Point cloud segmentation method of semi-supervised k-means based on edge enhancement

Abstract

The invention provides a semi-supervised k-means point cloud segmentation method based on edge enhancement, and belongs to the technical field of point cloud data processing. According to the method, the original point cloud data is denoised, the change of the normal vector in the point cloud data set is analyzed, the geometric edge in the point cloud data is extracted, the closed area formed by the edge point cloud set is obtained, the centroid is selected through the morphological closed operation, and the point cloud data is extracted. And segmenting the point cloud data set by using the optimized K-means clustering algorithm, thereby realizing self-adaptive high-quality segmentation of the point cloud data. The edge continuity is effectively improved through morphological closed operation, the number of iterations is reduced based on the mass center initialization strategy of the closed area, the number of clusters is prevented from being manually preset, and the boundary matching precision is improved through introduction of an edge constraint term. Compared with the traditional method, the method provided by the invention has higher stability and accuracy in the aspect of the point cloud segmentation method of the spliced detector.

Description

Technical Field

[0001] The present invention belongs to the technical field of point cloud data processing, and in particular relates to a point cloud segmentation method based on edge enhancement and semi-supervised k-means. Background Art

[0002] With the rapid development of astronomy, large-aperture, wide-field-of-view, and wide-band astronomical optical telescopes have become a key technological tool for solving cutting-edge scientific problems. To detect more distant and fainter celestial objects, the apertures of ground-based telescopes continue to increase. At the same time, to enhance observation capabilities over wider areas of the sky, their fields of view are also expanding. This trend has led to a significant increase in focal plane size. However, due to the current limitations of single detector size, tiled detector technology has become the mainstream solution for constructing large-area focal plane arrays.

[0003] In this context, the flatness of the stitched detector is extremely important. For a given optical system, the tilt error of the CCD target surface must be strictly controlled within the focal depth range. Otherwise, some areas will deviate from the optimal imaging surface, resulting in uneven image quality and difficulty meeting high-precision imaging requirements. Generally speaking, the flatness of the focal plane after stitching needs to be controlled within the range of 20-30μm. The actual integration process usually requires multiple measurements and fine-tuning to achieve the required stitching performance.

[0004] High-precision point cloud segmentation, a key step in point cloud data processing for stitched detectors, is central to achieving these requirements. In recent years, numerous improvements have been proposed for point cloud segmentation. Traditional point cloud segmentation methods rely primarily on geometric constraints and statistical rules to manually design object features, dividing the raw point cloud data into several non-overlapping regions corresponding to individual objects in the scene. For example, traditional segmentation methods based on edge information, model fitting, and region growing have low accuracy when processing complex point clouds. With the advancement of computer technology, techniques for processing point cloud data using deep learning have matured and achieved excellent results. These methods, such as projection-based, voxel-based, and point-based segmentation, require transforming the point cloud into a regular structure suitable for convolutional neural network processing and performing predictions using the network model. While deep learning-based methods demonstrate significant advantages in general scene segmentation, they present challenges in the specific application of large-aperture telescope stitched detectors. These methods require point-by-point annotation of the point cloud data, which is time-consuming and manual. This limitation ensures that traditional methods remain irreplaceable in the stitched detector field.

[0005] Existing traditional segmentation methods still have defects in their applicability to splicing detector engineering: the adaptive region growing algorithm based on neighborhood density reduces dependence on the initial seed point by dynamically adjusting the growth criterion. However, this method requires pre-extraction of ground point clouds and has poor adaptability to unprocessed spliced ​​point clouds. An urban point cloud segmentation process is proposed based on the PCL library, combining voxel filtering and Euclidean clustering to extract building and terrain features. This method has high computational efficiency, but is sensitive to the point cloud density distribution and prone to over-segmentation. A point cloud segmentation algorithm based on octree voxelization and multi-stage region growing. This method achieves efficient spatial indexing through an octree structure, but it is sensitive to the selection of initial seed points, which may affect stability. Summary of the Invention

[0006] The purpose of the present invention is to overcome the defects of the prior art and propose a point cloud segmentation method based on edge-enhanced semi-supervised k-means.

[0007] To achieve the above objectives, the present invention adopts the following specific technical solutions:

[0008] A point cloud segmentation method based on edge enhancement and semi-supervised k-means includes the following steps:

[0009] S1. Obtaining original point cloud data, performing denoising processing on the original point cloud data, and obtaining a three-dimensional point cloud data set;

[0010] S2. Analyze changes in normal vectors in the point cloud data set, extract geometric edges in the point cloud data, and obtain an edge point cloud set;

[0011] S3, in the closed area formed by the edge point cloud, selecting the centroid through morphological closing operation as the initial center point of the K-means clustering optimization function;

[0012] S4. Optimizing the K-means clustering algorithm based on the edge point cloud set;

[0013] S5. Use the optimized K-means clustering algorithm to segment the point cloud dataset and output the point cloud segmentation result.

[0014] Furthermore, the step of obtaining the edge point cloud set in step S2 is specifically as follows:

[0015] Select any point p in the point cloud dataset i , the local surface covariance matrix composed of its k neighboring points is as shown in formula (1):

[0016]

[0017] in, represents the centroid of the neighborhood point set, and k represents the number of neighborhood points;

[0018] Perform eigendecomposition on the covariance matrix, and the eigenvector corresponding to the minimum eigenvalue is the normal vector n i , then the normal vector of any point in the cloud dataset is as shown in formula (2):

[0019]

[0020] Among them, λ1, λ2, λ3 represent the eigenvalues ​​of the covariance matrix, and λ1≤λ2≤λ3;

[0021] By comparing p i and its neighboring point p j The normal vector difference is used to identify the boundary. When p i and its neighboring point p j When the normal vector of is greater than the preset angle threshold, the geometric edge in the point cloud data is extracted, and the edge point cloud set is shown in formula (3):

[0022]

[0023] E represents the edge point cloud set, n i Represents point p i The normal vector, n j Represents point p j The normal vector of , τ represents the preset angle threshold.

[0024] Furthermore, the step of obtaining the initial center point in step S3 is specifically as follows:

[0025] Projecting the edge point cloud onto a two-dimensional plane to obtain a two-dimensional edge projection point set, and identifying closed areas using a contour extraction method;

[0026] discretizing the two-dimensional edge projection point set into a binary image;

[0027] Using morphological closing operation to process the binary image, fill edge gaps, and form a closed area with continuous edges;

[0028] Match each closed region with continuous edges with the original point cloud to extract the point set inside the region;

[0029] The geometric centroid of each internal point set is selected as the initial center point of the K-means clustering optimization function.

[0030] Furthermore, step S4 is specifically as follows:

[0031] Optimize the objective function of the K-means clustering algorithm:

[0032] An edge penalty term is added to the objective function of the K-means clustering algorithm. The edge penalty term is shown in formula (4):

[0033]

[0034] Among them, w ij ∈{0,1} represents a binary membership variable function. When point p i It is 1 if it belongs to cluster j, otherwise it is 0; p i represents the i-th 3D point cloud data point in the closed area; c j represents the center point of the jth cluster; D E (p i )=min q∈E ||p i -q|| means p i The closest Euclidean distance to the edge point set E. λ represents the penalty coefficient, which controls the strength of the edge constraint and is greater than 0; if point p i If it is close to the edge area, the center c of the cluster to which it belongs will be penalized. j , forcing the cluster center to move away from the edge;

[0035] Optimize the cluster centers of the K-means clustering algorithm:

[0036] Take the partial derivative of the optimized objective function and set it to zero to obtain the optimized cluster center;

[0037] Optimize the iterative function of the K-means clustering algorithm:

[0038] The optimized cluster center is shifted in the opposite direction along the edge. If the optimized cluster center offset is less than the set threshold, the iteration is terminated and the optimization of the iterative function is completed.

[0039] Furthermore, the specific processing method of the morphological closing operation is: using the expansion-erosion operation of the circular structure element.

[0040] Furthermore, the method for obtaining the original point cloud data is: using a dual probe with a differential structure to perform scanning, and obtaining the original point cloud data through a real-time differential algorithm.

[0041] Furthermore, the denoising method for the original point cloud data is: a radius filtering algorithm based on a neighborhood.

[0042] The present invention can achieve the following technical effects:

[0043] 1. The point cloud segmentation method based on edge-enhanced semi-supervised k-means provided by the present invention can achieve adaptive high-quality segmentation of point cloud data by fusing morphological processing of edge features with an improved clustering algorithm.

[0044] 2. The point cloud segmentation method based on edge-enhanced semi-supervised k-means provided by the present invention can effectively improve edge continuity through morphological closing operation, reduce the number of iterations based on the centroid initialization strategy of the closed area, and avoid manually presetting the number of clusters. The introduction of edge constraint terms improves the accuracy of boundary matching.

[0045] 3. The point cloud segmentation method based on edge-enhanced semi-supervised k-means provided by the present invention can select the centroid as the initial center point in the closed area formed by the edge points, thereby reducing the uncertainty of the initial position selection, avoiding the centroid falling on the edge or empty area, controlling the centroid from being repeatedly selected, and enhancing stability.

[0046] 4. The point cloud segmentation method based on edge-enhanced semi-supervised k-means provided by the present invention can effectively extract spatial structure areas from edge information and determine a unique initial centroid for each area, thereby avoiding the clustering instability caused by random initialization in the K-means algorithm and improving the consistency with the real structure.

[0047] 5. Compared with the traditional method, the point cloud segmentation method based on edge-enhanced semi-supervised k-means provided by the present invention has higher stability and accuracy in the point cloud segmentation method of the splicing detector. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 1 is a schematic diagram of a method flow of a point cloud segmentation method based on edge-enhanced semi-supervised k-means according to an embodiment of the present invention;

[0049] Figure 2 is a schematic diagram of a large-aperture stitched detector point cloud dataset according to an embodiment of the present invention;

[0050] Figure 3 is a schematic diagram of an edge point cloud of a large-aperture stitching detector according to an embodiment of the present invention;

[0051] Figure 4 is a schematic diagram of the result of morphological operation on the edge point cloud set of the large-aperture stitching detector according to an embodiment of the present invention;

[0052] Figure 5 is a schematic diagram of the segmentation result of the large-aperture stitching detector according to an embodiment of the present invention;

[0053] Figure 6 is a schematic diagram of a height cloud map of a segmentation result of a large-aperture stitching detector according to an embodiment of the present invention;

[0054] Figure 7 3 is a schematic diagram of an algorithm flow of a point cloud segmentation method based on edge-enhanced semi-supervised k-means according to an embodiment of the present invention. DETAILED DESCRIPTION

[0055] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, identical modules are denoted by identical reference numerals. In the case of identical reference numerals, their names and functions are also identical. Therefore, their detailed description will not be repeated.

[0056] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0057] The following combination Figure 1-Figure 7 The specific working mode of the point cloud segmentation method based on edge-enhanced semi-supervised k-means in an embodiment of the present invention is described in detail:

[0058] This embodiment of the present invention uses point cloud data collected from a large-aperture telescope's spliced ​​detector as an example to describe how a semi-supervised k-means-based point cloud segmentation method with edge enhancement can effectively segment point clouds from a large-aperture telescope's spliced ​​detector dataset. In this embodiment of the present invention, the large-aperture telescope's spliced ​​detector data measurement device consists of an air-based platform system, a motion system, and a dual-probe optical measurement system, achieving submicron-level accuracy. Host computer control software was developed based on the QT framework to enable simultaneous and precise acquisition of the motion system's real-time position and dual-probe height data. The base platform system utilizes an active air-floating vibration-isolating optical platform to provide stability support for measurement. The motion system utilizes the MISUMI E-RSZ12-8S-X360-Y600-DE30-XC5-YC5 high-precision two-dimensional platform, achieving ±5μm repeatability and ±10μm linearity within a 600mm × 360mm travel range. The measurement system utilizes a dual-probe differential architecture: a Keyence CL-3000 controller drives two CLP015N / CLP015 probes for measurement. Reference probe #2 measures point cloud data of a plane reflector with a surface shape better than λ / 20 in real time with high-precision linearity of ±0.41μm, recording this data as platform error noise. Working probe #1 performs a wide-area scan of the stitched detector, generating the final 3D raw point cloud data of the large-aperture telescope stitched detector using a real-time differential algorithm. Other conventional methods can also be used to collect 3D raw point cloud data, and this invention is not limited to this method.

[0059] In the embodiment of the present invention, the point cloud segmentation method based on edge-enhanced semi-supervised k-means is as follows: Figure 1 As shown, the following steps are included:

[0060] S1. Obtain original point cloud data, perform denoising on the original point cloud data, and obtain a three-dimensional point cloud dataset.

[0061] The point cloud data collection process is easily affected by environmental interference, laser echo anomalies, sampling errors and other factors, resulting in isolated points, burrs and other noise. In order to remove outliers introduced by noise, occlusion or missampling, the technical solution provided by the embodiment of the present invention eliminates platform noise through dual-probe differential noise reduction technology, obtains original point cloud data, and pre-processes the original point cloud data based on the radius filtering algorithm of the neighborhood to eliminate outliers, and finally obtains a high-quality three-dimensional point cloud data set for subsequent processing. Figure 2 As shown in the figure, the colored bar on the right side represents the height information of the spliced ​​detector point cloud dataset, and X and Y represent the coordinates of the X-axis and Y-axis of the point cloud dataset respectively.

[0062] The basic idea of ​​the radius filtering algorithm based on neighborhood is: if the number of neighboring points of a point within a given radius r is less than the preset threshold N min , then the point is considered as an outlier and is removed.

[0063] Let a point in the point cloud be p i , whose coordinates in three-dimensional space are p i =(x i ,y i ,z i ), then count the number of neighbor points N within a given radius r i The judgment criteria of this method are shown in the following formula (1):

[0064] N i =|{p j ∈P|||p i -p j ||≤r}| (1)

[0065] If the following conditions are met, then p i is an outlier:

[0066] N i <N min

[0067] Among them, P represents the point cloud dataset, ||·|| represents the Euclidean distance, r represents the neighborhood radius, N min represents the neighborhood number threshold, p j Represents any point p i 's neighborhood points.

[0068] S2. Analyze changes in normal vectors in the point cloud data set, extract geometric edges in the point cloud data, and obtain an edge point cloud set.

[0069] In this embodiment, edge detection using normal vectors is used to identify object boundaries by analyzing the changes in normal vectors on local surface areas of a point cloud. In continuous surface areas, the direction of normal vectors between adjacent points changes gradually; however, at object edges, the direction of normal vectors changes dramatically. By detecting changes in the angle between normal vectors, geometric edges in point cloud datasets can be effectively detected.

[0070] Select any point p in the point cloud dataset i , the local surface covariance matrix composed of its k neighboring points is shown in formula (2):

[0071]

[0072] in, represents the centroid of the neighborhood point set; k represents the number of neighborhood points; p j Represents any point p i 's neighborhood points.

[0073] Perform eigendecomposition on the covariance matrix, and the eigenvector corresponding to the minimum eigenvalue is the normal vector n i , then the normal vector of any point in the cloud dataset is as shown in formula (3):

[0074]

[0075] Among them, λ1, λ2, λ3 represent the eigenvalues ​​of the covariance matrix, and λ1≤λ2≤λ3;

[0076] By comparing p i and its neighboring point p j The normal vector difference is used to identify the boundary. When p i and its neighboring point p j When the normal vector of is greater than the preset angle threshold, the geometric edge in the point cloud data is extracted, and the edge point cloud set is shown in formula (3):

[0077]

[0078] E represents the edge point cloud set, n i Represents point p i The normal vector, n j Represents point p j The normal vector of , τ represents the preset angle threshold.

[0079] In the embodiment of the present invention, the edge point cloud set is obtained, such as Figure 3 As shown, the units of X-axis and Y-axis coordinates are both millimeters.

[0080] S3. In the closed area formed by the edge point cloud, a centroid is selected through a morphological closing operation as the initial center point of the K-means clustering optimization function.

[0081] exist Figure 3 The edge point cloud obtained based on normal vector differences shows that the vector estimation method can accurately capture geometric edges with significant curvature changes in the point cloud, such as the red point set. However, the original edge points have localized gaps, such as the one on the left. Therefore, the technical solution of the embodiments of the present invention needs to fill the edge gaps to form a complete closed structure.

[0082] Project the edge point cloud onto a two-dimensional plane to obtain a two-dimensional edge projection point set ε XY , ε XY ={(x i ,y i )|P i ∈E i}, use contour extraction method to identify closed areas.

[0083] The two-dimensional edge projection point set ε XY Discretized into a binary image B(x,y), 1 represents the edge point position, which can be expressed as formula (4):

[0084]

[0085] The binary image is processed using morphological closing operation to fill the edge gaps, making the edges more continuous and forming a closed structure, which can be expressed as formula (5):

[0086]

[0087] in, represents the expansion operation, represents the corrosion operation; S is the structural element.

[0088] In the embodiment of the present invention, the edge gaps are successfully filled through the expansion-erosion operation of the 3×3 circular structure element to form a complete closed structure. The morphological operation results are as follows: Figure 4 shown.

[0089] Get the closed area P k , point set R in the region k It is defined as formula (6):

[0090] R k ={p i ∈P|proj xy (p i )∈P k} (6)

[0091] Among them, projxy Represents the projection of a point on the xy plane.

[0092] Match each closed region with the original point cloud and extract the internal point set of the region. k , the initial cluster center is selected from it to initialize the K-means algorithm. In the embodiment of the present invention, the geometric centroid is used as the initial center point, which is expressed as formula (7):

[0093]

[0094] where c k is the center of the kth region, |R k | represents the number of region points.

[0095] In this way, a set of cluster initialization centers is obtained, which is expressed as formula (8) and serves as the initial center points of the K-means clustering optimization function.

[0096] C={c1,c2,…,c k} (8)

[0097] After acquiring the edge point cloud, the present invention proposes a K-means clustering initialization method based on an edge supervision mechanism. Unlike traditional K-means initialization methods that use random centroids, the present invention selects the centroid from the closed area formed by the edge points as the initial center point, thereby reducing the uncertainty of the initial position selection, preventing the centroid from falling on the edge or in empty areas, and ensuring that the centroid is not repeatedly selected, thereby enhancing stability.

[0098] S4. Optimizing the K-means clustering algorithm based on the edge point cloud set.

[0099] Optimize the objective function of the K-means clustering algorithm:

[0100] An edge penalty term is added to the objective function of the K-means clustering algorithm. The edge penalty term is shown in formula (9):

[0101]

[0102] Among them, w ij ∈{0,1} represents a binary membership variable function. When point p i It is 1 if it belongs to cluster j, otherwise it is 0; p i represents the i-th 3D point cloud data point in the closed area; c j represents the center point of the jth cluster; D E (p i )=min q∈E ||p i -q|| means pi The closest Euclidean distance to the edge point set E. λ represents the penalty coefficient, which controls the strength of the edge constraint and is greater than 0; if point p i If it is close to the edge area, the center c of the cluster to which it belongs will be penalized. j , forcing the cluster centers to move away from the edges.

[0103] The K-means clustering algorithm in the prior art updates the center by minimizing the square distance within the cluster, but the method provided by the embodiment of the present invention needs to consider edge constraints at the same time and can effectively extract spatial structure areas from edge information.

[0104] Optimize the cluster centers of the K-means clustering algorithm:

[0105] By taking the partial derivative of the optimized objective function J and setting it to zero, we can get the optimized cluster center, as shown in formula (10):

[0106]

[0107] Among them, the gradient term Characterization point p i The closest direction to the edge can be approximately calculated as formula (11):

[0108]

[0109] Among them, q * Represents point p i The corresponding nearest edge point, that is, the nearest point on the edge.

[0110] The edge penalty term in the embodiment of the present invention is a correction term, which causes the cluster center to shift in the opposite direction along the edge to ensure the accuracy of the segmentation boundary.

[0111] The technical solution provided by the embodiment of the present invention combines edge information to adjust the optimization function, adds an edge penalty term to the traditional K-means objective function, and forms a new optimization target. This can avoid errors in edge detection, which may introduce incorrect constraints and affect the clustering segmentation results. The iterative function of the K-means clustering algorithm is optimized:

[0112] The optimized cluster center is shifted in the opposite direction of the edge. If the optimized cluster center offset is less than the set threshold, the iteration is terminated and the optimization of the iterative function is completed. The calculation formula (12) is shown below.

[0113]

[0114] in, represents the cluster center coordinates of t iterations; represents the cluster center coordinates of the t+1 iteration; ε represents the set threshold.

[0115] S5. Use the optimized K-means clustering algorithm to segment the point cloud dataset and output the point cloud segmentation result.

[0116] The point cloud segmentation result of the final large-aperture telescope splicing detector in the embodiment of the present invention is as follows: Figure 5 and Figure 6 As shown in the figure, the colored bar on the right side represents the height information of the spliced ​​detector point cloud data set. The point cloud segmentation result processed by the method provided by the embodiment of the present invention shows excellent geometric feature preservation ability. Figure 5 It can be seen that the splicing detector module can identify 16 segmentation results. Figure 6 It can be seen that in the three-dimensional height map, the height information of the color gradient mapping reproduces the differences in the planar deformation characteristics and tilt angles between different modules of the spliced ​​detector.

[0117] The method flow chart in the embodiment of the present invention is as follows Figure 7 As shown, the technical solution provided by the embodiment of the present invention can effectively extract spatial structure areas from edge information and determine a unique initial centroid for each area, thereby avoiding the clustering instability caused by random initialization in the K-means algorithm in the prior art and improving the consistency with the real structure.

[0118] The embodiment of the present invention verifies the effectiveness of the technical solution of the present invention through comparative experiments, and the comparison method is the traditional K-means algorithm and DBSCAN algorithm in the prior art. The embodiment of the present invention adopts the silhouette coefficient and edge matching as core evaluation indicators, and counts the number of iterations at the same time. The silhouette coefficient is a general indicator for segmentation quality assessment, and its value range is [-1,1]. It reflects the comprehensive performance of intra-class tightness and inter-class separation. The larger the value, the more reasonable the clustering structure. The edge matching degree is specially designed for the splicing detector scene, and quantifies the algorithm's ability to maintain geometric boundary features. The closer the value is to 1, the higher the degree of fit between the cluster boundary and the real physical edge. The comparison results are shown in Table 1:

[0119] Table 1

[0120] algorithm Initial centroid Number of iterations Silhouette coefficient Edge matching Traditional k-means yes 15 0.51 0.65 DBSCAN no -- 0.58 0.72 Technical solution of the present invention no 6 0.62 0.85

[0121] As can be seen from Table 1, the technical solution provided by the embodiment of the present invention uses edge-enhanced semi-supervised k-means with a silhouette coefficient of 0.62, which is 21.6% higher than the traditional K-means, and the edge matching degree reaches 0.85, which is better than the comparison algorithm. In terms of efficiency, the technical solution provided by the embodiment of the present invention only requires 6 iterations to converge, which is 56.8% less than the traditional K-means, and there is no need to manually preset the initial centroid. The traditional K-means algorithm is limited by the random initial centroid selection, and its silhouette coefficient is significantly greater than the technical solution provided by the embodiment of the present invention. Although DBSCAN does not require a preset number of clusters, it is sensitive to the neighborhood radius and the minimum number of points. As shown in Table 1, its edge matching degree drops to 0.72 in the light scattering area, and it is unable to directly output the centroid position, making it difficult to meet the requirements of precision measurement.

[0122] Therefore, the technical solution proposed in the embodiments of the present invention: a point cloud segmentation method based on edge-enhanced semi-supervised k-means, achieves adaptive, high-quality segmentation of point cloud data by integrating morphological processing of edge features with an improved clustering algorithm. Experiments show that the morphological closing operation effectively improves edge continuity, and the centroid initialization strategy based on closed regions reduces the number of iterations by 56.8%, eliminating the need for manually presetting the number of clusters. The introduction of edge constraints improves boundary matching accuracy by over 57%. Compared with traditional methods, this method demonstrates greater stability and accuracy in point cloud segmentation using spliced ​​detectors.

[0123] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0124] Although the embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art may make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.

[0125] The above specific embodiments of the present invention do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made based on the technical concept of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A point cloud segmentation method based on edge enhancement semi-supervised k-means, characterized in that: The steps include: S1. Obtaining original point cloud data, performing denoising processing on the original point cloud data, and obtaining a three-dimensional point cloud data set; S2. Analyze changes in normal vectors in the point cloud data set, extract geometric edges in the point cloud data, and obtain an edge point cloud set; S3, in the closed area formed by the edge point cloud, selecting the centroid through morphological closing operation as the initial center point of the K-means clustering optimization function; S4. Optimizing the K-means clustering algorithm based on the edge point cloud set; S5. Use the optimized K-means clustering algorithm to segment the point cloud dataset and output the point cloud segmentation result.

2. The point cloud segmentation method according to claim 1, characterized in that: The steps for obtaining the edge point cloud set in step S2 are specifically as follows: Select any point p in the point cloud dataset i , the local surface covariance matrix composed of its k neighboring points is as shown in formula (1): in, represents the centroid of the neighborhood point set, and k represents the number of neighborhood points; Perform eigendecomposition on the covariance matrix, and the eigenvector corresponding to the minimum eigenvalue is the normal vector n i , then the normal vector of any point in the cloud dataset is as shown in formula (2): Among them, λ1, λ2, λ3 represent the eigenvalues ​​of the covariance matrix, and λ1≤λ2≤λ3; By comparing p i and its neighboring point p j The normal vector difference is used to identify the boundary. When p i and its neighboring point p j When the normal vector of is greater than the preset angle threshold, the geometric edge in the point cloud data is extracted, and the edge point cloud set is shown in formula (3): E represents the edge point cloud set, n i Represents point p i Normal vector, n j Represents point p j The normal vector of , τ represents the preset angle threshold.

3. The point cloud segmentation method according to claim 2, characterized in that: The steps of obtaining the initial center point in step S3 are specifically as follows: Projecting the edge point cloud onto a two-dimensional plane to obtain a two-dimensional edge projection point set, and identifying closed areas using a contour extraction method; discretizing the two-dimensional edge projection point set into a binary image; Using morphological closing operation to process the binary image, fill edge gaps, and form a closed area with continuous edges; Match each closed region with continuous edges with the original point cloud to extract the point set inside the region; The geometric centroid of each internal point set is selected as the initial center point of the K-means clustering optimization function.

4. The point cloud segmentation method according to claim 3, characterized in that: Step S4 is specifically as follows: Optimize the objective function of the K-means clustering algorithm: An edge penalty term is added to the objective function of the K-means clustering algorithm. The edge penalty term is shown in formula (4): Among them, w ij ∈{0,1} represents a binary membership variable function. When point p i It is 1 if it belongs to cluster j, otherwise it is 0; p i represents the i-th 3D point cloud data point in the closed area; c j represents the center point of the jth cluster; D E (p i )=min q∈E ||p i -q|| means p i The closest Euclidean distance to the edge point set E. λ represents the penalty coefficient, which controls the strength of the edge constraint and is greater than 0; if point p i If it is close to the edge area, the center c of the cluster to which it belongs will be penalized. j , forcing the cluster center to move away from the edge; Optimize the cluster centers of the K-means clustering algorithm: Take the partial derivative of the optimized objective function and set it to zero to obtain the optimized cluster center; Optimize the iterative function of the K-means clustering algorithm: The optimized cluster center is shifted in the opposite direction along the edge. If the optimized cluster center offset is less than the set threshold, the iteration is terminated and the optimization of the iterative function is completed.

5. The point cloud segmentation method according to claim 3, characterized in that: The specific processing method of the morphological closing operation is: using the expansion-erosion operation of the circular structure element.

6. The point cloud segmentation method according to claim 1, characterized in that: The method for obtaining the original point cloud data is: using a dual probe with a differential structure to scan and obtain the original point cloud data through a real-time differential algorithm.

7. The point cloud segmentation method according to claim 6, characterized in that: The method for denoising the original point cloud data is: a radius filtering algorithm based on a neighborhood.

Images