Intelligent segmentation method and system for spatial steel structure nodes

Abstract

The present invention discloses an intelligent segmentation method and system for spatial steel structure nodes, belonging to the field of building construction technology. The method comprises: collecting point cloud data of steel structures, and preprocessing the point cloud data; clustering point clouds to form a number of supervoxels, and calculating the concavity and convexity between adjacent supervoxels; abstracting the boundary between adjacent supervoxels into a point cloud with convex and concave information, and assigning weight values ​​to concave points and convex points, and calculating the best segmentation plane; segmenting the supervoxels according to the best segmentation plane, and optimizing the segmentation results, so as to accurately segment the nodes from the entire steel structure model, so as to facilitate the monitoring of the nodes using a finite element model. The present invention can realize the intelligent and rapid identification and segmentation of complex spatial steel structure nodes, significantly improve the recognition and segmentation efficiency of nodes, ensure that the segmentation results have a high degree of consistency, and improve the accuracy of the segmentation results, thereby ensuring the comprehensiveness and accuracy of damage detection.

Description

Technical Field

[0001] The invention relates to an intelligent segmentation method and system for a space steel structure node, belonging to the technical field of building construction. Background Art

[0002] In large steel structures, there are numerous connection nodes between components. Damage to the connection nodes may cause the failure of the entire structure. Therefore, node damage detection and analysis are crucial. In order to conduct effective detection, the nodes need to be accurately segmented from the entire steel structure model for detailed analysis and evaluation. Traditional node segmentation methods usually rely on manual operation and are inefficient, especially in complex spatial structures with a large number of nodes and dense rods, making this process very time-consuming and labor-intensive, and difficult to meet the needs of large-scale or emergency projects. At the same time, manual operation is bound to have errors, especially in detail processing, which may miss some important damage or features, affecting the accuracy of subsequent analysis.

[0003] Therefore, it is necessary to provide an intelligent segmentation method and system for spatial steel structure nodes. Summary of the invention

[0004] In view of the problem that traditional node segmentation methods rely on manual operation, the present invention provides an intelligent segmentation method and system for spatial steel structure nodes, which can realize intelligent and rapid identification and segmentation of complex spatial steel structure nodes, significantly improve the detection efficiency of nodes, and ensure the comprehensiveness and accuracy of damage detection.

[0005] In order to solve the above technical problems, the present invention includes the following technical solutions:

[0006] An intelligent segmentation method for spatial steel structure nodes comprises the following steps:

[0007] S1. Collect point cloud data of the actual scene of the steel structure, and remove noise points, downsample voxels and smooth the point cloud data;

[0008] S2. clustering the point cloud data processed in step S1 according to spatial proximity and feature similarity to form a number of supervoxels, and calculating the concavity and convexity between adjacent supervoxels;

[0009] S3. The boundary between adjacent supervoxels is abstracted into a point cloud with convex and concave information, the concave points are assigned a weight value a, a>0, and the convex points are assigned a weight value 0 or a negative value, and the weighted random sampling consensus algorithm is used to obtain the best segmentation plane;

[0010] S4. Segment the supervoxel according to the best segmentation plane, and perform optimization operations on the segmentation result such as smoothing the segmentation boundary, removing small areas, and merging contact areas.

[0011] Furthermore, calculating the concavity and convexity between adjacent supervoxels specifically includes:

[0012] The centroid vector connecting two adjacent supervoxels is denoted as , the normal vector is recorded as , , and , The angle between α 1. α 2. If α 1> α 2, it is considered a concave connection if α 1< α 2, it is considered a convex connection;

[0013] The cross product of the normal vectors of two adjacent supervoxels is denoted as , , and The angle between v , set the angle threshold β Thresh ,if v < β Thresh , then the convexity determination is invalid, if v ≥ β Thresh , then the concavity and convexity determination is valid.

[0014] Further, in step S3, when the weighted random sampling consensus algorithm is used to obtain the best segmentation plane, specifically:

[0015] Select points with concave-convex information and weights as seed points of the segmentation plane, and use these seed points to fit a segmentation plane based on their three-dimensional coordinates;

[0016] Calculate the distance from all points to the current fitting plane, and combine the weight of the points to determine the fit of the plane;

[0017] After multiple iterations, the segmentation plane with the highest fit and the smallest error is retained as the candidate best model, and those outlier points that are too far from the plane are excluded;

[0018] The algorithm converges to a stable optimal splitting plane solution.

[0019] Accordingly, the present invention also provides an intelligent segmentation system for spatial steel structure nodes, including a point cloud input system, a local feature extraction system, a segmentation position extraction system, and a segmentation and optimization system;

[0020] The point cloud input system includes a point cloud acquisition module and a point cloud preprocessing module. The point cloud acquisition module can collect point cloud data of the actual scene of the steel structure, and the point cloud preprocessing module can remove noise points, downsample voxels and perform smoothing on the point cloud data.

[0021] The local feature extraction system includes a supervoxel clustering module and a concave-convex relationship calculation module. The supervoxel clustering module can cluster point clouds based on spatial proximity and feature similarity of point clouds to form a number of supervoxels. The concave-convex relationship calculation module is used to calculate the concave-convexity between adjacent supervoxels.

[0022] The segmentation position extraction system includes a Euclidean edge cloud construction module and a geometric constraint partitioning module. The Euclidean edge cloud construction module can abstract the edges between adjacent supervoxels into point clouds, and assign a weight value a, a>0 to concave points, and a weight value 0 or negative to convex points. The geometric constraint partitioning module can use a weighted random sampling consensus algorithm to obtain the best segmentation plane.

[0023] The segmentation and optimization system includes a local constraint cutting module and a post-processing module. The local constraint cutting module can segment the supervoxel according to the segmentation plane, and the post-processing module can optimize the segmentation result by smoothing the segmentation boundary, removing small areas and merging contact areas.

[0024] Due to the adoption of the above technical solution, the present invention has the following advantages and positive effects compared with the prior art: the intelligent segmentation method and system of spatial steel structure nodes provided by the present invention collects point cloud data of the steel structure, forms a number of supervoxels by clustering point clouds, and calculates the concavity and convexity between adjacent supervoxels, abstracts the connecting edges between adjacent supervoxels into point clouds and performs concavity and convexity negative values, finds the best segmentation plane, and segments the supervoxels according to the best segmentation plane, thereby accurately segmenting the nodes from the entire steel structure model, so as to facilitate the monitoring of the nodes using the finite element model. The present invention develops an improved point cloud clustering network segmentation algorithm based on concavity and convexity, realizes the intelligent and rapid identification and segmentation of complex spatial steel structure nodes and components, solves the shortcomings of traditional manual segmentation, significantly improves the recognition and segmentation efficiency of nodes, can ensure that the segmentation results have a high degree of consistency, avoids the subjective differences that may occur in manual operation, reduces human errors, and improves the accuracy of segmentation results, thereby ensuring the comprehensiveness and accuracy of damage detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of a method for intelligent segmentation of a spatial steel structure node in one embodiment of the present invention;

[0026] Figure 2 A schematic diagram of a supervoxel of a single-layer lattice shell structure in one embodiment of the present invention;

[0027] Figure 3 is a schematic diagram of a segmented single-layer lattice shell structure in one embodiment of the present invention;

[0028] Figure 4 for Figure 3 Enlarged view of the middle G region;

[0029] Figure 5 It is a structural block diagram of an intelligent segmentation system for spatial steel structure nodes in one embodiment of the present invention.

[0030] The numbers in the figure are as follows:

[0031] 1-Supervoxel of single-layer lattice shell structure; 2-Segmented single-layer lattice shell structure. DETAILED DESCRIPTION

[0032] The following is a further detailed description of the intelligent segmentation method and system for spatial steel structure nodes provided by the present invention in conjunction with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer in conjunction with the following description. It should be noted that the accompanying drawings are all in a very simplified form and are not in precise proportions, and are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention. Embodiment 1

[0033] In order to solve the problems existing in the traditional node identification that relies on manual operation, this embodiment provides an intelligent segmentation method for spatial steel structure nodes, such as Figure 1 As shown, the method comprises the following steps:

[0034] S1. Collect point cloud data of the actual scene of the steel structure, and remove noise points, perform voxel downsampling and smoothing on the point cloud data; the SOR algorithm can be used to remove noise points. The SOR (Successive Over Relaxation) algorithm is a classic iterative algorithm. The present invention uses this algorithm to remove noise points; Voxel downsampling (Voxel Grid Downsampling) is a point cloud processing technology used to reduce the density of the point cloud and the amount of data while maintaining the characteristics of the point cloud. It is achieved by defining a voxel grid in three-dimensional space and then replacing the point cloud data in each voxel with a single point;

[0035] S2. Cluster the point cloud data processed in step S1 according to spatial proximity and feature similarity to form several supervoxels, and calculate the concavity and convexity between adjacent supervoxels. Specifically, the SLIC algorithm can be used to cluster the point cloud based on the spatial proximity and feature similarity of the point cloud. SLIC (Simple Linear Iterative Clustering) is a simple linear iterative clustering that can generate compact and approximately uniform supervoxels. A supervoxel is a set whose elements are "bodies", which are essentially small squares.

[0036] S3. The boundary between adjacent supervoxels is abstracted into a point cloud with convex and concave information, and the concave points are given a weight value a, a>0, usually a is set to 1, and the convex points are given a weight value 0 or a negative value, and the weighted random sampling consensus algorithm is used to obtain the best segmentation plane;

[0037] S4. Segment the supervoxel according to the best segmentation plane, and perform optimization operations on the segmentation result by smoothing, removing small areas, and merging contact areas.

[0038] The intelligent segmentation method of spatial steel structure nodes provided by the present invention obtains point cloud data based on three-dimensional computer vision recognition technology and performs noise point removal, voxel downsampling and smoothing processing on the point cloud data, then uses a clustering point cloud algorithm to form supervoxels, and determines the concavity and convexity between adjacent supervoxels, and abstracts the point cloud of the edge of the supervoxel and assigns concavity and convexity, and uses a weighted random sampling consistency algorithm to obtain the best segmentation plane, and then performs supervoxel segmentation to separate the steel structure nodes from the components, so as to facilitate the subsequent detection and analysis of the steel structure nodes by finite elements, significantly improve the recognition and segmentation efficiency of the steel structure nodes, and can ensure that the segmentation results have a high degree of consistency, and improve the accuracy of the segmentation results, thereby ensuring the comprehensiveness and accuracy of damage detection.

[0039] Take the single-layer lattice shell structure as an example. Figure 2 A schematic diagram of a supervoxel of a single-layer lattice shell structure is shown in Figure 3 The schematic diagram of the segmented single-layer lattice shell structure is shown. Figure 3 The split components are distinguished by color. Figure 4 Shown Figure 3 Of course, the technical solution of this embodiment can also be applied to other steel structures such as space truss structure, space grid structure, etc.

[0040] When clustering point clouds based on spatial proximity and feature similarity for the point cloud data processed in step S1, the SLIC algorithm can be used. The specific steps are as follows: first, uniformly initialize seed points in the point cloud; then, locally optimize each seed point and move it to the place with the smallest gradient in the neighborhood to ensure that the seed point is not located on the contour boundary; then, calculate the distance between each point and the seed point. This distance is a combination of color and spatial distance. Based on this distance, assign a class label to each point and place it in the cluster corresponding to the nearest seed point. This process is iterated until the clustering result is stable, that is, the point allocation no longer changes, and finally a group of supervoxels with similar spatial positions and feature attributes are obtained.

[0041] The calculation of the concavity and convexity between adjacent supervoxels can be implemented by using existing technologies. This embodiment provides an improved judgment rule, that is, the judgment of the concavity and convexity relationship is made by combining the extended convexity criterion (CC) and the rationality criterion (SC), which specifically includes:

[0042] The extended convexity criterion (CC) is used to determine whether two adjacent supervoxels are convexly connected or concavely connected. Specifically, the centroid connecting vector of the two adjacent supervoxels is recorded as , the normal vector is recorded as , , and , The angle between α 1. α 2. If α 1> α 2, it is considered a concave connection if α 1< α 2, it is considered a convex connection;

[0043] The judgment result of the extended convexity criterion (CC) is modified by the rationality criterion (SC), specifically: the cross product of the normal vectors of two adjacent supervoxels is recorded as , , and The angle between v , through the angle v To determine whether these two supervoxels are truly connected; set the angle threshold β Thresh ,if v < β Thresh , then the previous convexity determination is invalid; if v ≥ β Thresh, then the convexity judgment in the previous text is valid; thus, unconnected supervoxels are avoided from being mistakenly identified as convex connections; this criterion helps to distinguish cases of surface discontinuity, such as the presence of a single face or surface singularity, in which cases it is meaningless to judge convexity;

[0044] Ultimately, only connections that satisfy both CC and SC criteria are considered valid concave-convex relations.

[0045] The weighted random sample consensus (RANSAC) algorithm is an algorithm that calculates the mathematical model parameters of the data and obtains valid sample data based on a set of sample data sets containing abnormal data. In this embodiment, the RANSAC algorithm is used to obtain the best segmentation plane. In each iteration, the algorithm randomly selects points with concave and convex information and weights from the EEC (Euclidean edge cloud) as seed points of the segmentation plane, and uses these seed points to fit a segmentation plane based on their three-dimensional coordinates; the algorithm calculates the distance from all points to the current fitting plane, and combines the weights of the points to determine the fit of the plane, where concave points have a greater impact on the model quality due to their higher weights. In multiple iterations, the algorithm retains those segmentation planes with the highest fit and the smallest error as candidate best models, and excludes those outlier points that are too far from the plane. Then, the algorithm converges to a stable best segmentation plane solution, which provides the best data fit in all iterations, conforms to the structural characteristics of the point cloud data, and can accurately segment the best planes of different parts. Finally, the best segmentation plane is used as the input of the local constraint cutting module to guide it to perform more detailed segmentation in a specific direction. Embodiment 2

[0046] Combination Figures 1 to 5 As shown, this embodiment provides an intelligent segmentation system for spatial steel structure nodes, including a point cloud input system, a local feature extraction system, a segmentation position extraction system, and a segmentation and optimization system.

[0047] The point cloud input system includes a point cloud acquisition module and a point cloud preprocessing module. The point cloud acquisition module can collect point cloud data of the actual scene of the steel structure, and specifically can collect point cloud data through a three-dimensional scanning device; the point cloud preprocessing module can remove noise points, downsample voxels and perform smoothing on the point cloud data.

[0048] The local feature extraction system includes a supervoxel clustering module and a concave-convex relationship calculation module. The supervoxel clustering module can cluster point clouds based on spatial proximity and feature similarity of point clouds to form several supervoxels. The concave-convex relationship calculation module is used to calculate the concave-convexity between adjacent supervoxels.

[0049] The segmentation position extraction system includes a Euclidean edge cloud (EEC) building module and a geometric constraint partitioning module. The Euclidean edge cloud building module can abstract the edges between adjacent supervoxels into a point cloud, and assign a weight value a to concave points, where a>0, and a weight value 0 or a negative value to convex points. The geometric constraint partitioning module can use a weighted random sampling consensus algorithm to obtain the optimal segmentation plane.

[0050] The segmentation and optimization system includes a local constraint cutting module and a post-processing module. The local constraint cutting module performs Euclidean segmentation on the concave points near the optimal segmentation plane, controls the segmentation process by limiting the growth upper limit, and ensures the locality and accuracy of the segmentation. The subdomains obtained after segmentation are further segmented according to the directionality of the concave edge, because the directionality of the concave edge provides important geometric information, which helps to identify and retain key features in the point cloud. In this way, the local constraint cutting module can segment more accurately along the natural boundaries of the object, thereby obtaining a segmentation result that is more in line with the actual geometric structure. Ultimately, this segmentation method that combines the information of the optimal segmentation plane and local constraints can achieve point cloud segmentation from coarse to detailed, and improve the accuracy and efficiency of segmentation. The post-processing module optimizes and refines the segmentation results to improve the quality and accuracy of the segmentation, including operations such as smoothing the segmentation boundaries, removing small segmentation areas, merging contact areas, and using morphological operations to improve the segmentation results.

[0051] The intelligent segmentation system for spatial steel structure nodes provided in this embodiment can obtain point cloud data of the actual scene of the steel structure through the point cloud input system, pre-process the point cloud data, and then cluster the point cloud through the local feature extraction system to form a number of supervoxels, and calculate the convexity and concavity between adjacent supervoxels, and obtain the best segmentation plane through the segmentation position extraction system, and segment the supervoxels through the segmentation and optimization system, so as to accurately segment the nodes from the entire steel structure model, so as to facilitate the monitoring of the nodes using the finite element model.

[0052] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described 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.

[0053] The above-mentioned embodiments only express several implementation modes of the present invention, and the description thereof is relatively specific and detailed, but it cannot be understood as limiting the scope of the invention. It should be pointed out that, for a person of ordinary skill in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the attached claims.

Claims

1. An intelligent segmentation method for spatial steel structure nodes, characterized in that: The steps include: S1. Collect point cloud data of the actual scene of the steel structure, and remove noise points, downsample voxels and smooth the point cloud data; S2. clustering the point cloud data processed in step S1 according to spatial proximity and feature similarity to form a number of supervoxels, and calculating the concavity and convexity between adjacent supervoxels; S3. The boundary between adjacent supervoxels is abstracted into a point cloud with convex and concave information, the concave points are assigned a weight value a, a>0, and the convex points are assigned a weight value 0 or a negative value, and the weighted random sampling consensus algorithm is used to obtain the best segmentation plane; S4. segmenting the supervoxel according to the best segmentation plane, and performing optimization operations on the segmentation result by smoothing the segmentation boundary, removing small areas, and merging contact areas; Among them, in step S3, when the weighted random sampling consensus algorithm is used to obtain the best segmentation plane, specifically: points with concave and convex information and weights are selected as seed points of the segmentation plane, and these seed points are used to fit a segmentation plane based on their three-dimensional coordinates; the distances from all points to the current fitting plane are calculated, and the fit of the plane is determined in combination with the weights of the points; after multiple iterations, the segmentation plane with the highest fit and the smallest error is retained as the candidate best model, and those outlier points that are too far from the plane are excluded; the algorithm converges to a stable best segmentation plane solution.

2. The intelligent segmentation method of spatial steel structure nodes according to claim 1, characterized in that: Calculating the concavity and convexity between adjacent supervoxels specifically includes: The centroid vector connecting two adjacent supervoxels is denoted as , the normal vector is recorded as , , and , The angle between α 1. α 2. If α 1> α 2, it is considered a concave connection if α 1< α 2, it is considered a convex connection; The cross product of the normal vectors of two adjacent supervoxels is denoted as , , and The angle between v , set the angle threshold β Thresh ,if v < β Thresh , then the convexity determination is invalid, if v ≥ β Thresh , then the concavity and convexity determination is valid.

3. An intelligent segmentation system for spatial steel structure nodes, characterized in that: Including point cloud input system, local feature extraction system, segmentation position extraction system and segmentation and optimization system; The point cloud input system includes a point cloud acquisition module and a point cloud preprocessing module. The point cloud acquisition module can collect point cloud data of the actual scene of the steel structure, and the point cloud preprocessing module can remove noise points, downsample voxels and perform smoothing on the point cloud data. The local feature extraction system includes a supervoxel clustering module and a concave-convex relationship calculation module. The supervoxel clustering module can cluster point clouds based on spatial proximity and feature similarity of point clouds to form a number of supervoxels. The concave-convex relationship calculation module is used to calculate the concave-convexity between adjacent supervoxels. The segmentation position extraction system includes a Euclidean edge cloud construction module and a geometric constraint partitioning module. The Euclidean edge cloud construction module can abstract the edges between adjacent supervoxels into point clouds, and assign a weight value a, a>0 to concave points, and a weight value 0 or negative to convex points. The geometric constraint partitioning module can use a weighted random sampling consensus algorithm to obtain the best segmentation plane. The segmentation and optimization system includes a local constraint cutting module and a post-processing module. The local constraint cutting module can segment the supervoxel according to the segmentation plane, and the post-processing module can perform optimization operations on the segmentation result such as smoothing the segmentation boundary, removing small areas, and merging contact areas. Among them, when the geometric constraint partitioning module uses the weighted random sampling consensus algorithm to obtain the best segmentation plane, specifically: select points with concave and convex information and weights as seed points of the segmentation plane, and use these seed points to fit a segmentation plane based on their three-dimensional coordinates; calculate the distance from all points to the current fitting plane, and combine the weights of the points to determine the fit of the plane; After multiple iterations, the splitting plane with the highest fit and the smallest error is retained as the candidate best model, and those outlier points that are too far from the plane are excluded; the algorithm converges to a stable optimal splitting plane solution.

Images