Method for segmenting point cloud data of rod-shaped ground object and related device
By combining voxel feature calculation and neighborhood clustering with deep learning, the problem of incorrect segmentation caused by fracture in point cloud segmentation of rod-shaped ground features was solved, achieving high-precision semantic and instance segmentation of rod-shaped ground features and improving the semantic discrimination accuracy of rod-shaped ground features.
Patent Information
- Application Number
- CN202511110094.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-08
- Publication Date
- 2025-11-18
AI Technical Summary
Existing point cloud segmentation methods for stalk-shaped features in urban scenes suffer from point cloud fragmentation due to foliage occlusion and scanning blind spots, resulting in erroneous segmentation of stalk-shaped features into multiple segments. Furthermore, they struggle to achieve semantic recognition of stalk-shaped features, trees, and different types of stalk-shaped features, requiring manual annotation.
By calculating voxel features, a set of voxels with significant vertical attributes is selected. A horizontal single-layer and vertical double-layer neighborhood clustering strategy is adopted. The semantic segmentation of rod-shaped ground features is performed by combining point cloud geometric features and image deep learning, which enhances the structural connectivity of rod-shaped features in the vertical direction. Semantic recognition is performed by using multi-view projection and deep learning models.
It effectively penetrates point cloud fracture zones that are obscured by branches and leaves or have scanning blind spots, improving the semantic discrimination accuracy of rod-shaped objects, reducing instance segmentation errors, and enhancing the semantic recognition accuracy of ground objects such as streetlights and traffic lights.
Smart Images
Figure CN120976545A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, in particular to a rod-shaped ground object point cloud data segmentation method and related device. BACKGROUND
[0002] The current mainstream rod-shaped ground object point cloud segmentation method is mainly based on geometric feature clustering technology, such as Euclidean distance clustering, region growing algorithm, etc. These methods analyze the spatial distribution of point cloud data, and aggregate objects according to the distance threshold between adjacent points or local curvature features. In typical applications, the original point cloud is calculated for basic geometric properties such as normal vector and curvature, and the rod-shaped object is separated through spatial neighborhood search.
[0003] Due to factors such as tree leaf occlusion and device scanning blind area in urban scenes, the rod-shaped object point cloud often has local missing in the vertical dimension. The traditional method uses fixed radius neighborhood search or single layer neighborhood clustering mechanism, and the rod-shaped object is incorrectly segmented into multiple fragments at the point cloud fracture due to insufficient neighborhood connectivity, which further causes the seed object to be incomplete and the attached structure to be unable to be effectively completed in the region growing stage, and the broken fragments are misjudged as independent ground objects in subsequent processing, resulting in a significant increase in instance segmentation error rate. And it is difficult to realize semantic recognition between rod-shaped ground objects and trees and different rod-shaped ground objects by simply using geometric features, which requires manual annotation. SUMMARY
[0004] The embodiments of the present application provide a rod-shaped ground object point cloud data segmentation method and related device, which can first calculate voxel features in the point cloud data and select a voxel set with significant vertical attributes, then use a horizontal single layer constraint combined with a vertical double layer expansion neighborhood clustering strategy to effectively penetrate the point cloud fracture caused by leaf occlusion or scanning blind area, enhance the structural connectivity of the rod-shaped object in the vertical direction, and finally realize semantic segmentation of the rod-shaped ground object by combining point cloud geometric features and image deep learning.
[0005] The first aspect of the embodiments of the present application provides a rod-shaped ground object point cloud data segmentation method, which comprises: acquiring point cloud data of a rod-shaped ground object to be measured; dividing the point cloud data to obtain voxelized point cloud; performing feature calculation on the voxelized point cloud to obtain voxel features; determining a seed rod-shaped super voxel set according to the voxel features by using a screening method based on linearity dominance, normal vector verticality and elevation difference, and a horizontal single layer vertical double layer neighborhood clustering mechanism; performing region growing processing with plane distance constraint on the seed rod-shaped super voxel of the seed rod-shaped super voxel set to obtain a complete rod-shaped super voxel set; Multi-view projection is performed on each complete rod-shaped supervoxel of the complete rod-shaped supervoxel set to obtain a multi-view image of the rod-shaped object. The multi-view image of the rod-shaped object is input into a pre-trained deep learning model for rod-shaped ground object image recognition to obtain the semantic information of each object corresponding to the multi-view image of the rod-shaped object; Based on the semantic information of each object, the point cloud data of the measured rod-shaped ground features are semantically and instance-segmented to obtain the semantic and instance segmentation results of the rod-shaped ground features.
[0006] In one possible implementation, the step of performing feature calculation on the voxelized point cloud to obtain voxel features includes: Based on the voxelized point cloud, construct the covariance matrix; The covariance matrix is decomposed to obtain three eigenvalues and corresponding eigenvectors; Calculate the characteristic coefficients based on the three characteristic values; The eigenvector corresponding to the smallest eigenvalue is the normal vector, and the angle between the normal vector and the horizontal plane is calculated as the perpendicular angle of the normal vector. Calculate the standard deviation of elevation for all points within the voxel.
[0007] In one possible implementation of voxel features, the step of determining a seed rod-shaped supervoxel set based on the voxel features, using a screening method based on linearity dominance, normal vector verticality, and elevation difference, and a neighborhood clustering mechanism through a horizontal single-layer and vertical double-layer structure, includes: Based on the voxel features, the voxelized point cloud is vertically voxel-filtered to obtain a vertical voxel set; The vertical voxel set is subjected to a horizontal single-layer and vertical double-layer neighborhood clustering to obtain a suspected rod-shaped supervoxel set; Feature calculations are performed on the suspected rod-shaped object supervoxels in the suspected rod-shaped object supervoxel set to obtain the supervoxel features corresponding to each suspected rod-shaped object supervoxel; Based on the characteristics of the supervoxel corresponding to each suspected rod-shaped object, the set of suspected rod-shaped objects is subjected to seed rod-shaped screening to obtain a seed rod-shaped supervoxel set.
[0008] In one possible implementation, the vertical voxel filtering of the voxelized point cloud is performed using the following filtering formula: in, , , These represent the linearity, flatness, and dispersion characteristics among the voxel feature coefficients. The perpendicular angle of the normal vector. The maximum threshold for the perpendicular angle of the normal vector. For elevation standard deviation, This is the minimum threshold for the elevation standard deviation. In one possible implementation, the suspected rod-shaped object supervoxel set is seeded using a rod-shaped screening method based on the supervoxel features corresponding to each suspected rod-shaped object supervoxel. The seed rod-shaped screening formula is as follows: in, The vertical angle of the main direction vector (the angle between the main direction vector and the horizontal plane). The minimum threshold for the perpendicular angle of the main direction vector. For height, The minimum height threshold, Height above the ground The maximum threshold for height above ground. Let the projection radius be 1. , These are the maximum and minimum threshold values for the projection radius, respectively. The aspect ratio of the projection. The maximum aspect ratio threshold for projection. For the projected area, , These are the maximum and minimum thresholds for the projected area, respectively.
[0009] This example provides a method for segmenting point cloud data of rod-shaped features. First, voxels are constructed from the point cloud and voxel features are calculated. Based on the voxel features, voxels with vertical features are selected. Then, vertical voxels are clustered and feature filtered to obtain seed rod-shaped objects. Next, region growing with constraints is performed on the seed rod-shaped objects to obtain complete rod-shaped point clouds. Finally, multi-view projection is performed on each rod-shaped point cloud, and the images are input into a trained image recognition deep learning model for semantic recognition. Finally, the recognition results are mapped back to the point cloud data to complete the semantic and instance segmentation of rod-shaped feature point clouds. This method can first calculate voxel features in the point cloud data and select voxel sets with significant vertical attributes. Then, a neighborhood clustering strategy combining horizontal single-layer constraints and vertical double-layer expansion is adopted to effectively penetrate point cloud fracture zones caused by foliage occlusion or scanning blind spots, enhancing the structural connectivity of rod-shaped features in the vertical direction. Through multi-view projection and deep learning cross-modal recognition, while eliminating instance segmentation errors caused by fractures in traditional methods, the semantic discrimination accuracy of features such as streetlights and traffic lights is significantly improved.
[0010] A second aspect of this application provides a segmentation apparatus for rod-shaped ground feature point cloud data, the apparatus comprising: The acquisition unit is used to acquire point cloud data of the rod-shaped ground object being measured. The first processing unit is used to divide the point cloud data to obtain a voxelized point cloud; The second processing unit is used to perform feature calculations on the voxelized point cloud to obtain voxel features. The third processing unit is used to determine the seed rod-shaped supervoxel set based on the voxel characteristics, using a screening method based on linearity dominance, normal vector verticality and elevation difference, and a neighborhood clustering mechanism through a horizontal single layer and a vertical double layer. The fourth processing unit is used to perform planar distance constraint-based region growth processing on the seed rod-shaped supervoxels of the seed rod-shaped supervoxel set to obtain a complete rod-shaped supervoxel set; The fifth processing unit is used to perform multi-view projection on each complete rod-shaped object supervoxel of the complete rod-shaped object supervoxel set to obtain a multi-view image of the rod-shaped object. The sixth processing unit is used to input the multi-view image of the pole-shaped object into a pre-trained deep learning model for pole-shaped ground object image recognition to obtain the semantic information of each object corresponding to the multi-view image of the pole-shaped object. The seventh processing unit is used to perform semantic and instance segmentation on the point cloud data of the measured rod-shaped ground features based on the semantic information of each object, and to obtain the semantic and instance segmentation results of the rod-shaped ground features.
[0011] In one possible implementation, in the aspect of performing feature calculation on the voxelized point cloud to obtain voxel features, the second processing unit is configured to: Based on the voxelized point cloud, construct the covariance matrix; The covariance matrix is decomposed to obtain three eigenvalues and corresponding eigenvectors; Calculate the characteristic coefficients based on the three characteristic values; The eigenvector corresponding to the smallest eigenvalue is the normal vector, and the angle between the normal vector and the horizontal plane is calculated as the perpendicular angle of the normal vector. Calculate the standard deviation of elevation for all points within the voxel.
[0012] In one possible implementation, based on the voxel characteristics, a selection method based on linearity dominance, normal vector verticality, and elevation difference is used, along with a neighborhood clustering mechanism involving a horizontal single-layer and a vertical double-layer structure to determine the seed rod-shaped supervoxel set. The third processing unit is used for: Based on the voxel features, the voxelized point cloud is vertically voxel-filtered to obtain a vertical voxel set; The vertical voxel set is subjected to a horizontal single-layer and vertical double-layer neighborhood clustering to obtain a suspected rod-shaped supervoxel set; Feature calculations are performed on the suspected rod-shaped object supervoxels in the suspected rod-shaped object supervoxel set to obtain the supervoxel features corresponding to each suspected rod-shaped object supervoxel; Based on the characteristics of the supervoxel corresponding to each suspected rod-shaped object, the set of suspected rod-shaped objects is subjected to seed rod-shaped screening to obtain a seed rod-shaped supervoxel set.
[0013] In one possible implementation, in the aspect of performing vertical voxel filtering on the voxelized point cloud based on the voxel features to obtain a vertical voxel set, the third processing unit uses the following filtering formula for performing vertical voxel filtering on the voxelized point cloud: in, , , These represent the linearity, flatness, and dispersion characteristics among the voxel feature coefficients. The perpendicular angle of the normal vector. The maximum threshold for the perpendicular angle of the normal vector. For elevation standard deviation, This is the minimum threshold for the standard deviation of elevation.
[0014] In one possible implementation, regarding the seed rod-shaped supervoxel set being seeded based on the supervoxel features corresponding to each suspected rod-shaped supervoxel, the third processing unit performs seed rod-shaped screening on the suspected rod-shaped supervoxel set, using the following seed rod-shaped screening formula: in, The vertical angle of the main direction vector (the angle between the main direction vector and the horizontal plane). The minimum threshold for the perpendicular angle of the main direction vector. For height, The minimum height threshold, Height above the ground The maximum threshold for height above ground. Let the projection radius be 1. , These are the maximum and minimum threshold values for the projection radius, respectively. The aspect ratio of the projection. The maximum aspect ratio threshold for projection. For the projected area, , These are the maximum and minimum thresholds for the projected area, respectively.
[0015] A third aspect of this application provides a terminal including a processor, an input device, an output device, and a memory, wherein the processor, input device, output device, and memory are interconnected, and the memory is used to store a computer program, the computer program including program instructions, and the processor is configured to invoke the program instructions to execute the step instructions as described in the rod-shaped feature point cloud data segmentation method of the first aspect of this application.
[0016] A fourth aspect of this application provides a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in the method for segmenting rod-shaped feature point cloud data in the first aspect of this application.
[0017] A fifth aspect of this application provides a computer program product, comprising a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in the method for segmenting rod-shaped feature point cloud data in the first aspect of this application. The computer program product may be a software installation package. Attached Figure Description
[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0019] Figure 1 This application provides a schematic diagram of the overall process for segmenting point cloud data of rod-shaped ground features. Figure 2 This application provides a schematic diagram of the overall structure of a segmentation device for point cloud data of rod-shaped ground features. Figure 3 This application provides a schematic diagram of the structure of a terminal. Figure label: Acquisition Unit-1, First Processing Unit-2, Second Processing Unit-3, Third Processing Unit-4, Fourth Processing Unit-5, Fifth Processing Unit-6, Sixth Processing Unit-7, Seventh Processing Unit-8. Detailed Implementation
[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0021] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.
[0022] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.
[0023] To better understand the segmentation method for pole-shaped feature point cloud data provided in this application embodiment, the scenarios in which this method is applied are briefly introduced below. Three-dimensional laser scanning, as an efficient three-dimensional geospatial information acquisition technology, can quickly acquire detailed three-dimensional spatial information of roads and features on both sides. Pole-shaped features such as streetlights, traffic lights, traffic signs, and utility poles are important components of urban road scenarios. Semantic and instance segmentation of pole-shaped features based on point cloud data is of great significance for the construction of digital cities and smart cities. Existing pole-shaped feature segmentation methods suffer from the following technical bottlenecks: traditional point cloud clustering methods (such as Euclidean clustering and RANSAC) rely solely on spatial distance or geometric features, making it difficult to distinguish similar-shaped pole-shaped objects (such as streetlights and trees); pole-shaped features exhibit significant vertical continuity, but existing neighborhood definitions do not fully utilize this feature, leading to cluster breaks; semantic recognition is lacking, and pure geometric clustering cannot achieve semantic differentiation between different categories such as streetlights and traffic lights, requiring manual annotation and post-processing.
[0024] To address the aforementioned technical problems, this embodiment provides a method for segmenting rod-shaped ground feature point cloud data, which is applied to a rod-shaped ground feature point cloud data segmentation device. Figure 1A schematic diagram illustrating the overall process of a method for segmenting point cloud data of rod-shaped ground features is shown. Figure 1 As shown, it includes: S1. Obtain point cloud data of the rod-shaped ground object being measured.
[0025] Point cloud data can be collected using stand-alone, backpack, handheld, or vehicle-mounted laser scanners.
[0026] S2. Divide the point cloud data to obtain a voxelized point cloud.
[0027] The voxel partitioning size is set to 0.2m×0.2m×0.2m. Then, voxels with fewer than 5 points in the point cloud are removed to obtain the voxelized point cloud.
[0028] S3. Perform feature calculation on the voxelized point cloud to obtain voxel features.
[0029] Among them, voxel features include feature coefficients (flatness) linearity Dispersion ), perpendicular angle of normal vector (Angle between the normal vector and the horizontal plane), standard deviation of elevation The eigenvalues are calculated as follows: First, construct the covariance matrix of the point cloud within the voxel, and then decompose the covariance matrix to obtain three eigenvalues. and the corresponding feature vector Then, the characteristic coefficients are calculated using the following formula.
[0030] S4. Based on the voxel characteristics, a seed rod-shaped supervoxel set is determined by a screening method based on linearity dominance, verticality of normal vectors, and elevation differences, and by a neighborhood clustering mechanism of horizontal single layer and vertical double layer.
[0031] First, the voxelized point cloud is vertically voxel-filtered based on voxel features. Then, the voxels after vertical voxel filtering are clustered in their neighborhoods. After clustering, the supervoxel objects that are suspected to be rod-shaped objects obtained from the neighborhood clustering are seed objects to obtain a seed rod-shaped supervoxel set.
[0032] S5. Perform a region growth process with planar distance constraints on the seed rod-shaped supervoxels of the seed rod-shaped supervoxel set to obtain a complete rod-shaped supervoxel set.
[0033] Specifically, region growth is performed on the seed rod-shaped object within the global voxel scope. For the initial rod object... Query all neighboring voxels of its voxel, and for each neighboring voxel... If to the object The distance between the centroid planes is less than the maximum radius threshold of the rod. Then include the current object. By limiting the growth radius, excessive object growth can be prevented. If a voxel is contained by multiple objects, the object to which the voxel belongs is determined by the nearest neighbor principle.
[0034] S6. Perform multi-view projection on each complete rod-shaped object supervoxel of the complete rod-shaped object supervoxel set to obtain a multi-view image of the rod-shaped object.
[0035] In this process, the point cloud of each rod-shaped object is projected from multiple perspectives to obtain a multi-view image of the rod-shaped object. Generally, frontal, side, and top views are projected to obtain a three-view image.
[0036] S7. Input the multi-view image of the pole-shaped object into a pre-trained deep learning model for pole-shaped ground object image recognition to obtain the semantic information of each object corresponding to the multi-view image of the pole-shaped object.
[0037] Among them, ResNet50 can be selected as the deep learning model for image recognition. During training, at least 200 samples of various rod-shaped land features of each type are input for training with multiple view images. Data augmentation strategies such as random rotation and brightness adjustment are used during training to improve the generalization performance of the model.
[0038] In this process, multiple view images of each object can be input into a trained image recognition deep learning model for image recognition to obtain the semantic information of each object.
[0039] S8. Based on the semantic information of each object, perform semantic and instance segmentation on the point cloud data of the measured rod-shaped ground features to obtain the semantic and instance segmentation results of the rod-shaped ground features.
[0040] By mapping the multi-view semantic information of each object to the corresponding point cloud data, the semantic and instance segmentation of the pole-shaped ground feature point cloud can be completed, and the semantic and instance segmentation results of the pole-shaped ground feature can be obtained.
[0041] In this example, voxels are first constructed from the point cloud and their features are calculated. Voxels with vertical features are selected based on these features. Then, the vertical voxels are clustered and feature filtered to obtain seed rod-shaped objects. Next, region growing with constraints is performed on the seed rod-shaped objects to obtain a complete rod-shaped point cloud. Finally, multi-view projection is performed on each rod-shaped point cloud, and the images are input into a trained image recognition deep learning model for semantic recognition. The recognition results are then mapped back to the point cloud data to complete the semantic and instance segmentation of the rod-shaped ground feature point cloud. This method can first calculate voxel features in the point cloud data and select a set of voxels with significant vertical attributes. Then, a neighborhood clustering strategy combining horizontal single-layer constraints and vertical double-layer expansion is adopted to effectively penetrate the point cloud fracture zones caused by foliage occlusion or scanning blind spots, enhancing the structural connectivity of rod-shaped objects in the vertical direction. Through multi-view projection and deep learning cross-modal recognition, the method can eliminate instance segmentation errors caused by fractures in traditional methods while significantly improving the semantic discrimination accuracy of ground features such as streetlights and traffic lights.
[0042] In one possible implementation, the step of performing feature calculation on the voxelized point cloud to obtain voxel features includes: Based on the voxelized point cloud, construct the covariance matrix; The covariance matrix is decomposed to obtain three eigenvalues and corresponding eigenvectors; Calculate the characteristic coefficients based on the three characteristic values; The eigenvector corresponding to the smallest eigenvalue is the normal vector, and the angle between the normal vector and the horizontal plane is calculated as the perpendicular angle of the normal vector. Calculate the standard deviation of elevation for all points within the voxel.
[0043] In one possible implementation, based on the voxel characteristics, a seed rod-shaped supervoxel set is determined using a screening method based on linearity dominance, normal vector verticality, and elevation difference, and a neighborhood clustering mechanism through a horizontal single-layer and vertical double-layer structure, including: S301. Based on the voxel features, perform vertical voxel filtering on the voxelized point cloud to obtain a vertical voxel set; Specifically, the screening formula used is as follows: in, , , These represent the linearity, flatness, and dispersion characteristics among the voxel feature coefficients. The perpendicular angle of the normal vector. The maximum threshold for the perpendicular angle of the normal vector. For elevation standard deviation, This is the minimum threshold for the standard deviation of elevation.
[0044] Among them, by statistically analyzing 1000+ rod-shaped samples, setting the angle to 10° and the depth to 0.3m, a set of vertical voxels was obtained after screening. .
[0045] S302. Perform neighborhood clustering on the vertical voxel set with a horizontal single layer and a vertical double layer to obtain a suspected rod-shaped supervoxel set; Among them, in the vertical voxel set Neighborhood clustering is performed within the target voxel to accommodate the continuous characteristics of the rod-shaped object in the vertical dimension. Its neighborhood voxels are defined as follows: In the horizontal direction (XY plane), adjacent voxels only include directly adjacent voxels with a common surface (i.e., voxels adjacent in the X or Y axis direction in 3D space); in the vertical direction (Z axis direction), adjacent voxels include: directly adjacent voxels (voxels with a common surface in the Z axis direction) and adjacent voxels of adjacent voxels (i.e., voxels that are further adjacent to the directly adjacent voxels in the Z axis direction, forming two layers of neighborhood extension). in, Let be the three-dimensional coordinates of the target voxel. For the horizontal neighborhood voxels of the target voxel, The vertical neighborhood voxels of the target voxel.
[0046] The clustering method in this example can enhance vertical structural connectivity while maintaining horizontal clustering accuracy, thus improving the integrity of rod-shaped clusters. Multiple supervoxel objects suspected to be rod-shaped are obtained through clustering.
[0047] S303. Perform feature calculation on the suspected rod-shaped object supervoxels in the suspected rod-shaped object supervoxel set to obtain the supervoxel features corresponding to each suspected rod-shaped object supervoxel.
[0048] This involves calculating the features of each hypervoxel, including the vertical angle of the principal direction vector. (Angle between the principal direction vector and the horizontal plane), height Height above ground Projection radius Projected aspect ratio Projected area .
[0049] S304. Based on the characteristics of the supervoxel corresponding to each suspected rod-shaped object supervoxel, seed rod-shaped screening is performed on the suspected rod-shaped object supervoxel set to obtain a seed rod-shaped supervoxel set.
[0050] Among them, the hypervoxels that satisfy the following formula are used as seed rod objects, and the set of seed rod objects is used to obtain the seed rod hypervoxel set.
[0051] in, The vertical angle of the main direction vector (the angle between the main direction vector and the horizontal plane). The minimum threshold for the perpendicular angle of the main direction vector. For height, The minimum height threshold, Height above the ground The maximum threshold for height above ground. Let the projection radius be 1. , These are the maximum and minimum threshold values for the projection radius, respectively. The aspect ratio of the projection. The maximum aspect ratio threshold for projection. For the projected area, , These are the maximum and minimum thresholds for the projected area, respectively.
[0052] For those consistent with the above, please refer to Figure 2 , Figure 2 This application provides a schematic diagram of the structure of a device for segmenting point cloud data of rod-shaped ground features. For example... Figure 2 As shown, the device includes: Acquisition unit 1 is used to acquire point cloud data of the rod-shaped ground object being measured; The first processing unit 2 is used to divide the point cloud data to obtain a voxelized point cloud; The second processing unit 3 is used to perform feature calculations on the voxelized point cloud to obtain voxel features; The third processing unit 4 is used to determine the seed rod-shaped supervoxel set based on the voxel characteristics, using a screening method based on linearity dominance, normal vector verticality and elevation difference, and a neighborhood clustering mechanism through a horizontal single layer and a vertical double layer. The fourth processing unit 5 is used to perform planar distance constraint region growth processing on the seed rod-shaped supervoxels of the seed rod-shaped supervoxel set to obtain a complete rod-shaped supervoxel set; The fifth processing unit 6 is used to perform multi-view projection on each complete rod-shaped object supervoxel of the complete rod-shaped object supervoxel set to obtain a multi-view image of the rod-shaped object. The sixth processing unit 7 is used to input the multi-view image of the pole-shaped object into a pre-trained deep learning model for pole-shaped ground object image recognition to obtain the semantic information of each object corresponding to the multi-view image of the pole-shaped object; The seventh processing unit 8 is used to perform semantic and instance segmentation on the point cloud data of the measured rod-shaped ground features based on the semantic information of each object, and to obtain the semantic and instance segmentation results of the rod-shaped ground features.
[0053] In one possible implementation, in the aspect of performing feature calculation on the voxelized point cloud to obtain voxel features, the second processing unit 3 is used to: Based on the voxelized point cloud, construct the covariance matrix; The covariance matrix is decomposed to obtain three eigenvalues and corresponding eigenvectors; Calculate the characteristic coefficients based on the three characteristic values; The eigenvector corresponding to the smallest eigenvalue is the normal vector, and the angle between the normal vector and the horizontal plane is calculated as the perpendicular angle of the normal vector. Calculate the standard deviation of elevation for all points within the voxel.
[0054] In one possible implementation, regarding the determination of the seed rod-shaped supervoxel set based on the voxel characteristics using a screening method based on linearity dominance, normal vector verticality, and elevation difference, and through a neighborhood clustering mechanism of horizontal single-layer and vertical double-layer, the third processing unit 4 is used for: Based on the voxel features, the voxelized point cloud is vertically voxel-filtered to obtain a vertical voxel set; The vertical voxel set is subjected to a horizontal single-layer and vertical double-layer neighborhood clustering to obtain a suspected rod-shaped supervoxel set; Feature calculations are performed on the suspected rod-shaped object supervoxels in the suspected rod-shaped object supervoxel set to obtain the supervoxel features corresponding to each suspected rod-shaped object supervoxel; Based on the characteristics of the supervoxel corresponding to each suspected rod-shaped object, the set of suspected rod-shaped objects is subjected to seed rod-shaped screening to obtain a seed rod-shaped supervoxel set.
[0055] In one possible implementation, regarding the aspect of performing vertical voxel filtering on the voxelized point cloud based on the voxel features to obtain a vertical voxel set, in the third processing unit 4, the filtering formula used for performing vertical voxel filtering on the voxelized point cloud is as follows: in, These represent the linearity, flatness, and dispersion characteristics among the voxel feature coefficients. The perpendicular angle of the normal vector. The maximum threshold for the perpendicular angle of the normal vector. For elevation standard deviation, This is the minimum threshold for the standard deviation of elevation.
[0056] In one possible implementation, in the aspect of performing vertical voxel screening on the voxelized point cloud based on the voxel features to obtain a vertical voxel set, in the third processing unit 4, the seeded rod-shaped screening is performed on the suspected rod-shaped supervoxel set, and the seeded rod-shaped screening formula is as follows: in, The vertical angle of the main direction vector (the angle between the main direction vector and the horizontal plane). The minimum threshold for the perpendicular angle of the main direction vector. For height, The minimum height threshold, Height above the ground The maximum threshold for height above ground. Let the projection radius be 1. , These are the maximum and minimum threshold values for the projection radius, respectively. The aspect ratio of the projection. The maximum aspect ratio threshold for projection. For the projected area, , These are the maximum and minimum thresholds for the projected area, respectively.
[0057] For examples consistent with the above embodiments, please refer to... Figure 3 , Figure 3 A schematic diagram of a terminal structure provided in an embodiment of this application is shown in the figure. It includes a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions. The program includes instructions for performing the following steps. Acquire point cloud data of the rod-shaped ground object being measured; The point cloud data is divided to obtain a voxelized point cloud; The voxelized point cloud is subjected to feature calculation to obtain voxel features; Based on the voxel characteristics, a selection method based on linearity dominance, verticality of normal vectors, and elevation differences is adopted, and a neighborhood clustering mechanism with horizontal single layer and vertical double layer is used to determine the seed rod-shaped supervoxel set; The seed rod-shaped supervoxels of the seed rod-shaped supervoxel set are subjected to a region growth process with planar distance constraints to obtain a complete rod-shaped supervoxel set; Multi-view projection is performed on each complete rod-shaped supervoxel of the complete rod-shaped supervoxel set to obtain a multi-view image of the rod-shaped object. The multi-view image of the rod-shaped object is input into a pre-trained deep learning model for rod-shaped ground object image recognition to obtain the semantic information of each object corresponding to the multi-view image of the rod-shaped object; Based on the semantic information of each object, the point cloud data of the measured rod-shaped ground features are semantically and instance-segmented to obtain the semantic and instance segmentation results of the rod-shaped ground features.
[0058] In this example, voxels are first constructed from the point cloud and their features are calculated. Voxels with vertical features are then selected based on these features. Next, vertical voxels are clustered and feature-filtered to obtain seed rod-shaped objects. Constrained region growing is then performed on these seed rod-shaped objects to obtain a complete rod-shaped point cloud. Finally, multi-view projection is performed on each rod-shaped point cloud, and the images are input into a trained image recognition deep learning model for semantic recognition. The recognition results are then mapped back to the point cloud data, completing the semantic and instance segmentation of the rod-shaped ground feature point cloud. This approach first calculates voxel features from the point cloud data and selects voxel sets with significant vertical attributes. Then, a neighborhood clustering strategy combining horizontal single-layer constraints and vertical double-layer expansion is employed to effectively penetrate point cloud fractures caused by foliage occlusion or scanning blind spots, enhancing the structural connectivity of rod-shaped objects in the vertical direction. Through multi-view projection and deep learning cross-modal recognition, the method eliminates instance segmentation errors caused by fractures in traditional methods while significantly improving the semantic discrimination accuracy of ground features such as streetlights and traffic lights.
[0059] The above mainly describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the terminal includes the corresponding hardware structure and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed in hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0060] This application embodiment can divide the terminal into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.
[0061] This application also provides a computer storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the rod-shaped feature point cloud data segmentation methods described in the above method embodiments.
[0062] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program that causes a computer to perform some or all of the steps of any of the rod-shaped feature point cloud data segmentation methods described in the above method embodiments.
[0063] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.
[0064] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.
[0065] In the several embodiments provided in this application, it should be understood that the disclosed apparatus can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical or other forms.
[0066] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.
[0067] Furthermore, the functional units in the various embodiments of the application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software program module.
[0068] If the integrated unit is implemented as a software program module and sold or used as an independent product, it can be stored in a computer-readable storage device (CMD). Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a memory and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned memory includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.
[0069] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage device, which may include: a flash drive, a read-only memory, a random access memory, a magnetic disk, or an optical disk, etc.
[0070] The embodiments of this application have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of this application. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of this application. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A method for segmenting point cloud data of rod-shaped ground features, characterized in that, include: Acquire point cloud data of the rod-shaped ground object being measured; The point cloud data is divided into voxels to obtain voxelized point clouds; The voxelized point cloud is subjected to feature calculation to obtain voxel features; Based on the voxel characteristics, a selection method based on linearity dominance, verticality of normal vectors, and elevation differences is adopted, and a neighborhood clustering mechanism with horizontal single layer and vertical double layer is used to determine the seed rod-shaped supervoxel set; The seed rod-shaped supervoxels of the seed rod-shaped supervoxel set are subjected to a region growth process with planar distance constraints to obtain a complete rod-shaped supervoxel set; Multi-view projection is performed on each complete rod-shaped supervoxel of the complete rod-shaped supervoxel set to obtain a multi-view image of the rod-shaped object. The multi-view image of the rod-shaped object is input into a pre-trained deep learning model for rod-shaped ground object image recognition to obtain the semantic information of each object corresponding to the multi-view image of the rod-shaped object. Based on the semantic information of each object, the point cloud data of the measured rod-shaped ground features are semantically and instance-segmented to obtain the semantic and instance segmentation results of the rod-shaped ground features.
2. The method for segmenting point cloud data of rod-shaped ground features according to claim 1, characterized in that, The process of calculating features from the voxelized point cloud to obtain voxel features includes: Based on the voxelized point cloud, a covariance matrix is constructed using voxel features; The covariance matrix is decomposed to obtain three eigenvalues and corresponding eigenvectors; Calculate the characteristic coefficients based on the three characteristic values; The eigenvector corresponding to the smallest eigenvalue is the normal vector, and the angle between the normal vector and the horizontal plane is calculated as the perpendicular angle of the normal vector. Calculate the standard deviation of elevation for all points within the voxel.
3. The method for segmenting point cloud data of rod-shaped ground features according to claim 1, characterized in that, The process of determining the seed rod-shaped supervoxel set based on the voxel characteristics, using a screening method based on linearity dominance, normal vector verticality, and elevation difference, and a neighborhood clustering mechanism through a horizontal single-layer and vertical double-layer structure, includes: Based on the voxel features, the voxelized point cloud is vertically voxel-filtered to obtain a vertical voxel set; The vertical voxel set is subjected to horizontal single-layer and vertical double-layer neighborhood clustering to obtain a suspected rod-shaped supervoxel set; Feature calculations are performed on the suspected rod-shaped object supervoxels in the suspected rod-shaped object supervoxel set to obtain the supervoxel features corresponding to each suspected rod-shaped object supervoxel; Based on the characteristics of the supervoxel corresponding to each suspected rod-shaped object, the set of suspected rod-shaped objects is subjected to seed rod-shaped screening to obtain a seed rod-shaped supervoxel set.
4. The method for segmenting point cloud data of rod-shaped ground features according to claim 3, characterized in that, The vertical voxel filtering of the voxelized point cloud is performed using the following filtering formula: in, , , These represent the linearity, flatness, and dispersion characteristics among the voxel feature coefficients. The perpendicular angle of the normal vector. The maximum threshold for the perpendicular angle of the normal vector. For elevation standard deviation, This is the minimum threshold for the standard deviation of elevation.
5. The method for segmenting point cloud data of rod-shaped ground features according to claim 3, characterized in that, The seed rod-shaped screening process is performed on the set of suspected rod-shaped supervoxels based on the supervoxel features corresponding to each suspected rod-shaped supervoxel. The seed rod-shaped screening formula is as follows: in, The vertical angle of the main direction vector (the angle between the main direction vector and the horizontal plane). The minimum threshold for the perpendicular angle of the main direction vector. For height, The minimum height threshold, Height above the ground The maximum threshold for height above ground. Let the projection radius be 1. , These are the maximum and minimum threshold values for the projection radius, respectively. The aspect ratio of the projection. The maximum aspect ratio threshold for projection. For the projected area, , These are the maximum and minimum thresholds for the projected area, respectively.
6. A device for segmenting point cloud data of rod-shaped ground features, characterized in that, include: The acquisition unit is used to acquire point cloud data of the rod-shaped ground object being measured. The first processing unit is used to divide the point cloud data to obtain a voxelized point cloud; The second processing unit is used to perform feature calculations on the voxelized point cloud to obtain voxel features. The third processing unit is used to determine the seed rod-shaped supervoxel set based on the voxel characteristics, using a screening method based on linearity dominance, normal vector verticality and elevation difference, and a neighborhood clustering mechanism through a horizontal single layer and a vertical double layer. The fourth processing unit is used to perform planar distance constraint-based region growth processing on the seed rod-shaped supervoxels of the seed rod-shaped supervoxel set to obtain a complete rod-shaped supervoxel set; The fifth processing unit is used to perform multi-view projection on each complete rod-shaped object supervoxel of the complete rod-shaped object supervoxel set to obtain a multi-view image of the rod-shaped object. The sixth processing unit is used to input the multi-view image of the pole-shaped object into a pre-trained deep learning model for pole-shaped ground object image recognition to obtain the semantic information of each object corresponding to the multi-view image of the pole-shaped object. The seventh processing unit is used to perform semantic and instance segmentation on the point cloud data of the measured rod-shaped ground features based on the semantic information of each object, and to obtain the semantic and instance segmentation results of the rod-shaped ground features.
7. The segmentation device for rod-shaped ground feature point cloud data according to claim 6, characterized in that, The second processing unit performs feature calculations on the voxelized point cloud to obtain aspects of the voxel features, and is used to: Based on the voxelized point cloud, construct the covariance matrix; The covariance matrix is decomposed to obtain three eigenvalues and corresponding eigenvectors; Calculate the characteristic coefficients based on the three characteristic values; The eigenvector corresponding to the smallest eigenvalue is the normal vector, and the angle between the normal vector and the horizontal plane is calculated as the perpendicular angle of the normal vector. Calculate the standard deviation of elevation for all points within the voxel.
8. The segmentation device for rod-shaped ground feature point cloud data according to claim 6, characterized in that, In determining the seed rod-shaped supervoxel set based on the voxel characteristics using a screening method based on linearity dominance, normal vector verticality, and elevation difference, and through a neighborhood clustering mechanism of horizontal single-layer and vertical double-layer, the third processing unit is used for: Based on the voxel features, the voxelized point cloud is vertically voxel-filtered to obtain a vertical voxel set; The vertical voxel set is subjected to horizontal single-layer and vertical double-layer neighborhood clustering to obtain a suspected rod-shaped supervoxel set; Feature calculations are performed on the suspected rod-shaped object supervoxels in the suspected rod-shaped object supervoxel set to obtain the supervoxel features corresponding to each suspected rod-shaped object supervoxel; Based on the characteristics of the supervoxel corresponding to each suspected rod-shaped object, the set of suspected rod-shaped objects is subjected to seed rod-shaped screening to obtain a seed rod-shaped supervoxel set.
9. A terminal, characterized in that, The system includes a processor, an input device, an output device, and a memory, which are interconnected. The memory stores a computer program, which includes program instructions. The processor is configured to invoke the program instructions to execute the segmentation method for rod-shaped feature point cloud data as described in any one of claims 1-5.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the segmentation method for rod-shaped feature point cloud data as described in any one of claims 1-5.