An unsupervised point cloud segmentation method and system

Abstract

The application relates to an unsupervised point cloud segmentation method and system, and relates to the technical field of target segmentation, and comprises the following steps: S1, acquiring point cloud data, estimating a neighborhood radius, and making the ratio of a first target eigenvalue and a second target eigenvalue not less than a preset value when a directionless normal of each point is extracted based on PCA; S2, acquiring a local representative subset through range search based on the estimated neighborhood radius; estimating a local tangent plane of each point in the local representative subset based on PCA, and extracting a directionless normal of each point based on the local tangent plane of each point; and S3, performing plane segmentation based on the directionless normal of each point. The neighborhood radius is estimated by calculating eigenvalues based on PCA analysis, so that the plane information in original data is fully utilized when the normal of each point is estimated. The unsupervised system does not need training data or manual intervention, and can obtain a point cloud segmentation result in a short time and at a low calculation cost.

Description

Technical Field

[0001] The present application relates to the field of target segmentation technology, and in particular to an unsupervised point cloud segmentation method and system. Background Art

[0002] Point cloud segmentation is primarily based on rigorous handcrafted features derived from geometric constraints, heuristics, and rules. It aims to group raw 3D points into distinct regions corresponding to specific structures or objects that are representative of a scene. To reduce computational costs before classification, a common strategy is to segment the raw point cloud into small regions. Four approaches can be categorized. The first is edge-based segmentation, which detects the boundaries of certain regions of the point cloud by locating intensity changes at edge points. While this method is fast, it is highly sensitive to noise and cannot guarantee accuracy. The second is graph-based segmentation, which treats the point cloud as a graph where each point represents a node and the connections between points represent their relationships. Graph-based algorithms are used to segment the point cloud into distinct components. While this method offers high segmentation accuracy, it is less efficient. The third is deep learning-based segmentation, which uses a large amount of labeled point cloud data to train a deep learning model to segment unlabeled point cloud data. While this method offers excellent accuracy, it requires a large amount of labeled data and computational resources. The fourth segmentation method is based on region growing. By selecting seed points and using the normal vector angle or curvature as a threshold, region growing is performed until the preset conditions are met. This method maintains good boundary integrity and regional connectivity, resulting in high segmentation efficiency, and the algorithm's performance depends on seed selection. All four of these methods require a labeled dataset based on a specific domain, which increases costs. Reducing the cost of point cloud segmentation without relying on labeled datasets is a pressing technical challenge for those skilled in the art. Summary of the Invention

[0003] (1) Purpose of application

[0004] In view of this, the purpose of this application is to provide an unsupervised point cloud segmentation method and system to solve the technical problem of how to perform point cloud segmentation without a labeled dataset and reduce costs.

[0005] (2) Technical solution

[0006] This application discloses an unsupervised point cloud segmentation method, comprising the following steps:

[0007] S1. Obtain point cloud data and estimate the domain radius so that when extracting the undirected normal of each point based on PCA, the ratio of the first target eigenvalue to the second target eigenvalue is not less than a preset value;

[0008] S2. Based on the estimated neighborhood radius, obtain a local representative subset through range search; estimate the local tangent plane of each point in the local representative subset based on PCA, and extract the undirected normal of each point based on the local tangent plane of each point;

[0009] S3. Perform plane segmentation based on the undirected normal of each point.

[0010] In a possible implementation, the first target eigenvalue is the minimum eigenvalue among the eigenvalues. , the second target eigenvalue is the second largest eigenvalue among the eigenvalues The ratio of the first target characteristic value to the second target characteristic value is not less than the preset value .

[0011] In a possible implementation, estimating the domain radius includes: finding a domain radius range such that each point , sets the estimated domain radius to the median of the range of domain radii.

[0012] In one possible embodiment, the plane segmentation based on the undirected normal of each point includes: starting from a single seed point by seed growth, by judging whether the adjacent points are in the plane P, the qualified points are added to the region, and the index of the added point is recorded at the same time, the plane P is defined by the center of mass c and normal n of the region, the center of mass c and normal n of the plane P are initialized based on the undirected normal and the position of the seed point, the qualified point is a point whose normal has an angle difference with the normal n of the plane P that is not greater than α and whose position is at most 3d from the local tangent plane to the region, and the size of d is equal to the radius of the determined domain.

[0013] In one possible embodiment, α is a region of size τ created on the surface of a sphere with a radius of 2ϵ, so that all its points fit into a cone with an opening angle of α, the size of ϵ is equal to the radius of the determined domain, and τ is the threshold of the minimum number of points required for the region formed based on seed growth.

[0014] As a second aspect of the present application, an unsupervised point cloud segmentation system is also disclosed, comprising:

[0015] A domain radius determination module is used to obtain point cloud data and estimate the domain radius so that when extracting the undirected normal of each point based on PCA, the ratio of the first target eigenvalue to the second target eigenvalue is not less than a preset value;

[0016] The point normal extraction module is used to obtain a local representative subset through range search based on the estimated neighborhood radius; the local tangent plane of each point in the local representative subset is estimated based on PCA, and the undirected normal of each point is extracted based on the local tangent plane of each point;

[0017] The plane segmentation module is used to perform plane segmentation based on the undirected normal of each point.

[0018] In a possible implementation, the first target eigenvalue is the minimum eigenvalue among the eigenvalues. , the second target eigenvalue is the second largest eigenvalue among the eigenvalues The ratio of the first target characteristic value to the second target characteristic value is not less than the preset value .

[0019] In a possible implementation, estimating the domain radius includes: finding a domain radius range such that each point , sets the estimated domain radius to the median of the range of domain radii.

[0020] In one possible embodiment, the plane segmentation based on the undirected normal of each point includes: starting from a single seed point by seed growth, by judging whether the adjacent points are in the plane P, the qualified points are added to the region, and the index of the added point is recorded at the same time, the plane P is defined by the center of mass c and normal n of the region, the center of mass c and normal n of the plane P are initialized based on the undirected normal and the position of the seed point, the qualified point is a point whose normal has an angle difference with the normal n of the plane P that is not greater than α and whose position is at most 3d from the local tangent plane to the region, and the size of d is equal to the radius of the determined domain.

[0021] In one possible embodiment, the plane segmentation based on the undirected normal of each point includes: starting from a single seed point by seed growth, by judging whether the adjacent points are in the plane P, the qualified points are added to the region, and the index of the added point is recorded at the same time, the plane P is defined by the center of mass c and normal n of the region, the center of mass c and normal n of the plane P are initialized based on the undirected normal and the position of the seed point, the qualified point is a point whose normal has an angle difference with the normal n of the plane P that is not greater than α and whose position is at most 3d from the local tangent plane to the region, and the size of d is equal to the radius of the determined domain.

[0022] In one possible embodiment, α is a region of size τ created on the surface of a sphere with a radius of 2ϵ, so that all its points fit into a cone with an opening angle of α, the size of ϵ is equal to the radius of the determined domain, and τ is the threshold of the minimum number of points required for the region formed based on seed growth.

[0023] (3) Beneficial effects

[0024] By calculating eigenvalues ​​based on PCA analysis to estimate the neighborhood radius, the normal estimation for each point fully utilizes the planar information in the original data. This unsupervised system, which requires no training data or manual intervention, can achieve point cloud segmentation results in a short time and at a low computational cost.

[0025] Other advantages, objectives, and features of the present application will be described in detail in the following description and, to some extent, will be apparent to those skilled in the art upon examination and study of the following, or may be taught from practice of the present application. The objectives and other advantages of the present application may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain and illustrate the present application, but should not be construed as limiting the scope of protection of the present application.

[0027] Figure 1 is the system flow chart of this application;

[0028] Figure 2 This is a comparison chart of the actual surface normal and the estimated normal based on PCA in this application;

[0029] Figure 3 This is a schematic diagram of the growth of the application area;

[0030] Figure 4 This is the system structure diagram of this application.

[0031] Among them, 1. Field radius determination module; 2. Point normal extraction module; 3. Plane segmentation module. DETAILED DESCRIPTION

[0032] To make the objectives, technical solutions, and advantages of the embodiments of the present application more clear, the technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Generally, the components of the embodiments of the present application described and shown in the drawings herein can be arranged and designed in various different configurations.

[0033] Therefore, the following detailed description of the embodiments of the present application provided in the accompanying drawings is not intended to limit the scope of the present application for protection, but merely represents selected embodiments of the present application. All other embodiments obtained by persons of ordinary skill in the art based on the embodiments in the present application without creative work are within the scope of protection of the present application.

[0034] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, it does not need to be further defined or explained in subsequent drawings.

[0035] In the above description of the present application, the terms “first”, “second”, etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0036] like Figure 1 As shown, this embodiment provides an unsupervised point cloud segmentation method, comprising the following steps:

[0037] S1. Obtain point cloud data and estimate the domain radius so that when extracting the undirected normal of each point based on PCA, the ratio of the first target eigenvalue to the second target eigenvalue is not less than a preset value.

[0038] A K-dimensional tree spatial structure based on point cloud data is constructed to facilitate fast nearest neighbor searches for each point, and a range search is used to obtain a local representative subset of the point cloud. To balance runtime and extracting the unoriented normal for each point during normal estimation, an appropriate neighborhood radius is required. Assuming uniform sampling of points in a 10×10×0.01 unit volume, this serves as a probabilistic noise model for the point cloud, which has local normal directions. Experiments on runtime and average normal deviation using different neighborhood radii show that the size of the neighborhood radius directly affects the quality of the normals and the runtime of normal estimation. Smaller values ​​result in shorter runtimes, but when the radius approaches the noise level of the point cloud, the resulting normals are affected by noise and fail to capture any useful information. Conversely, larger values ​​not only significantly increase the runtime of normal estimation but also cause normals near prominent features in the point cloud to be smoothed out at larger radii, reducing the quality of the normal estimation.

[0039] In some embodiments, the first target eigenvalue is the minimum eigenvalue among the eigenvalues , the second target eigenvalue is the second largest eigenvalue among the eigenvalues The ratio of the first target characteristic value to the second target characteristic value is not less than the preset value The estimation of the field radius includes: finding a field radius range so that each point , set the estimated domain radius to the median of the range of domain radius. The domain radius determined based on the eigenvalue needs to be much higher than the radius of the noise in the point cloud data, because PCA is used to calculate the normal of the data point later. After the feature analysis of the point neighborhood, the eigenvector corresponding to the minimum eigenvalue becomes the directionless normal of the point. Due to the characteristics of PCA itself, the second largest eigenvalue is used here. and the minimum eigenvalue The ratio between them is used to measure the planarity of the neighborhood, thereby measuring the reliability of the normal estimation. In order to estimate the neighborhood radius, it is necessary to find the optimal neighborhood radius. First, find a radius that makes each point The neighborhood radius range is then set to the median of the range. This median is the estimated neighborhood radius, because it is usually assumed that about 50% of the points in the point cloud are located on the plane area. It mainly affects the grouping of small areas with their larger corresponding areas, resulting in fewer segments, but may lose the finer decomposition of low-variance planar areas.

[0040] S2. Based on the estimated neighborhood radius, a local representative subset is obtained through range search; the local tangent plane of each point in the local representative subset is estimated based on PCA, and the undirected normal of each point is extracted based on the local tangent plane of each point.

[0041] Principal Component Analysis (PCA) is used to estimate the local tangent plane for each point and extract the unoriented normal of that point. The main limitation of using PCA for normal estimation is that the method only uses a single tangent plane to approximate the local neighborhood of each point in the point cloud. Although it can provide high-quality normals for points far away from prominent features, any point within a selected radius around an edge or corner will be assigned a biased normal because two or more planes of the original geometry will be approximated by a single plane, such as Figure 2 Although the PCA method will lead to a certain decline in the quality of normal estimation, compared with the large runtime overhead and extremely long runtime of other existing methods, it can obtain results quickly and with minimal resource consumption.

[0042] S3. Perform plane segmentation based on the undirected normal of each point. Take real-world point cloud data as input and output the corresponding segmentation cloud and its neighborhood graph. The output segmentation cloud is a scalar field that stores the index of each point in the X, Y, and Z plane area.

[0043] The point cloud data is segmented into planes based on the undirected normal of each point, primarily performing region growing, edge point refinement, and neighbor relationship extraction. First, random points that have not yet been assigned to a region are repeatedly selected from the point cloud and the region to which they belong is determined. If the number of points in a calculated region exceeds a heuristic threshold, it is considered a significant region. This process is repeated until the probability of finding all planes in the dataset reaches a threshold of 99%.

[0044] Region growing, in some embodiments, the plane segmentation based on the undirected normal of each point includes: starting from a single seed point, adding the qualifying points to the region by determining whether the neighboring points are in a plane P, and recording the index of the added points, wherein the plane P is defined by the centroid c and normal n of the region, and initializing the centroid c and normal n of the plane P based on the undirected normal and the position of the seed point, wherein the qualifying points are points whose normals differ from the normal n of the plane P by an angle no greater than α and whose position is at most 3d from the local tangent plane to the region, where d is equal to the radius of the determined domain. α is to create a region of size τ on the surface of a sphere of radius 2ϵ, so that all its points fit into a cone with an opening angle α, where ϵ is equal to the radius of the determined domain, and τ is a threshold for the minimum number of points required to form a region based on seed growing.

[0045] Single region growth, starting from a single seed point, the growth of the region is driven by two index sets R and F, R contains the indexes of all points currently considered to be part of the region, and F contains the indexes of those points added to R in the last iteration. Let the plane normal be n and its center of mass be c, which together define the plane P, as shown in Figure 3 shown.

[0046] Seed points are first added to R and F, and then n and c are initialized using the normals and seed point positions obtained by PCA. At each step of the growing process, all points in a k-neighborhood around any point in F are identified. Then, among the unassigned points in these neighborhoods, points whose normals differ from n by at most α and whose positions are at most 3d away from the local tangent plane to region R are added, replacing the previous points in F. The maximum width of the extracted region is chosen to be 3d to avoid over-segmentation. To reduce the influence of seed point selection on the extracted region, the plane estimate is refitted to R using PCA on the point positions at exponentially increasing intervals. After each refit, points that fail to meet the normal and distance criteria of the new plane estimate are discarded from the region.

[0047] α represents the maximum angular deviation between a point normal and a region normal. Assuming a sphere with a radius of ϵ, the feature point affects the neighboring points within the radius. According to the characteristics of PCA, the points within the radius of ϵ may have unreliable normal estimates. Therefore, the points in the ϵ neighborhood that are not affected by unreliable normal estimates must be at least 2ϵ away from the feature point.

[0048] It can also be interpreted as a safety margin of 2ϵ around the prominent features, the area below 2ϵ is a potential feature to be washed out, and the area above 2ϵ may be an extractable fragment. Therefore, the choice of α needs to be able to ensure that line segments of size ≥ τ can grow on a sphere with a radius of 2ϵ. The necessary condition for α is to create an area of ​​size τ on the surface of a sphere with a radius of 2ϵ so that all its points fit into a cone with an opening angle of α, and the tip coincides with the center of the sphere. Where τ is the threshold for the minimum number of points required to form a region based on seed growth, is the threshold for the minimum number of points required to form a valid planar region, and contains more than The value is more of a user preference, for example, the user may want a region above a certain value, so the choice of τ is only provided as a guide rather than being completely fixed by dataset analysis. is set to be much higher than the noise in the dataset, so the minimum area of ​​a region should be a radius of The order of magnitude of the circle, thus preventing the region composed purely of noise data from being accepted. Since the area of ​​the region is not clearly defined, the threshold number of points τ is used instead, and the number of points around each point is calculated. The number of points within the sphere, then select as the median of these point counts. If the user wants the area A above a certain value, we can do so by setting To merge the corresponding points, it is found through experiments that a higher τ value is beneficial to a larger plane area. Therefore, the area will form an area of ball cap, and compare it with the ball cap area For comparison, setting r = 2ϵ and solving for h yields ,because , α can be derived as . During system initialization , α defaults to 29°.

[0049] The termination of the region growing process is similar to the random sampling consensus algorithm RANSAC, which randomly selects two points to obtain the corresponding model, puts all data points into the calculated error, and finds all points that meet the error threshold, repeating until a model with the most matching points is found. It can be reconstructed from any sample set of size k, where k represents the minimum set of points required to define a shape candidate within its N points. k is a fixed number that depends on the shape type, and for a plane k = 3. The probability of randomly selecting this primitive is

[0050]

[0051] The probability of finding a region of size g in s trials is

[0052]

[0053] From the above two equations, we can see that the number of regions considered during segmentation grows exponentially with the number of sample points k required to identify these regions. Because the method periodically refits the plane estimate, it is possible to construct regions from a single sample point, that is, k=1.

[0054] Edge point refinement,Due to the PCA characteristics used for normal estimation,,the normals of hard edges in the point data will be smoothed.,When applying the region growing method to a point cloud with PCA-derived,normals, points close to hard edges will usually not become part of any region.,The edge point refinement step is to compensate for the limitations of,PCA-based normal estimation.

[0055] Edge point refinement corrects this effect by reordering the assignments of previously unassigned points in the cloud. For each unassigned point, consider all regions that appear near the point and, among all discovered regions, find the region that the point falls within 3 of. If a point is in a band, it is assigned to the region closest to it. Compared to the original region growing method, the normal angle criterion is abandoned while maintaining the validity of the planar distance criterion. This refinement step is fast and easy to implement, and it can obtain high-quality segmentation even around prominent features without relying on computationally expensive methods to obtain standard normals.

[0056] Neighborhood relationship extraction, in addition to extracting the plane area from the point cloud, it is also necessary to extract the neighborhood relationship graph, and obtain a rough multi-level topology estimate based on the relationship between the regional entities. The judgment criterion is that if there are two points and So that q is located at p In the neighborhood, the opposite is also true, then the two areas A and B are neighbors.

[0057] Three heuristic parameters, namely the distance threshold for region growing, the threshold for the minimum number of points required to form a valid planar region, and the decision criterion for adding points to a region, are quickly determined in an unsupervised manner. These parameters are then applied to a region growing method based on normal estimation to complete the point cloud segmentation task in a short time and at a low cost.

[0058] like Figure 4 As a second aspect of the present application, an unsupervised point cloud segmentation system is also disclosed, comprising:

[0059] The domain radius determination module 1 is used to obtain point cloud data and estimate the domain radius so that when extracting the undirected normal of each point based on PCA, the ratio of the first target eigenvalue to the second target eigenvalue is not less than a preset value;

[0060] Point normal extraction module 2 is used to obtain a local representative subset through range search based on the estimated neighborhood radius; estimate the local tangent plane of each point in the local representative subset based on PCA, and extract the undirected normal of each point based on the local tangent plane of each point;

[0061] The plane segmentation module 3 is configured to perform plane segmentation based on the non-directional normal of each point.

[0062] In some embodiments, the first target eigenvalue is the minimum eigenvalue among the eigenvalues , the second target eigenvalue is the second largest eigenvalue among the eigenvalues The ratio of the first target characteristic value to the second target characteristic value is not less than the preset value .

[0063] In some embodiments, estimating the area radius includes: finding a range of area radius such that each point , sets the estimated domain radius to the median of the range of domain radii.

[0064] In some embodiments, the plane segmentation based on the undirected normal of each point includes: starting from a single seed point by seed growth, and adding the qualified points to the region by judging whether the adjacent points are in the plane P, and recording the index of the added points at the same time, the plane P is defined by the center of mass c and normal n of the region, and the center of mass c and normal n of the plane P are initialized based on the undirected normal and the position of the seed point. The qualified point is a point whose normal has an angle difference with the normal n of the plane P that is not greater than α and whose position is at most 3d away from the local tangent plane to the region R, and the size of d is equal to the radius of the determined area.

[0065] In some embodiments, α is a region of size τ created on the surface of a sphere with a radius of 2ϵ, so that all its points fit into a cone with an opening angle of α, the size of ϵ is equal to the radius of the determined domain, and τ is the threshold of the minimum number of points required for the region to be formed based on seed growth.

[0066] Those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not other features, the combination of features from different embodiments is intended to be within the scope of this application and to form different embodiments.

[0067] Those skilled in the art will understand that the description of each embodiment has its own focus, and for parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0068] Although the embodiments of the present application are described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present application, and such modifications and variations shall fall within the scope defined by the appended claims. The above are only specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present invention, and such modifications or replacements shall be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be subject to the scope of protection of the claims.

Claims

1. An unsupervised point cloud segmentation method, characterized in that The steps include: S1. Obtain point cloud data and estimate the neighborhood radius so that when extracting the undirected normal of each point based on PCA, the ratio of the first target eigenvalue to the second target eigenvalue is not less than a preset value, and the first target eigenvalue is the minimum eigenvalue among the eigenvalues. , the second target eigenvalue is the second largest eigenvalue among the eigenvalues , the ratio is not less than the preset value and is expressed as , The estimation of the neighborhood radius includes: finding a neighborhood radius range so that each point , set the estimated neighborhood radius to the median of the range of neighborhood radii; S2. Based on the estimated neighborhood radius, obtain a local representative subset through range search; estimate the local tangent plane of each point in the local representative subset based on PCA, and extract the undirected normal of each point based on the local tangent plane of each point; S3. performing plane segmentation based on the undirected normal of each point, specifically comprising: At a radius of Create a sphere of size The area of ​​​​the All points in the region are fitted into an opening angle of The cone, The size is equal to the neighborhood radius, the The threshold value of the minimum number of points required for the region formed based on seed growth; starting from a single seed point, the neighboring points are judged to be on the plane. In the plane, the points that meet the conditions are added to the specific area, and the index of the added points is recorded. By the centroid of a specific area and normals Definition, initialize the plane based on the undirected normal and the position of the seed point The center of mass and normals , the point that meets the conditions is the normal of the point and the plane Normal The angle difference is no greater than And its position is at most the distance from the local tangent plane to the region point, The size is equal to the neighborhood radius.

2. An unsupervised point cloud segmentation system, characterized in that include: The neighborhood radius determination module is used to obtain point cloud data and estimate the neighborhood radius so that when extracting the undirected normal of each point based on PCA, the ratio of its first target eigenvalue to the second target eigenvalue is not less than a preset value, and the first target eigenvalue is the minimum eigenvalue among the eigenvalues. , the second target eigenvalue is the second largest eigenvalue among the eigenvalues , the ratio is not less than the preset value and is expressed as , The estimation of the neighborhood radius includes: finding a neighborhood radius range so that each point , set the estimated neighborhood radius to the median of the range of neighborhood radii; The point normal extraction module is used to obtain a local representative subset through range search based on the estimated neighborhood radius; the local tangent plane of each point in the local representative subset is estimated based on PCA, and the undirected normal of each point is extracted based on the local tangent plane of each point; A plane segmentation module is used to perform plane segmentation based on the undirected normal of each point, specifically comprising: At a radius of Create a sphere of size The area of ​​​​the All points in the region are fitted into an opening angle of The cone, The size is equal to the neighborhood radius, the The threshold value of the minimum number of points required for the region formed based on seed growth; starting from a single seed point, the neighboring points are judged to be on the plane. In the plane, the points that meet the conditions are added to the specific area, and the index of the added points is recorded. By the centroid of a specific area and normals Definition, initialize the plane based on the undirected normal and the position of the seed point The center of mass and normals , the point that meets the conditions is the normal of the point and the plane Normal The angle difference is no greater than And its position is at most the distance from the local tangent plane to the region point, The size is equal to the neighborhood radius.

Images