Point cloud segmentation method and system based on super voxel clustering

Abstract

The application discloses a point cloud segmentation method and system based on super voxel clustering, and relates to the technical field of point cloud segmentation.The method comprises the following steps: obtaining a target grid model by performing voxel down-sampling and gridding processing on scene point cloud; determining an initial seed point according to the target grid model; dividing the down-sampled voxels into super voxels to obtain a super voxel set according to the similarity between adjacent voxels based on the initial seed point through an iterative mode; constructing an undirected weighted graph according to the adjacency relationship between super voxels, taking the distance between adjacent super voxels in the current undirected weighted graph as an adjacent weight, and taking each adjacent weight as an adjacent weight set; and updating the current super voxel set by clustering similar super voxels according to the minimum adjacent weight in the current undirected weighted graph through an iterative mode until the minimum adjacent weight in the current undirected weighted graph is greater than or equal to a set threshold, so as to obtain a scene point cloud segmentation result.The application reduces the segmentation complexity.

Description

Technical Field

[0001] This invention relates to the field of point cloud segmentation technology, and in particular to a point cloud segmentation method and system based on supervoxel clustering. Background Technology

[0002] Augmented reality (AR) is a technology that merges the virtual and real worlds, involving multiple research fields such as computer vision, computer graphics, and optics. It has wide applications in smart healthcare, industrial manufacturing, assisted assembly, robot path planning, and natural sciences. AR is an interactive system that integrates the virtual and real worlds and achieves virtual-real alignment. Achieving precise alignment between the virtual and real worlds is a fundamental problem that AR systems must solve. Accurate registration is a key technology limiting the development of AR systems. Existing AR applications based on stereo vision perception typically rely on point cloud information to identify target models and estimate target poses to achieve virtual-real alignment.

[0003] The development of depth sensors has provided a solution for virtual-real registration tasks in augmented reality applications based on stereo vision perception. Unlike marker recognition schemes based on 2D images, stereo vision-based virtual-real registration schemes can determine the pose of a given target model in the camera coordinate system based on the features of the target model within a scene point cloud constructed from images returned by the sensor. Due to the computationally intensive nature of point cloud data, target pose estimation algorithms based on stereo vision perception choose to extract instance point clouds from scene point cloud information and estimate the target model pose based on these instance point clouds. Using scene point cloud segmentation algorithms to cluster scene point clouds according to defined feature similarity is a fundamental and crucial task in computer vision, robot perception, remote sensing mapping, and augmented reality, with widespread research and applications. Currently, existing traditional point cloud segmentation algorithms are mostly applied to planar detection and basic primitive fitting in indoor and outdoor scenes, resulting in low computational efficiency. Deep learning-based point cloud segmentation algorithms also suffer from problems such as insufficient real datasets, high dataset annotation costs, and poor training effects on virtual datasets, limiting the application of deep learning algorithms in the augmented reality field.

[0004] Augmented reality applications under stereo vision conditions employ depth cameras to acquire dense scene point clouds. Recognizing and estimating the pose of target models within these dense point clouds suffers from high computational cost and low estimation efficiency. Extracting instance point clouds from the scene point cloud and then identifying and estimating the target pose based on these instance point clouds can reduce computational cost and improve estimation efficiency. However, feature-based Euclidean clustering methods for extracting instance point clouds suffer from problems such as high computational cost and over-segmentation due to point-by-point judgment. Summary of the Invention

[0005] The purpose of this invention is to provide a point cloud segmentation method and system based on supervoxel clustering, which reduces the segmentation complexity.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] This invention discloses a point cloud segmentation method based on supervoxel clustering, comprising:

[0008] By performing voxel downsampling and meshing on the scene point cloud, a target mesh model is obtained, and initial seed points are determined based on the target mesh model.

[0009] Based on the initial seed point, the downsampled voxels are divided into super voxels according to the similarity between neighboring voxels through an iterative method, resulting in a super voxel set;

[0010] Based on the set of supervoxels, an undirected weighted graph is constructed according to the adjacency relationship between the supervoxels. The distance between adjacent supervoxels in the current undirected weighted graph is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set.

[0011] The current set of super voxels is updated iteratively based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold. The current set of super voxels is then used as the scene point cloud segmentation result.

[0012] Optionally, a target mesh model is obtained by voxel downsampling and meshing the scene point cloud, and initial seed points are determined based on the target mesh model, specifically including:

[0013] Voxel downsampling is performed on the scene point cloud according to a first preset step size;

[0014] The scene point cloud is meshed according to the second preset step size, and the downsampled voxels are mapped onto the coordinate system of the meshed scene point cloud to obtain the target mesh model.

[0015] The center of the target mesh model is used as the initial seed point.

[0016] Optionally, based on the initial seed point, the downsampled voxels are iteratively divided into supervoxels according to the similarity between neighboring voxels to obtain a supervoxel set, specifically including:

[0017] To construct a kd-tree for the downsampled voxels, denoted as the first kd-tree;

[0018] For each seed point, query the first kd-tree to find the neighbor voxels of the seed point;

[0019] Calculate the adjacent color difference distance, normalized geometric distance, and normalized Euclidean distance between each adjacent voxel;

[0020] Calculate the characteristic distance between each neighboring voxel based on the adjacent color difference distance, normalized geometric distance, and normalized Euclidean distance between each neighboring voxel;

[0021] The voxel closest to each sub-point is classified into the cluster corresponding to the seed point;

[0022] For each cluster, the cluster's center point is calculated as the corresponding seed point.

[0023] Return to the step of querying the first kd-tree to find the neighboring voxels of the seed point for each seed point, until all voxels are classified into clusters after downsampling, and the classified clusters no longer change. Then, treat each cluster as a supervoxel to obtain a supervoxel set.

[0024] Optionally, based on the set of supervoxels, an undirected weighted graph is constructed according to the adjacency relationships between the supervoxels. The distance between adjacent supervoxels in the current undirected weighted graph is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set, specifically including:

[0025] Construct an undirected weighted graph based on the adjacency relationships between adjacent supervoxels in the supervoxel set;

[0026] Calculate the normalized chromatic aberration distance and normalized geometric distance between each adjacent supervoxel in the supervoxel set;

[0027] Calculate the adjacency fusion feature distance between each adjacent supervoxel based on the normalized chromatic distance and normalized geometric distance between each adjacent supervoxel;

[0028] The adjacency fusion feature distance between each adjacent supervoxel is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set.

[0029] Optionally, the current set of supervoxels is updated iteratively based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold. The current set of supervoxels is then used as the point cloud segmentation result. Specifically, this includes:

[0030] Merge the adjacent super-voids corresponding to the minimum adjacent weight in the current undirected weighted graph to obtain the constructed super-void. Remove the adjacent super-voids corresponding to the minimum adjacent weight from the current super-void set, add the current constructed super-void to the current super-void set, and update the current undirected weighted graph according to the current super-void set.

[0031] Iterate through the steps of merging the neighboring supervoxels corresponding to the minimum neighbor weight in the current undirected weighted graph until the minimum neighbor weight in the current undirected weighted graph is greater than or equal to a set threshold, and then use the current set of supervoxels as the point cloud segmentation result.

[0032] This invention also discloses a point cloud segmentation system based on supervoxel clustering, comprising:

[0033] The downsampling and meshing module is used to obtain a target mesh model by performing voxel downsampling and meshing on the scene point cloud, and to determine the initial seed points based on the target mesh model.

[0034] The supervoxel partitioning module is used to partition downsampled voxels into supervoxels based on the initial seed point and the similarity between neighboring voxels in an iterative manner, thereby obtaining a supervoxel set.

[0035] An undirected weighted graph construction module is used to construct an undirected weighted graph based on the set of supervoxels and according to the adjacency relationship between the supervoxels. The distance between adjacent supervoxels in the current undirected weighted graph is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set.

[0036] The scene point cloud segmentation module updates the current set of super voxels iteratively based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold, and then uses the current set of super voxels as the scene point cloud segmentation result.

[0037] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0038] This invention achieves hierarchical segmentation of scene point clouds by performing voxel downsampling and meshing, thereby reducing segmentation complexity. Based on the similarity between adjacent voxels, the downsampled voxels are divided into super voxels to obtain a set of super voxels. The instance affiliation is determined based on the adjacency relationship between super voxels, further reducing segmentation complexity and improving computation speed. Attached Figure Description

[0039] To more clearly illustrate the technical solutions in the embodiments of the present invention 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 the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0040] Figure 1 A schematic diagram of a point cloud segmentation method based on supervoxel clustering provided in an embodiment of the present invention;

[0041] Figure 2 A schematic diagram illustrating the specific process of a point cloud segmentation method based on supervoxel clustering provided in an embodiment of the present invention;

[0042] Figure 3This is a schematic diagram of a point cloud segmentation system based on supervoxel clustering, provided as an embodiment of the present invention. Detailed Implementation

[0043] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0044] The purpose of this invention is to provide a point cloud segmentation method and system based on supervoxel clustering, which reduces the segmentation complexity.

[0045] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] This invention constructs a point cloud of an assembly site scene based on RGB-D images returned by a depth camera. For the task of segmenting instance point clouds in the scene point cloud, it implements a scene instance point cloud segmentation method based on supervoxel clustering, which provides scene understanding information for subsequent identification and estimation of target model pose based on instance point clouds.

[0047] The basic idea and workflow of this invention are as follows: Figure 2 As shown, the entire process is divided into two stages: scene point cloud hypervoxarization and hypervoxar clustering.

[0048] In the scene point cloud super-voxelization stage, the scene point cloud after voxel downsampling is meshed. The scene point closest to the center of the mesh is used as the seed point of the initial cluster. The distance from the adjacent scene points to the center of each cluster is calculated iteratively. The scene point closest to the center of each cluster is selected and merged into the corresponding cluster. The seed point of the cluster is recalculated and a new round of iteration is carried out until all points are classified stably.

[0049] During the supervoxel clustering stage, after calculating the distance between adjacent supervoxels based on a custom distance metric, the adjacent supervoxels corresponding to the minimum distance are iteratively merged until the minimum distance is greater than the threshold. The merged result is then output as the instance point cloud segmentation result.

[0050] Example 1

[0051] like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a point cloud segmentation method based on supervoxel clustering, which specifically includes the following steps.

[0052] Step 101: Obtain the target mesh model by voxel downsampling and meshing the scene point cloud, and determine the initial seed points based on the target mesh model.

[0053] Step 102: Based on the initial seed point, the downsampled voxels are divided into supervoxels according to the similarity between neighboring voxels through iteration, and a supervoxel set is obtained.

[0054] Step 103: Based on the set of supervoxels, construct an undirected weighted graph according to the adjacency relationship between the supervoxels. Use the distance between adjacent supervoxels in the current undirected weighted graph as the adjacency weight, and each adjacency weight constitutes an adjacency weight set.

[0055] Step 104: Iteratively update the current supervoxel set based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold, and then use the current supervoxel set as the scene point cloud segmentation result.

[0056] Step 101 specifically includes:

[0057] Construct scene point clouds based on RGB-D images returned by a depth camera.

[0058] The scene point cloud is downsampled using voxels according to a first preset step size. The first preset step size is 0.005d. M d M This is the diameter of the target mesh model. This step uses 0.005d. M A voxel mesh is established, and the downsampled scene point cloud is mapped to a voxel coordinate system. Voxels containing scene points are considered effective voxels. The centroids of each voxel containing scene points in the voxel model composed of effective voxels are calculated as the voxelization result, and the voxel centers replace the local point cloud. A voxel is a downsampled unit. A voxel divided with a custom step size may contain multiple points. The centroid of the downsampled voxel is calculated based on all data points within the voxel and used as the voxelization result.

[0059] The scene point cloud is meshed according to a second preset step size, and the downsampled voxels are mapped onto the coordinate system of the meshed scene point cloud to obtain the target mesh model. The second preset step size is 0.05d. M This step uses 0.05d M The scene point cloud is meshed with a step size, and the camera position is used as the origin of the mesh coordinate system. The voxelization result is then mapped to the mesh coordinate system.

[0060] The center of the target mesh model is used as the initial seed point.

[0061] Step 102 specifically includes:

[0062] A kd-tree is constructed for the downsampled voxels, denoted as the first kd-tree. The first kd-tree structures the downsampled scene point cloud and is used to query the adjacency relationships between voxels.

[0063] For each seed point, query the first kd-tree to find the neighbor voxels of the seed point.

[0064] Calculate the adjacent color difference distance, normalized geometric distance, and normalized Euclidean distance between each adjacent voxel, specifically including:

[0065] The formula for calculating the adjacent color difference distance between adjacent voxels is expressed as follows:

[0066]

[0067] Among them, D c (p i ,p j ) represents the adjacent color difference distance between voxels i and j, where voxels i and j are adjacent voxels, p i p j For any adjacent voxel center, p i p is the center of voxel i. j The center of voxel j; ΔE 00 (p i ,p j () represents the CIEDE2000 standard color difference value; The color difference range for CIEDE2000; D c (p i ,p j ) is the normalized standard chromatic aberration distance (adjacent chromatic aberration distance) between adjacent voxels.

[0068] The formula for calculating the normalized geometric distance between adjacent voxels is as follows:

[0069] D n (n i ,n j )=1-|cos(n i ,n j )|;

[0070] Among them, D n (n i ,n j ) represents the normalized geometric distance between voxels i and j, n i n j Let n be the normal vector corresponding to the adjacent voxels. i Let n be the normal vector corresponding to voxel i. j Let be the normal vector corresponding to voxel j; the geometric similarity between adjacent voxels is calculated as the similarity of the normal vectors between adjacent voxels.

[0071] The formula for calculating the normalized Euclidean distance between adjacent voxels is as follows:

[0072]

[0073] Among them, D s (p i ,p j ) represents the normalized Euclidean distance between voxels i and j, p i p j Let r be any adjacent voxel center; normalize the distance between adjacent voxels using the maximum distance between adjacent mesh centers, r s For the gridding step size, r s Set to 0.05d M .

[0074] The characteristic distance between each neighboring voxel is calculated based on the adjacent color difference distance, normalized geometric distance, and normalized Euclidean distance between each neighboring voxel.

[0075] The formula for calculating the feature distance is:

[0076]

[0077] Among them, D1(p i ,p j D represents the characteristic distance between voxels i and j. c =D c (p i ,p j ), D n =D n (n i ,n j ), D s =D s (p i ,p j ), p i p j For any adjacent voxel center.

[0078] The voxel closest to each sub-point is classified into the cluster corresponding to the seed point.

[0079] For each cluster, the center point of the cluster is calculated as the seed point p corresponding to the cluster.

[0080]

[0081] Where n is the amount of data in the cluster, the Lab (color model) features, spatial distance, and mean normal vector of n points in the cluster are calculated as seed point features, where (x i ,y i ,z i) represents the coordinates of a point, L i a i b i These are the L, a, and b values ​​of the point in Lab space, respectively.

[0082] Return to the step of querying the first kd-tree to find the neighboring voxels of the seed point for each seed point, until all voxels are classified into clusters after downsampling, and the classified clusters no longer change. Then, treat each cluster as a supervoxel to obtain a supervoxel set.

[0083] Step 103 specifically includes:

[0084] The set of supervoxels is structured using a kd-tree to obtain a second kd-tree, which is used for adjacent supervoxel lookup.

[0085] An undirected weighted graph is constructed based on the adjacency relationships between adjacent supervoxels in the supervoxel set.

[0086] Calculating the normalized chromatic aberration distance and normalized geometric distance between adjacent supervoxels in the supervoxel set specifically includes:

[0087] 1) For any supervoxel r p Calculate its relationship with adjacent supervoxels r q Normalized chromatic aberration distance between:

[0088]

[0089] In the formula, D e (p p ,p q ) represents supervoxel r p and supervoxel r q Normalized chromatic aberration distance between, supervoxel r p and supervoxel r q For adjacent supervoxels, p p p q r p r q Corresponding supervoxel center; ΔE 00 The color difference value is the CIEDE2000 standard value. This refers to the color difference range of CIEDE2000.

[0090] 2) Determine r p r q Let n be the adjacency state between them. i n j r p r q Corresponding supervoxel center p p p q The corresponding normal vector; c ijFor p p Point to p q The unit vector; when n i ·c ij <n j ·c ij When to determine r p and r q concave adjacency, determining the adjacency of r p r q Belonging to different instances; when n i ·c ij ≥n j ·c ij When to determine r p r q Convex adjacency, r is considered p r q Belonging to the same instance;

[0091] 3) According to r p r q Adjacency state calculation r p r q Normalized geometric distance between:

[0092]

[0093] Among them, D r (p p ,p q ) represents supervoxel r p and supervoxel r q Normalized geometric distance between them, r p r q When concave adjacency is determined, r is considered p r q They belong to different instances, and their geometric distance is defined based on their geometric relationship; r p r q When convex adjacency is defined, r is determined. p r q If they are the same instance, reduce the geometric distance between them.

[0094] Based on the normalized chromatic aberration distance and normalized geometric distance between each adjacent supervoxel, the adjacency fusion feature distance between each adjacent supervoxel is calculated, specifically including:

[0095]

[0096] Wherein, D2(r p ,r q ) represents supervoxel r p and supervoxel r q The adjacency fusion feature distance between the two features, where λ is the feature fusion weight, and D is the feature fusion distance between the two features. e =De (p p ,p q ), D r =D r (p p ,p q Let the mean of the statistical distribution of color and geometric distance between all adjacent supervoxels be μ. c μ g Let λ = μ g / (μ c +μ g ).

[0097] The adjacency fusion feature distance between each adjacent supervoxel is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set.

[0098] Query the second kd-tree, calculate the fusion feature distances between all neighboring voxels, and form the distance set W. k That is, the set of adjacency weights.

[0099] Construct an undirected weighted graph G based on the adjacency relationships between adjacent supervoxels. k ={R k W k The adjacency weight is calculated as the fusion feature distance between adjacent supervoxels.

[0100] Step 104 specifically includes:

[0101] The process involves merging the adjacent supervoids corresponding to the minimum adjacent weight in the current undirected weighted graph to obtain the constructed supervoid. The adjacent supervoids corresponding to the minimum adjacent weight are then removed from the current supervoid set, and the constructed supervoid is added to the current supervoid set. Finally, the current undirected weighted graph is updated based on the current supervoid set. Specifically, this includes:

[0102] The adjacency weight set W k The adjacent elements are sorted from smallest to largest by their respective weights.

[0103] Let W k Minimum weighted distance w min w min Corresponding adjacent supervoxels merge Constructing new supervoxels:

[0104]

[0105] In the formula: For the merging of supervoxels (construction supervoxels).

[0106] Remove the minimum weighted distance w from the supervoxel set. min Add the corresponding adjacent supervoxels to the corresponding construction set and update the supervoxel set:

[0107] In the formula: R k-1 For the original set R k The updated set of supervoxels is obtained by removing the adjacent supervoxels corresponding to the minimum weighted distance and then adding them to the merged supervoxels.

[0108] Calculate the adjacency distance (adjacency fusion feature distance) between the constructed supervoxel and other supervoxels:

[0109]

[0110] In the formula: for The adjacency distance between adjacent supervoxels. Indicates adjacent supervoxels The distance between adjacent fusion features.

[0111] Update the distance set and the undirected weighted graph:

[0112] G k-1 ={R k-1 W k-1}

[0113] Among them, G k-1 Let W represent the updated undirected weighted graph. k-1 Let R represent the updated set of adjacency weights. k-1 This represents the updated set of supervoxels.

[0114] The process of merging the adjacent supervoxels corresponding to the minimum adjacent weight in the current undirected weighted graph is iterated until the minimum adjacent weight in the current undirected weighted graph is greater than or equal to a set threshold ξ(w) is returned. min ≥ξ), the current set of supervoxels is used as the point cloud segmentation result.

[0115] The beneficial effects of this invention are as follows:

[0116] This invention addresses the problems of large amounts of scene point cloud data acquired in augmented reality applications, long running time, and low segmentation efficiency of existing point cloud segmentation methods. It employs a hierarchical segmentation method to reduce the segmentation complexity by downsampling the scene point cloud. By using geometric and color features of scene points, it clusters scene point clouds with similar local features into supervoxels. The instance affiliation is determined based on the adjacency relationship between supervoxels. Instance point cloud extraction is achieved by using geometric and color features to cluster similar supervoxels. This simplifies the scene point cloud instance segmentation problem, improves segmentation accuracy, shortens estimation time, and increases computational efficiency.

[0117] Example 2

[0118] like Figure 3As shown, this embodiment provides a point cloud segmentation system based on supervoxel clustering, which specifically includes:

[0119] The downsampling and meshing module 201 is used to obtain a target mesh model by performing voxel downsampling and meshing on the scene point cloud, and to determine the initial seed points based on the target mesh model.

[0120] The supervoxel partitioning module 202 is used to partition downsampled voxels into supervoxels based on the initial seed point and the similarity between adjacent voxels in an iterative manner, thereby obtaining a supervoxel set.

[0121] The undirected weighted graph construction module 203 is used to construct an undirected weighted graph based on the set of supervoxels and according to the adjacency relationship between the supervoxels. The distance between adjacent supervoxels in the current undirected weighted graph is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set.

[0122] The scene point cloud segmentation module 204 is used to update the current super voxel set by iteratively based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold, and then use the current super voxel set as the scene point cloud segmentation result.

[0123] Example 3

[0124] This embodiment provides an electronic device, including a memory and a processor. The memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the point cloud segmentation method based on supervoxel clustering according to Embodiment 1.

[0125] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the point cloud segmentation method based on supervoxel clustering as described in Embodiment 1.

[0126] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0127] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, 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 the present invention.

Claims

1. A point cloud segmentation method based on supervoxel clustering, characterized in that, include: By performing voxel downsampling and meshing on the scene point cloud, a target mesh model is obtained, and initial seed points are determined based on the target mesh model. Based on the initial seed point, the downsampled voxels are divided into super voxels according to the similarity between neighboring voxels through an iterative method, resulting in a super voxel set; Based on the set of supervoxels, an undirected weighted graph is constructed according to the adjacency relationship between the supervoxels. The distance between adjacent supervoxels in the current undirected weighted graph is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set. The current set of super voxels is updated iteratively based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold. The current set of super voxels is then used as the scene point cloud segmentation result. Specifically, based on the set of supervoxels, an undirected weighted graph is constructed according to the adjacency relationships between supervoxels. The distance between adjacent supervoxels in the current undirected weighted graph is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set. This process includes: constructing an undirected weighted graph based on the adjacency relationships between adjacent supervoxels in the set of supervoxels; calculating the normalized color difference distance and normalized geometric distance between adjacent supervoxels in the set of supervoxels; calculating the adjacency fusion feature distance between adjacent supervoxels based on the normalized color difference distance and normalized geometric distance between adjacent supervoxels; and using the adjacency fusion feature distance between adjacent supervoxels as the adjacency weight, with each adjacency weight constituting an adjacency weight set. The process iteratively updates the current set of supervoxels based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold. The current set of supervoxels is then used as the point cloud segmentation result. Specifically, this involves: merging the adjacent supervoxels corresponding to the minimum adjacency weight in the current undirected weighted graph to obtain a constructed supervoxel; removing the adjacent supervoxel corresponding to the minimum adjacency weight from the current set of supervoxels; adding the constructed supervoxel to the current set of supervoxels; and updating the current undirected weighted graph based on the current set of supervoxels. The process then iterates back to the step of merging the adjacent supervoxels corresponding to the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold. The current set of supervoxels is then used as the point cloud segmentation result.

2. The point cloud segmentation method based on supervoxel clustering according to claim 1, characterized in that, By performing voxel downsampling and meshing on the scene point cloud, a target mesh model is obtained. Initial seed points are then determined based on the target mesh model, specifically including: Voxel downsampling is performed on the scene point cloud according to a first preset step size; The scene point cloud is meshed according to the second preset step size, and the downsampled voxels are mapped onto the coordinate system of the meshed scene point cloud to obtain the target mesh model. The center of the target mesh model is used as the initial seed point.

3. The point cloud segmentation method based on supervoxel clustering according to claim 1, characterized in that, Based on the initial seed point, the downsampled voxels are iteratively divided into supervoxels according to the similarity between neighboring voxels, resulting in a supervoxel set, specifically including: To construct a kd-tree for the downsampled voxels, denoted as the first kd-tree; For each seed point, query the first kd-tree to find the neighbor voxels of the seed point; Calculate the adjacent color difference distance, normalized geometric distance, and normalized Euclidean distance between each adjacent voxel; Calculate the characteristic distance between each neighboring voxel based on the adjacent color difference distance, normalized geometric distance, and normalized Euclidean distance between each neighboring voxel; The voxel closest to each sub-point is classified into the cluster corresponding to the seed point; For each cluster, the cluster's center point is calculated as the corresponding seed point. Return to the step of querying the first kd-tree to find the neighboring voxels of the seed point for each seed point, until all voxels are classified into clusters after downsampling, and the classified clusters no longer change. Then, treat each cluster as a supervoxel to obtain a supervoxel set.

4. A point cloud segmentation system based on supervoxel clustering, characterized in that, The point cloud segmentation system based on supervoxel clustering applies the point cloud segmentation method based on supervoxel clustering as described in claim 1, wherein the point cloud segmentation system based on supervoxel clustering includes: The downsampling and meshing module is used to obtain a target mesh model by performing voxel downsampling and meshing on the scene point cloud, and to determine the initial seed points based on the target mesh model. The supervoxel partitioning module is used to partition downsampled voxels into supervoxels based on the initial seed point and the similarity between neighboring voxels in an iterative manner, thereby obtaining a supervoxel set. An undirected weighted graph construction module is used to construct an undirected weighted graph based on the set of supervoxels and according to the adjacency relationship between the supervoxels. The distance between adjacent supervoxels in the current undirected weighted graph is used as the adjacency weight, and each adjacency weight constitutes an adjacency weight set. The scene point cloud segmentation module updates the current set of super voxels iteratively based on the minimum adjacency weight in the current undirected weighted graph until the minimum adjacency weight in the current undirected weighted graph is greater than or equal to a set threshold, and then uses the current set of super voxels as the scene point cloud segmentation result.

Images