Part point cloud hierarchical super voxel segmentation method based on adaptive boundary perception
By adopting an adaptive boundary-aware hierarchical supervoxel segmentation method for component point clouds, and utilizing multi-dimensional geometric features and a global energy optimization framework, the problem of insufficient boundary preservation capability in complex equipment point cloud data is solved, and more stable and accurate supervoxel segmentation is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-05-11
- Publication Date
- 2026-06-05
Smart Images
Figure CN122156502A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of 3D point cloud supervoxel segmentation technology, and in particular to a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary perception. Background Technology
[0002] With the development of 3D scanning technology, 3D point cloud data has been widely used in structural analysis, accuracy assessment, and virtual assembly of complex equipment. Complex equipment typically features large scale differences, highly nested features, and smooth transitions in geometric boundaries. During actual data acquisition, factors such as occlusion, reflection, and viewing angle limitations result in point cloud data with localized missing data and uneven density, leading to a coexistence of weak boundaries and incomplete regions. Under these conditions, supervoxel segmentation, as a crucial step in point cloud preprocessing, structural organization, and feature representation, directly impacts the reliability of subsequent refined analysis.
[0003] Early supervoxel segmentation methods were mainly based on voxelization strategies, discretizing the point cloud space into a regular voxel mesh and combining features such as color, normal vector, curvature, and spatial distance to achieve voxel merging through region growing or clustering algorithms. These methods are simple to implement and computationally efficient, but they are significantly dependent on voxel resolution and feature weight parameters. When point clouds have scale differences or density variations, a fixed resolution struggles to balance local details with overall structural consistency, easily leading to oversegmentation or undersegmentation.
[0004] Subsequent research introduced graph model optimization mechanisms, treating voxels or points as graph nodes, constructing edge weights using adjacency relationships and feature differences, and solving for the optimal segmentation through methods such as graph cut, normalized cut, or energy minimization. These methods enhance the consistency within regions and the overall integrity of the segmentation results to some extent. However, when processing point cloud data with blurred boundaries, the quality of the graph model construction deteriorates, and edge weights based on simple feature metrics struggle to effectively distinguish the regions on either side of the boundary, leading to algorithm performance degradation and insufficient boundary preservation capabilities.
[0005] In recent years, some studies have attempted to introduce deep learning techniques, using neural networks to learn local feature representations to assist in constructing more discriminative region segmentation strategies. While these methods have made progress in some semantic scene segmentation, they rely on large amounts of labeled data for training and incur significant computational overhead, limiting their applicability and generalization capabilities in complex industrial equipment scenarios. Summary of the Invention
[0006] The purpose of this application is to provide a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness, which can effectively maintain object boundaries and improve the stability and accuracy of supervoxel segmentation in complex equipment point cloud scenarios where weak boundaries and incomplete data coexist.
[0007] To achieve the above objectives, this application provides the following solution: A hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness includes the following steps: For the parts to be processed, collect the three-dimensional point cloud data of the parts to be processed.
[0008] The 3D point cloud data is input into the point cloud boundary weight solution algorithm to calculate the multi-dimensional geometric features of each point in the 3D point cloud data and perform weighted fusion and nonlinear transformation on the multi-dimensional geometric features to obtain the point cloud boundary weight model. The point cloud boundary weight model includes the boundary weight value of each point. Points with boundary weight values greater than the boundary point threshold are defined as boundary points.
[0009] An octree spatial partitioning mechanism is used to voxelize the 3D point cloud data, and each voxel is classified into boundary voxels or seed voxels according to the boundary weight value of each point; voxels containing boundary points are boundary voxels, and voxels without boundary points are seed voxels.
[0010] For all voxels obtained by voxelization, all points within each voxel are merged into the set to which the corresponding voxel representative point belongs, generating several initial super voxels; the voxel representative point of the boundary voxel is the point with the largest boundary weight value within the voxel, and the voxel representative point of the seed voxel is the point with the smallest boundary weight value within the voxel.
[0011] Based on the predetermined target supervoxel resolution and the distance metric function including boundary weight difference terms, a hierarchical iterative fusion of representative points of each initial supervoxel is performed using a global energy optimization framework to obtain several supervoxels.
[0012] After boundary assignment correction and spatial connectivity correction of each supervoxel obtained by hierarchical iterative fusion, it is output as the multi-resolution supervoxel segmentation result.
[0013] Optionally, the multi-dimensional geometric features include geometric consistency features, normalized density consistency features, and normalized orientation entropy features; calculating the multi-dimensional geometric features of each point in the 3D point cloud data specifically includes the following steps: The geometric consistency characteristics of each point are calculated based on the average normal angle between each point and its neighboring points.
[0014] Based on the average Euclidean distance between each point and its neighboring points, the density consistency characteristics of each point are calculated and then normalized.
[0015] The directional entropy characteristics of each point are calculated based on the spherical quantization statistical model, and the directional entropy characteristics are then normalized.
[0016] Optionally, a point cloud boundary weight model is obtained by weighted fusion and nonlinear transformation of multi-dimensional geometric features, specifically including the following steps: The weight coefficients corresponding to each geometric feature are adaptively determined based on the variance of each geometric feature.
[0017] The geometric consistency feature, the normalized density consistency feature, and the normalized directional entropy feature are weighted and fused according to the weight coefficients corresponding to the geometric features of each dimension to obtain the fused feature value.
[0018] The fused feature values are subjected to a Sigmoid nonlinear mapping to generate the boundary weight values corresponding to each point, thus obtaining the point cloud boundary weight model.
[0019] Optionally, an octree spatial partitioning mechanism is used to perform voxelization processing on the 3D point cloud data, specifically including the following steps: Based on the number of points contained in a single voxel, calculate the maximum Euclidean distance between each point in the 3D point cloud data and its neighboring points, and take twice the average of the global maximum Euclidean distance as the voxel mesh side length.
[0020] Based on the voxel grid edge length, the 3D point cloud data is divided into voxel grids using an octree space partitioning mechanism to generate several voxels. Each voxel contains a preset number of points that do not overlap. The voxel adjacency relationship of each voxel adopts a 6-adjacency relationship, which means that a single voxel forms a neighborhood with its adjacent voxels in the front, back, left, right, up, and down directions.
[0021] Optionally, based on the boundary weight values of each point, each voxel is classified as a boundary voxel or a seed voxel, specifically including the following steps: Points whose boundary weight values are greater than the boundary point threshold are marked as boundary points.
[0022] Voxels containing boundary points are classified as boundary voxels, and voxels without boundary points are classified as seed voxels.
[0023] For boundary voxels, the point with the largest boundary weight value inside the voxel is selected as the representative voxel point, and for seed voxels, the point with the smallest boundary weight value inside the voxel is selected as the representative voxel point.
[0024] Alternatively, the distance metric function is as follows: .
[0025] in, and The coordinates of two points in the 3D point cloud data of the component to be processed. For point With point Distance metric between and Points With point Boundary weight values, and Points With point The normal vector, For target super-voxel resolution, For adjustment coefficients, To find the absolute value, To find the Euclidean distance between two points.
[0026] Optionally, based on the predetermined target supervoxel resolution and a distance metric function including boundary weight differences, a hierarchical iterative fusion of representative points of each initial supervoxel is performed using a global energy optimization framework to obtain several supervoxels. This specifically includes the following steps: The objective function of the global energy optimization framework is constructed with the goal of minimizing the sum of the distance metrics between the supervoxel interior points and their corresponding representative points.
[0027] The number of target supervoxels is determined based on the predetermined target supervoxel resolution, and the fusion step size parameter is initialized.
[0028] In each iteration, all representative points are traversed in ascending order of boundary weight values to obtain the current representative point and its corresponding neighboring representative points.
[0029] Based on the merging criteria of the global energy optimization framework, it is determined in turn whether the current representative point and each neighboring representative point meet the merging conditions.
[0030] If the fusion conditions are met, the fusion of the supervoxels corresponding to the two representative points is completed, the voxel neighborhood is expanded, and the representative points of the fused supervoxels are updated; the fusion of supervoxels satisfies the supervoxel size constraint strategy; if the fusion conditions are not met, the two representative points are not fused.
[0031] After completing the hierarchical fusion of the current representative point and each neighboring representative point, the next representative point is obtained as the current representative point, and each neighboring representative point corresponding to the current representative point is obtained. Then, the process jumps to the step "Based on the merging criteria of the global energy optimization framework, it is determined in turn whether the current representative point and each neighboring representative point meet the fusion conditions" until a single iteration of all representative points is completed.
[0032] After a single iteration is completed, the fusion step size parameter is dynamically adjusted according to the global energy decrease rate, and the process jumps to the step "In each iteration, all representative points are traversed in ascending order of boundary weight values to obtain the voxel neighborhood of the current representative point and the corresponding neighboring representative points", until the current number of super voxels reaches the target number of super voxels or the global energy decreases by zero in a single iteration, at which point the iteration is terminated and several super voxels are obtained.
[0033] Optionally, updating the representative point of the fused supervoxel specifically involves updating the representative point of the fused supervoxel to the point with the same voxel type as the current representative point and the lowest boundary weight value.
[0034] Optionally, the supervoxel size constraint strategy is as follows: For the supervoxel corresponding to the boundary voxel, a maximum size constraint threshold is set. When the number of points inside the supervoxel reaches the maximum size constraint threshold, the fusion expansion of the supervoxel is stopped.
[0035] For the supervoxel corresponding to the seed voxel, it is allowed to break through the maximum size constraint threshold and continue to fuse the same type of supervoxel in the neighborhood.
[0036] Optionally, after boundary assignment correction and spatial connectivity correction of each supervoxel obtained after hierarchical iterative fusion, it is output as the multi-resolution supervoxel segmentation result, which specifically includes the following steps: Traverse all points in the 3D point cloud data, mark points with different supervoxel labels in the neighborhood as boundary points to be corrected, and reassign the boundary points to be corrected to the supervoxel to which the nearest neighbor representative point belongs, thus completing the boundary assignment correction.
[0037] Based on boundary attribution correction, the spatial connectivity within each supervoxel is detected based on the octree voxel connectivity relationship. For supervoxels with multiple disconnected substructures, they are split into the corresponding number of independent supervoxels to complete spatial connectivity correction and output multi-resolution supervoxel segmentation results.
[0038] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness. In this method, for the acquired 3D point cloud data of the component to be processed, firstly, multi-dimensional geometric features of each point in the point cloud are calculated and weighted fusion and nonlinear transformation are performed to construct a point cloud boundary weight model to accurately define boundary points, effectively improving the reliability of weak boundary region identification; then, point cloud voxelization is completed based on an octree spatial partitioning mechanism, and voxels are divided into boundary voxels and seed voxels according to boundary weight values, adapting to complex industrial scenarios with uneven point cloud density and local incompleteness; then... The disjoint-set data structure is used to merge the points within each voxel into a set of representative voxel points to generate an initial supervoxel, ensuring the regularity of the fusion of the point sets within the voxels. Then, relying on the global energy optimization framework, a hierarchical iterative fusion of the initial supervoxel representative points is carried out by combining the target supervoxel resolution with a distance metric function that includes boundary weight differences, to suppress cross-boundary erroneous merging and improve segmentation stability. Finally, boundary assignment correction and spatial connectivity correction are performed on the fused supervoxels to optimize the continuity and accuracy of the segmentation results, thus outputting multi-resolution supervoxel segmentation results adapted to complex component structures. Attached Figure Description
[0039] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments 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.
[0040] Figure 1 This is a flowchart illustrating a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness, provided as an embodiment of this application.
[0041] Figure 2 This is a flowchart of step S2 in a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness, provided in an embodiment of this application.
[0042] Figure 3 This is a flowchart of step S3 in a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness, provided in an embodiment of this application.
[0043] Figure 4 This is a schematic diagram illustrating the visualization results of a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness, applied to point cloud data of turbine machinery inner cylinder parts, as provided in an embodiment of this application. Detailed Implementation
[0044] 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.
[0045] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.
[0046] This application provides a hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness. In an exemplary embodiment, such as... Figure 1 As shown, it includes the following steps: S1. For the component to be processed, collect its 3D point cloud data. Specifically, in this embodiment, the component to be processed is a turbine machinery inner cylinder part, and its 3D point cloud data is obtained through a 3D scanning device. Of course, the solution provided in this embodiment can also be adapted to the 3D point cloud data of other complex equipment components.
[0047] S2. Input the 3D point cloud data into the point cloud boundary weight solution algorithm, calculate the multi-dimensional geometric features of each point in the 3D point cloud data, and perform weighted fusion and nonlinear transformation on the multi-dimensional geometric features to obtain the point cloud boundary weight model; the point cloud boundary weight model includes the boundary weight value of each point; points with boundary weight values greater than the boundary point threshold are defined as boundary points.
[0048] Specifically, multi-dimensional geometric features include geometric consistency features, normalized density consistency features, and normalized orientation entropy features; step S2, "calculating the multi-dimensional geometric features of each point in the 3D point cloud data," such as... Figure 2 As shown, the specific steps include: S21. Based on the average normal angle between each point and its neighboring points, the geometric consistency characteristics of each point are calculated. In this embodiment, the geometric consistency characteristics are obtained by calculating the average normal angle between the current point and its neighboring points. The average normal angle between neighboring points is obtained as shown in the following formula: .
[0049] in, For point Geometric consistency eigenvalues, For point The normal vector, For point The The recommended value for the normal vector of the neighboring points is... .
[0050] S22. Based on the average Euclidean distance between each point and its neighboring points, the density consistency feature of each point is calculated, and the density consistency feature is normalized. In this embodiment, the density consistency feature is calculated by measuring the average Euclidean distance between the current point and its neighboring points. The average Euclidean distance between neighboring points is obtained as shown in the following formula: .
[0051] in, For point The density consistency characteristic value, For point spatial coordinates, For point The The recommended values for the spatial coordinates of the neighboring points are... k =25.
[0052] To eliminate dimensional differences, the density consistency characteristic is normalized: .
[0053] in, These are the normalized density consistency eigenvalues. and These represent the minimum and maximum values of the density consistency eigenvalues for all points, respectively.
[0054] S23. The directional entropy characteristics of each point are calculated based on the spherical quantization statistical model, and the directional entropy characteristics are normalized.
[0055] In this embodiment, the directional entropy feature is calculated based on a spherical quantization statistical model, specifically including the following steps: S231, Calculate the current point and its... The unit direction vector between neighboring points is shown in the following equation: .
[0056] in, u x , u y , u z They represent the unit direction vectors respectively. u Recommended values for x, y, z coordinates. k =25.
[0057] S232. Convert the unit direction vector to unit spherical coordinates, as shown in the following formula: , .
[0058] in, θ Represents the polar angle. Represents azimuth, ( θ , ) is the form of expression for spherical coordinates.
[0059] S233, Divide the spherical space into Given a discrete block, count the number of neighboring points within each block. And calculate the probability distribution, as shown in the following formula: .
[0060] in, To divide the polar angle of spherical space into a number, Recommended value for the number of divisions of azimuth angle in spherical space , ; S234. Calculate the directional entropy eigenvalue based on the probability distribution, as shown in the following formula: .
[0061] in, For point directional entropy eigenvalues To prevent tiny constants from being zero in logarithmic operations.
[0062] To eliminate the influence of neighborhood size, the directional entropy feature is normalized, as shown in the following equation: .
[0063] in, This represents the normalized directional entropy eigenvalue.
[0064] In step S2, "weighted fusion and nonlinear transformation of multi-dimensional geometric features are performed to obtain a point cloud boundary weight model," as shown below. Figure 2 As shown, the specific steps include: S24. Based on the variance of each dimension's geometric features, adaptively determine the weight coefficients corresponding to each dimension's geometric features. The weight coefficients for each dimension's geometric features are shown in the following formula. , , Determined adaptively by the variance of each dimension of geometric features: .
[0065] in, , , These are the variances of geometric consistency, density consistency, and directional entropy features, respectively.
[0066] S25. Based on the weight coefficients corresponding to the geometric features of each dimension, the geometric consistency feature, the normalized density consistency feature, and the normalized directional entropy feature are weighted and fused to obtain the fused feature value. As shown in the following formula: .
[0067] in, To fuse feature values.
[0068] S26. Apply a Sigmoid nonlinear mapping to the fused feature values to generate boundary weight values for each point, thus obtaining the point cloud boundary weight model. The fused feature values are then weighted using the following formula. Perform a Sigmoid nonlinear mapping: in, For point Boundary weight values, This is the slope control parameter. For the center offset parameter, the recommended value is... , , This indicates that the average value of the input data is taken.
[0069] S3. The octree spatial partitioning mechanism is used to perform voxelization on the 3D point cloud data, and each voxel is classified into boundary voxels or seed voxels according to the boundary weight value of each point; voxels containing boundary points are boundary voxels, and voxels without boundary points are seed voxels.
[0070] Specifically, in this embodiment, step S3, "using an octree spatial partitioning mechanism to perform voxelization processing on the 3D point cloud data," means, for example... Figure 3 As shown, it includes the following steps: S31. Based on the number of points contained in a target within a single voxel, calculate the maximum Euclidean distance between each point in the 3D point cloud data and its neighboring points, and take twice the average of the global maximum Euclidean distances as the voxel mesh side length. In this embodiment, the average number of points contained in a target within a voxel is 4. Calculate the maximum Euclidean distance between each point and its four neighboring points in the point cloud data, and take twice the average of the global maximum Euclidean distances as the voxel mesh side length.
[0071] S32. Based on the voxel grid edge length, the three-dimensional point cloud data is divided into voxel grids using an octree space partitioning mechanism to generate several voxels. Each voxel contains a preset number of points that do not overlap. The voxel adjacency relationship of each voxel adopts a 6-adjacency relationship, where a single voxel forms a neighborhood with adjacent voxels in the front, back, left, right, up, and down directions.
[0072] In step S3, "classify each voxel into a boundary voxel or a seed voxel based on the boundary weight value of each point," as shown in the example... Figure 3 As shown, the specific steps include: S33. Mark points whose boundary weight values are greater than the boundary point threshold as boundary points. In this embodiment, the boundary point threshold is 0.5.
[0073] S34. Voxels containing boundary points are classified as boundary voxels, and voxels without boundary points are classified as seed voxels.
[0074] S35. Select the point with the largest boundary weight value inside the voxel as the representative point of the voxel for the boundary voxel, and select the point with the smallest boundary weight value inside the voxel as the representative point of the seed voxel.
[0075] S4. For all voxels obtained from the voxelization process, merge all points within each voxel into the set to which the corresponding voxel representative point belongs, generating several initial supervoxels. Each initial supervoxel corresponds to the point set of a single voxel. The type of the supervoxel is consistent with the voxel type of the corresponding voxel representative point. That is, if the voxel to which the representative point belongs is a boundary voxel, the resulting initial supervoxel is a boundary supervoxel; if the voxel to which the representative point belongs is a seed voxel, the resulting initial supervoxel is a seed supervoxel.
[0076] S5. Based on the predetermined target supervoxel resolution and the distance metric function including boundary weight difference terms, a hierarchical iterative fusion is performed on the representative points of each initial supervoxel using a global energy optimization framework to obtain several supervoxels. In this embodiment, the distance metric function is as follows: .
[0077] in, and The coordinates of two points in the 3D point cloud data of the component to be processed. For point With point Distance metric between and Points With point Boundary weight values, and Points With point The normal vector, For target super-voxel resolution, For adjustment coefficients, To find the absolute value, To find the Euclidean distance between two points; when point With point When the voxel types are the same, the recommended value is... Conversely, the recommended value is... .
[0078] In this embodiment, step S5 specifically includes the following steps: S51. The objective function of the global energy optimization framework is constructed with the goal of minimizing the sum of distance metrics between points within the supervoxel and their corresponding representative points. In this embodiment, the objective function of the global energy optimization framework is as follows: .
[0079] in, Let be the energy function. If it is a binary indicator variable, then , then represents a point Subordinate to the point The supervoxel is represented by the point; The number of points in the 3D point cloud data; It is a distance metric function; This represents the current number of hypervoxels. For the target number of supervoxels, Indicates supervoxel, This is for the fusion step size parameter.
[0080] S52. Determine the number of target supervoxels based on the predetermined target supervoxel resolution, and initialize the fusion step size parameters.
[0081] First, using an octree spatial mechanism, with the target supervoxel resolution as the grid edge length, the number of grids containing point clouds is counted based on the resulting spatial grids, thus obtaining the target supervoxel count. Then, based on the initial supervoxels, the minimum distance metric between each representative point and its neighboring representative points is calculated, and this distance metric is multiplied by the number of points within the corresponding neighboring voxels. The median of all calculated results is then used as the initial value for the fusion step size parameter.
[0082] S53. In each iteration, all representative points are traversed in ascending order of boundary weight values to obtain the current representative point and its corresponding neighboring representative points.
[0083] S54. Based on the global energy optimization framework, the merging criterion sequentially determines whether the current representative point and each of its neighboring representative points meet the merging conditions. Specifically, the merging criterion formula is: .
[0084] in, The energy decreases when the two representative points merge. To fuse step size parameters, The number of points contained in the neighborhood supervoxel. For the aforementioned distance metric function that includes boundary weight difference terms, As the current representative point, This represents the neighborhood point. When... When the two representative points meet the fusion condition, then the two representative points satisfy the fusion condition.
[0085] If the fusion conditions are met, the fusion of the supervoxels corresponding to the two representative points is completed, and step S55 is executed; if the fusion conditions are not met, the two representative points are not fused.
[0086] If the current representative point and its neighboring representative points are successfully merged, the neighboring voxel is removed from the voxel adjacency relation of the current representative point, and the voxel adjacency relation of the neighboring representative point is obtained. The new voxel adjacency relation is assigned to the current representative point, thereby expanding the voxel neighborhood and continuing the previous fusion process. If the fusion of the current representative point and its neighboring representative point fails, the voxel adjacency relation is retained, and no neighborhood expansion is performed. That is, after two representative points are successfully merged, all point clouds in the supervoxels to which the two representative points belong are merged into one, forming a new supervoxel, and the representative point is updated according to subsequent steps.
[0087] S55. Expand the voxel neighborhood and update the representative points of the fused supervoxels. For the current representative point that has been successfully fused, expand its voxel neighborhood and sort the neighboring representative points: if the current representative point belongs to a boundary voxel, the neighboring representative points are sorted in descending order of boundary weight; if the current representative point belongs to a seed voxel, the neighboring representative points are sorted in ascending order of boundary weight. The fusion of supervoxels in step S55 satisfies the supervoxel size constraint strategy.
[0088] Specifically, updating the representative point of the fused supervoxel involves replacing it with the point of the same voxel type as the current representative point and having the lowest boundary weight value. This is because, for seed supervoxels, the representative point with the lowest boundary weight is selected during the update to ensure that the fusion process keeps the supervoxel as far away from the boundary as possible, thus ensuring the expansion and extension of the internal region. For boundary supervoxels, the representative point with the lowest boundary weight is selected during the update to ensure that the representative point is always distributed in the boundary region and gradually tends towards the flatter region at the edge of the boundary, avoiding cross-boundary segmentation errors during subsequent boundary assignment correction.
[0089] S56. After completing the hierarchical fusion of the current representative point and each neighboring representative point, obtain the next representative point as the current representative point, and obtain each neighboring representative point corresponding to the current representative point. Jump to step S54 until a single iteration of all representative points is completed.
[0090] S57. After a single iteration is completed, the fusion step size parameter is dynamically adjusted according to the global energy decrease rate, and the process jumps to step S53 until the current number of supervoxels reaches the target number of supervoxels or the global energy decreases by zero in a single iteration. Then the iteration is terminated, and a number of supervoxels are obtained.
[0091] In this embodiment, after each iteration, the fusion step size parameter is dynamically adjusted according to the global energy decrease rate, as shown in the following formula: .
[0092] in, and These are the fusion step size parameters for the current and next iterations, respectively. and These represent the energy decrease in the current iteration and the previous iteration, respectively; that is, the total energy decrease due to successful fusion during a single iteration of fusion. It is a tiny constant. To adjust the coefficient, the recommended value is... .
[0093] In an exemplary embodiment, the supervoxel size constraint strategy mentioned in step S55 is specifically as follows: For the supervoxel corresponding to the boundary voxel, a maximum size constraint threshold is set. When the number of points inside the supervoxel reaches the maximum size constraint threshold, the fusion expansion of the supervoxel is stopped.
[0094] For the supervoxel corresponding to the seed voxel, it is allowed to break through the maximum size constraint threshold and continue to fuse the same type of supervoxel in the neighborhood.
[0095] In this embodiment, the maximum size constraint threshold is defined as follows: .
[0096] in, The maximum size constraint threshold for supervoxels. For target super-voxel resolution, This represents the maximum number of points within a single voxel. Let be the side length of the voxel mesh. To adjust the coefficient, the recommended value is... .
[0097] S6. After performing boundary assignment correction and spatial connectivity correction on each supervoxel obtained through hierarchical iterative fusion, the result is output as the multi-resolution supervoxel segmentation result. In this embodiment, step S6 specifically includes the following steps: S61. Traverse all points in the 3D point cloud data. Mark points with different supervoxel labels within their neighborhood as boundary points to be corrected. Reassign these boundary points to the supervoxel belonging to the nearest neighboring representative point, thus completing the boundary assignment correction. Traverse each point in the point cloud and take its... kThe neighborhood step marks points with different supervoxel labels within the neighborhood as boundary points to be corrected, and reassigns them to the supervoxel of the nearest neighbor representative point in terms of distance metric. Simultaneously, points with inconsistent labels within their neighborhood are also marked as boundary points to be corrected, achieving local continuity correction. In this step, the recommended value is... k =10.
[0098] S62. Based on the boundary attribution correction, and based on the octree voxel connectivity relationship, detect the spatial connectivity within each supervoxel. For supervoxels with multiple disconnected substructures, split them into the corresponding number of independent supervoxels, complete the spatial connectivity correction, and output the multi-resolution supervoxel segmentation results.
[0099] This embodiment uses actual collected point cloud data of turbine machinery inner cylinder parts to verify the proposed hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary perception. The resulting visualization results are as follows: Figure 4 As shown, where, Figure 4 Part (a) is the result of multi-resolution super-voxel segmentation. Figure 4 Part (b) is the multi-resolution supervoxel segmentation boundary.
[0100] The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness provided in the above embodiments of this application constructs a boundary weight model by fusing geometric consistency, density consistency, and directional entropy features, and classifies voxel types based on boundary weights. This allows representative point distributions to be preserved in micro-structural regions, enhancing the structural representation capability of weakly boundaryed and incomplete regions. By introducing a distance metric including boundary weight differences within a global energy optimization framework and dynamically adjusting the fusion step size based on the energy decay rate, similar regions are preferentially fused, and cross-boundary erroneous merging is suppressed, improving the boundary preservation capability and stability of supervoxel segmentation. By dynamically updating the representative points of supervoxels during the fusion process and introducing supervoxel size constraints, combined with boundary attribution correction and spatial connectivity correction, a compact multi-resolution supervoxel segmentation result is formed, improving the segmentation accuracy and adaptability for point clouds with complex geometric structures. The solution of this application can adapt to complex industrial scenarios with uneven point cloud density, local missing points, and weak boundaries, effectively suppressing cross-boundary erroneous merging, improving the boundary preservation capability, accuracy, and stability of supervoxel segmentation, solving the problems of over-segmentation and under-segmentation in traditional methods, and enhancing the adaptability to complex equipment component structures.
[0101] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Moreover, the collection, use and processing of the relevant data are carried out in compliance with the relevant data protection laws and policies of the country where the location is located, and with the authorization granted by the owner of the corresponding device.
[0102] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).
[0103] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0104] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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.
[0105] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.
Claims
1. A hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness, characterized in that, include: For the component to be processed, collect the three-dimensional point cloud data of the component to be processed; The three-dimensional point cloud data is input into the point cloud boundary weight solution algorithm to calculate the multi-dimensional geometric features of each point in the three-dimensional point cloud data. The multi-dimensional geometric features are then weighted, fused, and nonlinearly transformed to obtain the point cloud boundary weight model. The point cloud boundary weight model includes the boundary weight value of each point. Points with boundary weight values greater than the boundary point threshold are defined as boundary points. The 3D point cloud data is voxelized using an octree spatial partitioning mechanism, and each voxel is classified into boundary voxels or seed voxels according to the boundary weight value of each point; voxels containing boundary points are boundary voxels, and voxels without boundary points are seed voxels. For all voxels obtained by voxelization, all points within each voxel are merged into the set to which the corresponding voxel representative point belongs, generating several initial super voxels; the voxel representative point of the boundary voxel is the point with the largest boundary weight value within the voxel, and the voxel representative point of the seed voxel is the point with the smallest boundary weight value within the voxel. Based on the predetermined target supervoxel resolution and the distance metric function including the boundary weight difference term, the representative points of each initial supervoxel are hierarchically iteratively fused according to the global energy optimization framework to obtain several supervoxels. After boundary assignment correction and spatial connectivity correction of each supervoxel obtained by hierarchical iterative fusion, it is output as the multi-resolution supervoxel segmentation result.
2. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 1, characterized in that, The multidimensional geometric features include geometric consistency features, normalized density consistency features, and normalized directional entropy features; Calculating the multi-dimensional geometric features of each point in the three-dimensional point cloud data specifically includes: The geometric consistency characteristics of each point are calculated based on the average normal angle between each point and its neighboring points. Based on the average Euclidean distance between each point and its neighboring points, the density consistency characteristics of each point are calculated, and the density consistency characteristics are normalized. The directional entropy features of each point are calculated based on the spherical quantization statistical model, and the directional entropy features are then normalized.
3. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 2, characterized in that, The multi-dimensional geometric features are weighted, fused, and nonlinearly transformed to obtain a point cloud boundary weight model, specifically including: Based on the variance of each dimension of geometric features, the weight coefficients corresponding to each dimension of geometric features are adaptively determined. Based on the weight coefficients corresponding to the geometric features of each dimension, the geometric consistency feature, the normalized density consistency feature, and the normalized directional entropy feature are weighted and fused to obtain the fused feature value. The fused feature values are subjected to a Sigmoid nonlinear mapping to generate boundary weight values for each point, thus obtaining a point cloud boundary weight model.
4. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 1, characterized in that, The 3D point cloud data is voxelized using an octree spatial partitioning mechanism, specifically including: Based on the number of points contained in a single voxel, calculate the maximum Euclidean distance between each point in the 3D point cloud data and its neighboring points, and take twice the average value of the global maximum Euclidean distance as the voxel mesh side length. Based on the voxel grid side length, the three-dimensional point cloud data is divided into voxel grids using an octree space partitioning mechanism to generate several voxels; each voxel includes a preset number of points that do not overlap; the voxel adjacency relationship of each voxel adopts a 6-adjacency relationship, where a single voxel forms a neighborhood with adjacent voxels in the front, back, left, right, up, and down directions.
5. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 1, characterized in that, Based on the boundary weight values of each point, each voxel is classified into boundary voxels or seed voxels, specifically including: Points with boundary weight values greater than the boundary point threshold are marked as boundary points; Voxels containing boundary points are classified as boundary voxels, and voxels without boundary points are classified as seed voxels. For boundary voxels, the point with the largest boundary weight value inside the voxel is selected as the representative voxel point, and for seed voxels, the point with the smallest boundary weight value inside the voxel is selected as the representative voxel point.
6. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 1, characterized in that, The distance metric function is shown in the following equation: ； in, and The coordinates of two points in the 3D point cloud data of the component to be processed. For point With point Distance metric between and Points With point Boundary weight values, and Points With point The normal vector, For target super-voxel resolution, For adjustment coefficients, To find the absolute value, To find the Euclidean distance between two points.
7. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 1, characterized in that, Based on the predetermined target supervoxel resolution and a distance metric function including boundary weight differences, a hierarchical iterative fusion of representative points of each initial supervoxel is performed using a global energy optimization framework to obtain several supervoxels, specifically including: The objective function of the global energy optimization framework is constructed with the goal of minimizing the sum of the distance metrics between the supervoxel in-situ points and their corresponding representative points. The number of target supervoxels is determined based on the predetermined target supervoxel resolution, and the fusion step size parameter is initialized. In each iteration, all representative points are traversed in ascending order of boundary weight values to obtain the current representative point and its corresponding neighboring representative points. Based on the merging criteria of the global energy optimization framework, it is determined in turn whether the current representative point and each neighboring representative point meet the merging conditions. If the fusion conditions are met, the fusion of the supervoxels corresponding to the two representative points is completed, the voxel neighborhood is expanded, and the representative points of the fused supervoxels are updated; the fusion of supervoxels satisfies the supervoxel size constraint strategy; if the fusion conditions are not met, the two representative points are not fused. After completing the hierarchical fusion of the current representative point and each neighboring representative point, the next representative point is obtained as the current representative point, and each neighboring representative point corresponding to the current representative point is obtained. Then, the process jumps to the step "Based on the merging criteria of the global energy optimization framework, it is determined in turn whether the current representative point and each neighboring representative point meet the fusion conditions" until a single iteration of all representative points is completed. After a single iteration is completed, the fusion step size parameter is dynamically adjusted according to the global energy decrease rate, and the process jumps to the step "In each iteration, all representative points are traversed in ascending order of boundary weight values to obtain the voxel neighborhood of the current representative point and the corresponding neighboring representative points" until the current number of super voxels reaches the target number of super voxels or the global energy decreases by zero in a single iteration, at which point the iteration is terminated and several super voxels are obtained.
8. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 7, characterized in that, The specific steps for updating the representative point of the fused supervox are as follows: update the representative point of the fused supervox to the point with the same voxel type as the current representative point and the lowest boundary weight value.
9. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 7, characterized in that, The specific hypervoxel size constraint strategy is as follows: For the supervoxel corresponding to the boundary voxel, a maximum size constraint threshold is set. When the number of points inside the supervoxel reaches the maximum size constraint threshold, the fusion expansion of the supervoxel is stopped. For the supervoxel corresponding to the seed voxel, it is allowed to break through the maximum size constraint threshold and continue to fuse the same type of supervoxel in the neighborhood.
10. The hierarchical supervoxel segmentation method for component point clouds based on adaptive boundary awareness according to claim 1, characterized in that, After boundary assignment correction and spatial connectivity correction of each supervoxel obtained through hierarchical iterative fusion, the result is output as the multi-resolution supervoxel segmentation result, specifically including: Traverse all points in the 3D point cloud data, mark points with different supervoxel labels in the neighborhood as boundary points to be corrected, and reassign the boundary points to be corrected to the supervoxel to which the nearest neighbor representative point belongs, thus completing the boundary assignment correction. Based on boundary attribution correction, the spatial connectivity within each supervoxel is detected based on the octree voxel connectivity relationship. For supervoxels with multiple disconnected substructures, they are split into the corresponding number of independent supervoxels to complete spatial connectivity correction and output multi-resolution supervoxel segmentation results.