature to obtain a fused feature; A classification task and / or a segmentation task is performed based on the fused features, and a classification result and / or a segmentation result is output.

2. The elevation-aware multi-resolution network processing method for transmission corridor point cloud classification according to claim 1 is characterized in that: The elevation embedding module includes: The Z-Sampling layer is used to search for the nearest neighbor of each point, calculate the local elevation difference, and concatenate the local elevation difference with the original coordinates to obtain the concatenated features; At least one feature processing layer is used to process the splicing features and output the elevation enhancement features.

3. The elevation-aware multi-resolution network processing method for transmission corridor point cloud classification according to claim 1 is characterized in that: The multi-resolution module includes: A dual-branch downsampling structure, comprising a main sampling path and an offset sampling path, for performing a downsampling operation on the initial features to obtain features of different resolutions; The feature fusion structure is used to upsample the features of different resolutions to the original number of points, and perform cross-resolution splicing and fusion to obtain the multi-resolution features.

4. The elevation-aware multi-resolution network processing method for transmission corridor point cloud classification according to claim 1 is characterized in that: The feature extraction module includes an attention feature extractor, which is used to extract attention features from the initial features and enhance the correlation of local features.

5. The elevation-aware multi-resolution network processing method for transmission corridor point cloud classification according to claim 1 is characterized in that: The performing of a classification task based on the fused features and outputting a classification result includes: A global maximum pooling operation is performed on the fused features to obtain global features, and the scene type label is output through a fully connected layer.

6. The elevation-aware multi-resolution network processing method for transmission corridor point cloud classification according to claim 1, characterized in that: The performing of a segmentation task based on the fused features and outputting a segmentation result includes: The fused features are spliced ​​point by point, and spliced ​​with the global features to obtain enhanced features; Based on the enhanced features, the category probability of each point is output through a point-by-point classification operation.

7. The elevation-aware multi-resolution network processing method for transmission corridor point cloud classification according to any one of claims 1 to 6, characterized in that: The preprocessing includes filtering out outliers and noise points, and normalizing the point cloud data.

8. An elevation-aware multi-resolution network processing system for transmission corridor point cloud classification, characterized by: include: A data preprocessing unit, used to preprocess the LiDAR point cloud data of the transmission corridor to obtain preprocessed point cloud data; A feature extraction module is used to input the preprocessed point cloud data into a feature embedding layer, map the point cloud coordinates to a high-dimensional feature space, and obtain initial features; An elevation embedding module, configured to perform feature enhancement on the initial features to obtain elevation enhanced features; A multi-resolution module, used to extract features from the initial features to obtain multi-resolution features; A feature fusion unit, configured to fuse the elevation enhancement feature and the multi-resolution feature to obtain a fused feature; A classification unit is used to perform a classification task and / or a segmentation task based on the fused features and output a classification result and / or a segmentation result.

9. An electronic device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the steps of the method according to any one of claims 1 to 7.

Images