A long-range ground laser point cloud plane segmentation method

Abstract

The application discloses a long-range ground laser point cloud plane segmentation method, and specifically comprises the following steps: T1, determining the best neighborhood based on a dynamic search range, T2, designing a region growing algorithm based on dimensional features and the best neighborhood, and relates to the technical field of laser point cloud data processing application. The long-range ground laser point cloud plane segmentation method has density adaptability, the neighborhood search range is automatically generated based on a theoretical point spacing, and the needs of point cloud neighborhood searching at different scanning distances can be met; the construction of the best neighborhood guarantees that the calculation of point cloud local features is not affected by density changes; the estimation of the sampling interval conforms to the ground laser scanner acquisition principle, and the rationality of the theoretical point spacing is enhanced; the best neighborhood is added to the region growing, the plane segmentation algorithm is adapted to high-density change point clouds, has wide applicability of the dynamic search range, and the determination of the neighborhood search range starts from a single point, and the overall density change of the point cloud does not need to be analyzed.

Description

Technical Field

[0001] This invention relates to the field of laser point cloud data processing and application technology, specifically a long-range ground laser point cloud plane segmentation method. Background Technology

[0002] Laser scanning technology is one of the main methods for building information acquisition. Since the point clouds collected by laser scanners are scattered and disordered, effectively extracting building information from them is of great research significance for smart city construction.

[0003] Laser scanning technology is categorized into airborne laser scanning technology, vehicle-mounted laser scanning technology, and ground-based laser scanning technology, depending on the carrier. Based on the data acquisition method, airborne laser scanning technology can collect information about building rooftops from the air, but lacks information about building facades. Ground-based laser scanning technology, on the other hand, can collect rich information about building facades, making it an important data source for refined 3D reconstruction.

[0004] Planar segmentation is an important method for extracting building facades, including region growing, model fitting, and feature clustering. Region growing is the process of growing seed points in their neighborhoods according to certain rules. First, seed points are selected based on criteria such as minimum curvature [Besl, 1988] and highest surface saliency [Wu Hao, 2019]. Then, growth points are searched in their neighborhoods according to growth rules. Features such as normal vector, Euclidean distance [CL Kang, 2020], point-to-surface distance [Lu Weixin, 2015], and plane fitting residual [XNing, 2009] are widely used in growth rules. At the same time, growth points that have not been grown are used as new seed points until all points are traversed. Model fitting achieves segmentation by fitting the original point cloud to a specific model, which is composed of geometric primitives [S Xia, 2020]. Among the algorithms with good performance are Hough transform [D Borrmann, 2011], random sampling consensus [LLi, 2017] and its derivative algorithms [B Zhao, 2020]; feature clustering methods consist of two steps: boundary extraction and clustering. Boundaries can be effectively detected based on the local features of points. For example, the angle between the normal vectors of boundary points and adjacent points is large, while the angle between the normal vectors of points on the same plane is small [E Che, 2017], and the angle between the start and end directions of the enclosing path of the neighborhood centered on the boundary point is less than 360° [C Mineo, 2019]. Clustering algorithms include k-means, k-nearest neighbor, and density-based algorithms (DBSCAN) [Jin Jianguo, 2014].

[0005] Current methods for dividing building plans have the following shortcomings:

[0006] 1. Insufficient consideration of the impact of high-density variations on segmentation. With continuous optimization of scanner hardware, the range of terrestrial laser scanning has increased, leading to the problem of high-density variations. For example, a scan line angle of 0.06°, a maximum scan distance of 1km, and a point spacing ranging from millimeters to meters result in high-density variations within the scene, making it difficult to standardize neighborhood sizes. Region growing and feature clustering are highly sensitive to neighborhood selection, and the choice of neighborhood size affects the quality of planar segmentation.

[0007] 2. Difficulty in balancing efficiency and accuracy. Model fitting methods offer higher robustness but require higher quality original point clouds. Due to ground cover occlusion, ground-based laser point clouds have certain deficiencies in integrity, and the collected points become sparser as the scanning distance increases. Therefore, using this algorithm on long-distance ground-based point clouds requires point cloud optimization, impacting algorithm efficiency. Region growing algorithms are simple in principle and highly efficient, widely used for large-scene point cloud segmentation. Seed point selection and growth rules affect segmentation efficiency and quality. A reasonable seed point selection method can significantly reduce algorithm time, and the quality of the segmentation plane is determined by the growth rules. Furthermore, feature fitting requires a uniform density in the point cloud; uneven or excessively low density can lead to missing boundaries, affecting the quality and stability of segmentation. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] To address the shortcomings of existing technologies, this invention provides a long-range ground laser point cloud planar segmentation method, which solves the problem that the point cloud density changes significantly with the increase of scanning distance in long-range ground-based point clouds. Existing planar segmentation methods rarely address this situation. Therefore, this invention designs a density-adaptive segmentation method to solve the following problems:

[0010] 1) A dynamic search range is designed to find the optimal neighborhood, addressing the issue of neighborhood size selection caused by density variations. The high density variation characteristics of long-range ground-based point clouds result in different neighborhood sizes for point clouds in different density regions, affecting planar segmentation. Setting a search range to find the optimal neighborhood for each point is an effective approach to solving this problem. Furthermore, the search range is determined by the theoretical point spacing, and the search range is constructed based on a single point, allowing the algorithm to adapt to ground laser point clouds with different longest scanning distances.

[0011] 2) Improve the region growing algorithm to balance accuracy and efficiency. Model fitting and feature clustering have certain requirements for the quality of the original point cloud. The quality of ground laser point clouds is difficult to meet these requirements, and the design of optimization measures is difficult. However, the principle of region growing is simple, can be improved according to different requirements, and has high efficiency. To improve the segmentation quality, principal component analysis is used to calculate dimensional features to solve the problem of boundary point misclassification. Growth rules are designed based on the best neighborhood to improve the accuracy of local feature calculation and make the planar segmentation results more reasonable.

[0012] (II) Technical Solution

[0013] To achieve the above objectives, the present invention provides the following technical solution: a long-range ground laser point cloud planar segmentation method, specifically comprising the following steps:

[0014] T1. Determine the optimal neighborhood based on the dynamic search range. The specific determination method includes the following steps:

[0015] e1. Ground filtering: Use cloth filtering algorithm to separate non-ground points from ground points and remove non-ground point clouds;

[0016] e2. Sampling Interval Estimation: Based on the acquisition principle, N points are randomly selected from the non-ground point cloud, and the k nearest neighbor algorithm is used to find k nearest neighbor points for each point. The formula is as follows:

[0017] m=1,2,…,N

[0018] In the formula, p m For random points, p mn For p m The nearest neighbor, S(p) m ) is the set of nearest neighbors.

[0019] In the scanner coordinate system, according to p m and p mn Calculate the horizontal angle Δα between the three-dimensional coordinates. mn and the vertical angle Δβ mn ,Δα mn Δβ mn ∈[0°,180°], if Δα mn For angles ≥180°, use 360° - Δα. mn Indicates the horizontal angle

[0020] ;

[0021] e3. Establishing the Dynamic Search Range: The point spacing of the point cloud acquired by the terrestrial laser scanner is related to the scanning distance and sampling interval. The point spacing directly affects the point cloud density. Utilizing this characteristic, the theoretical point spacing is calculated based on the sampling interval, combined with the scanning distance. p m Theoretical distance between points for ;

[0022] Based on existing experiments, neighborhoods are selected within a range of point spacings from 1 to 10 to maximize local features. Therefore, p... m Dynamic search range of points Determined as: ;

[0023] e4. Optimal neighborhood search using joint clustering analysis of internal and external indices and Shannon entropy: Select int* (int=1,2,…,10) is used as the clustering radius. The KNN method is used to cluster the point cloud, and the internal and external indices are used to evaluate the clustering results. The best clustering result is selected. The internal index evaluates the results based on the spatial relationship of each class after clustering. Considering the impact of the massive amount of point cloud on the algorithm efficiency, the DB index is selected among the internal indices, and the Rand index is selected among the external indices to verify the rationality of the DB index.

[0024] If clustering divides the point cloud into m classes M={M1,M2,…,Mm}, and manual classification divides it into n classes N={N1,N2,…,Nn}, the DB index and Rand index are calculated as follows:

[0025] ;

[0026] Among them, avg(M i ), avg(M j ) represents M i M j The average point spacing, dcen(M) i M j ) is M i With M j The distance between the centroids, n ij For M i With N j The number of commonalities, nj and ni, are M respectively. i N j The number of points, N is the total number of points in the point cloud; the smaller the DB index and the larger the Rand index, the better the clustering result.

[0027] At this point, the integer int corresponding to the neighborhood radius of the best clustering result is... Extract the effective subset from the dynamic search range for the left and right spans;

[0028] Shannon entropy represents the amount of information in the neighborhood of a point. The smaller the value, the more obvious the local features. Shannon entropy is calculated as follows:

[0029] Ef=- ln( )- ln( )- ln( ) ;

[0030] Where Ef is the Shannon entropy, δ1, δ2, and δ3 are the plane fitting residuals of the neighborhood of that point, calculated by principal component analysis, and δ1 ≥ δ2 ≥ δ3. Within the effective subset of the dynamic search range, 0.1* The interval values ​​are taken as the neighborhood radius. The covariance matrix of the neighborhood fitting plane is calculated using PCA. The eigenvalues ​​of the matrix are squared to obtain the fitting residuals δ1, δ2, and δ3 in the three axes. The Shannon entropy is calculated according to the formula for Shannon entropy. The neighborhood radius where Ef is minimized is selected. That is, the optimal neighborhood is determined as follows: ;

[0031] At this point, the local features are most prominent. The dimensional features (Dim) of the point are analyzed based on the fitting residuals.

[0032] ;

[0033] Where Dim∈[0,2], corresponding to linear points, planar points and discrete points respectively, and λ1≥λ2≥λ3.

[0034] T2. Design of a region growing algorithm based on dimensional features and optimal neighborhood, the specific method of which includes the following steps:

[0035] p1. Building Extraction: The clustering results include not only building classes, but also interference classes such as streetlights and trees. Since most of the points in the building class are planar points, while interference objects such as streetlights and trees have more discrete points and linear points, the building class is selected according to the proportion of planar points in each class. Different classes in the figure have different colors, and gray represents the rendering result.

[0036] p2. Seed point selection and growth rule setting: In the building, seed points are selected based on the best neighborhood, dimensional features, normal vector, and point-to-surface distance;

[0037] p3. Oversegmented plane merging: Due to the transmissivity of laser, windows appear as holes in the point cloud. The spatial discontinuity causes the initial plane obtained by segmentation to be oversegmented.

[0038] Preferably, the sampling interval in e2 is the horizontal or vertical angle between adjacent scan lines of the scanner, and the corresponding radians are the angular resolution.

[0039] Preferably, in step e2, based on the acquisition principle, in the spherical neighborhood S(p) determined by KNN... mIn ), p m The nearest neighbor points are most frequent on adjacent scan lines. A histogram is constructed by statistically analyzing the distribution of horizontal angles across several intervals, with an interval width of Δ. The interval with the most points is selected from the histogram, and the average horizontal angle α within that interval is calculated. avg .

[0040] Preferably, the rationality of the parameter Δ setting affects the estimation quality of the sampling interval. Δ should be set to values ​​in intervals of 0.001° within the range of [0.005°, 0.015°], and α should be calculated accordingly. avg Finally, take all α avg The median value med(α) avg ) represents the final sampling interval.

[0041] Preferably, in step e4, the internal indicators are Dunn index, silhouette coefficient and Davidson-Bolding index based on the spatial relationship of each cluster. The external indicators use the manual classification results as reference data and compare them with the clustering results to analyze the effectiveness of the clustering results. They are divided into standard mutual information index, F-Measure index, Rand index and Jacarrd coefficient.

[0042] Preferably, the specific method for setting the growth rules in step p2 is as follows;

[0043] U1. Selecting Seed Points: In the building category, the planar point is used as the starting seed point. During the growth process, all growth points are used as new seed points. Any point is only subjected to a growth operation once.

[0044] U2. Set growth rules: First, the growth point must be a planar point as a prerequisite. Then, the growth point must meet the distance requirements, normal vector requirements, and point-to-plane distance requirements.

[0045] ;

[0046] Where D is the Euclidean distance between the growth point and the seed point. φ is the optimal neighborhood radius of the seed point, and φ is the angle between the normal vectors of the growth point and the seed point. th It is the threshold of the angle between the normal vectors, d ’ It is the larger value of the point-to-surface distance, d ’ th It is the point-to-surface distance threshold.

[0047] U3. Assigning Linear Points: Linear points are also part of the plane. After all plane points have grown, the linear points are assigned to the nearest plane to obtain the initial segmentation result.

[0048] Preferably, the oversegmentation problem in step p3 is addressed by optimizing the initial result through planar merging. Planar merging needs to meet the following requirements:

[0049] (1) Requirement for the angle between the normal vectors of two planes: Calculate the angle between the normal vectors of the two planes. The angle must be less than the threshold φ. th ;

[0050] (2) Point-to-plane distance requirement: For two planes P1 and P2, find the geometric centers C1 and C2 respectively. The point-to-plane distance is calculated as follows:

[0051] ;

[0052] Where D c Let C1 and C2 be the spatial distance. , The point-to-plane distance must be less than a threshold normal vector between the two planes. ;

[0053] (3) Planar distance requirement: Oversegmentation is caused by missing windows. The maximum height of the windows in the scene is used as the distance threshold.

[0054] Preferably, a balanced binary tree is first established with a point in P1 as the center, and a search range is constructed using a distance threshold. Then, points in P2 within this range are queried, and if a point exists, the planar distance requirement is satisfied.

[0055] (III) Beneficial Effects

[0056] This invention provides a long-range ground laser point cloud planar segmentation method. Compared with existing technologies, it has the following advantages:

[0057] (1) The long-range ground laser point cloud planar segmentation method has density adaptability. The neighborhood search range is automatically generated based on the theoretical point spacing (this scheme uses 1-10 times the theoretical point spacing as the dynamic search range of the neighborhood), which can meet the needs of finding the neighborhood of the point cloud at different scanning distances. The construction of the best neighborhood ensures that the calculation of the local features of the point cloud is not affected by density changes. The estimation of the sampling interval conforms to the acquisition principle of the ground laser scanner, which enhances the rationality of the theoretical point spacing. The best neighborhood is added to the region growth, so that the planar segmentation algorithm can adapt to high-density changing point clouds.

[0058] (2) The long-range ground laser point cloud plane segmentation method has a wide applicability of dynamic search range. The determination of the neighborhood search range starts from a single point, without the need to analyze the overall density change of the point cloud, thus avoiding repeated setting of the search range for different distance-based ground point clouds and having stronger applicability.

[0059] (3) This long-range ground laser point cloud plane segmentation method, through the use of internal and external indicators of cluster analysis and reasonable analysis of cluster results, eliminates invalid subsets of dynamic search range while ensuring the quality of optimal neighborhood extraction, avoids the calculation of Shannon entropy on invalid subsets, and reduces the search time of optimal neighborhood.

[0060] (4) This long-range ground laser point cloud plane segmentation method has both efficiency and accuracy. By extracting boundary points through dimensional features, it solves the problem of inaccurate boundary normal vector estimation. Based on the clustering results, it extracts building classes through the overall dimensional features of each class and performs plane segmentation only on building classes, which significantly reduces the amount of data and improves processing efficiency. By selecting plane points as seed points and growth points, it effectively eliminates discrete points. Finally, it assigns the boundary points to the nearest plane, which improves the accuracy of the segmentation results at the boundary. Based on the window size, it designs the patch merging conditions, which effectively solves the over-segmentation problem caused by missing window information and makes the plane segmentation results more reasonable. Attached Figure Description

[0061] Figure 1 This is a flowchart of the planar segmentation process based on dynamic search range according to an embodiment of the present invention;

[0062] Figure 2 This is a schematic diagram illustrating the calculation of the horizontal and vertical angles between a random point and its nearest neighbor points according to an embodiment of the present invention.

[0063] Figure 3 This is a schematic diagram of the construction of a horizontal angle histogram according to an embodiment of the present invention;

[0064] Figure 4 This is a schematic diagram illustrating the determination of the optimal neighborhood in an embodiment of the present invention;

[0065] Figure 5 This is a schematic diagram of building extraction according to an embodiment of the present invention;

[0066] Figure 6 This is a schematic diagram illustrating the region growth effect in an embodiment of the present invention;

[0067] Figure 7 This is a schematic diagram of patch merging in an embodiment of the present invention. Detailed Implementation

[0068] 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.

[0069] Please see Figure 1-7This invention provides a technical solution: a long-range ground laser point cloud plane segmentation method, specifically including the following steps:

[0070] T1. Determine the optimal neighborhood based on the dynamic search range. The specific determination method includes the following steps:

[0071] e1. Ground filtering: Use cloth filtering algorithm to separate non-ground points from ground points and remove non-ground point clouds.

[0072] e2. Sampling Interval Estimation: The sampling interval is the horizontal or vertical angle between adjacent scan lines of the scanner, and its corresponding radians are the angular resolution. According to the acquisition principle, N points are randomly selected in the non-ground point cloud, and the k nearest neighbor algorithm is used to find the k nearest neighbor points for each point. The formula is:

[0073] m=1,2,…,N

[0074] In the formula, p m For random points, p mn For p m The nearest neighbor, S(p) m ) is the set of nearest neighbors.

[0075] In the scanner coordinate system, according to p m and p mn Calculate the horizontal angle Δα between the three-dimensional coordinates. mn and the vertical angle Δβ mn ,Δα mn Δβ mn ∈[0°,180°], if Δα mn For angles ≥180°, use 360° - Δα. mn Indicates the horizontal angle

[0076] ;

[0077] Based on the acquisition principle, in the spherical neighborhood S(p) determined by KNN m In ), p m The nearest neighbor points are most abundant on adjacent scan lines, such as Figure 2 Simultaneously, a histogram is constructed to statistically analyze the distribution of horizontal angles across several intervals, with an interval width of Δ. The interval with the most points is selected from the histogram, and the average horizontal angle α within that interval is calculated. avg .

[0078] The appropriateness of the parameter Δ setting affects the estimation quality of the sampling interval. Δ should be set to values ​​in 0.001° intervals within the range [0.005°, 0.015°], and α should be calculated accordingly. avg Finally, take all α avg The median value med(α)avg ) represents the final sampling interval.

[0079] e3. Establishing the Dynamic Search Range: The point spacing of the point cloud acquired by the terrestrial laser scanner is related to the scanning distance and sampling interval. The point spacing directly affects the point cloud density. Utilizing this characteristic, the theoretical point spacing is calculated based on the sampling interval, combined with the scanning distance. p m Theoretical distance between points for ;

[0080] Based on existing experiments, neighborhoods are selected within a range of point spacings from 1 to 10 to maximize local features. Therefore, p... m Dynamic search range of points It can be determined as follows:

[0081] e4. Optimal neighborhood search using joint clustering analysis of internal and external indices and Shannon entropy: Select int* (int=1,2,…,10) is used as the clustering radius. The KNN method is used to cluster the point cloud, and the internal and external indices are used to evaluate the clustering results and select the best clustering result.

[0082] Internal metrics evaluate the results based on the spatial relationships of clustered classes. These metrics include Dunn index, silhouette coefficient, and Davidson-Bolding index. External metrics use manual classification results as reference data and compare them with the clustering results to analyze the effectiveness of the clustering results. These metrics include standard mutual information (NMI), F-Measure index, Rand index, and Jacarrd coefficient. Considering the impact of massive point cloud volume on algorithm efficiency, the DB index is selected among the internal metrics, and the Rand index is selected among the external metrics to verify the rationality of the DB index.

[0083] If clustering divides the point cloud into m classes M={M1,M2,…,Mm}, and manual classification divides it into n classes N={N1,N2,…,Nn}, the DB index and Rand index are calculated as follows:

[0084] ;

[0085] Among them, avg(M i ), avg(M j ) represents M i M j The average point spacing, dcen(M) i M j ) is M iWith M j The distance between the centroids, n ij For M i With N j The number of commonalities, nj and ni, are M respectively. i N j The number of points is N, where N is the total number of points in the point cloud. A smaller DB index and a larger Rand index indicate better clustering results. Figure 4 To determine the optimal neighborhood graph, where Figure 4 (a) represents internal and external indicators. Figure 4 (b) is the Shannon entropy.

[0086] At this point, the integer int corresponding to the neighborhood radius of the best clustering result is... Extracting a valid subset from the dynamic search range for left and right spans, such as Figure 4 (a);

[0087] Shannon entropy represents the amount of information in the neighborhood of a point. The smaller the value, the more obvious the local features. Shannon entropy is calculated as follows:

[0088] Ef=- ln( )- ln( )- ln( ) ;

[0089] Where Ef is the Shannon entropy, δ1, δ2, and δ3 are the plane fitting residuals of the neighborhood of the point, which can be calculated by principal component analysis (PCA) [5], and δ1≥δ2≥δ3. In the effective subset of the dynamic search range, 0.1* The interval values ​​are taken as the neighborhood radius. The covariance matrix of the neighborhood fitting plane is calculated by PCA. The eigenvalues ​​of the matrix are squared to obtain the fitting residuals δ1, δ2, and δ3 in the three axes. The Shannon entropy is calculated according to equation (6). Figure 4 (b) Select the neighborhood radius where Ef is minimized. That is, to determine the best neighborhood.

[0090] ;

[0091] At this point, the local features are most pronounced, and the dimensional features (Dim) of the point can be analyzed based on the fitting residuals.

[0092] ;

[0093] Where Dim∈[0,2], corresponding to linear points, planar points and discrete points respectively, and λ1≥λ2≥λ3.

[0094] T2. Design of a region growing algorithm based on dimensional features and optimal neighborhood.

[0095] p1. Building Extraction: The clustering results include not only building clusters but also interfering clusters such as streetlights and trees. Since the majority of points in the building cluster are planar points, while interfering objects like streetlights and trees have more discrete and linear points, the building cluster is selected based on the proportion of planar points in each cluster. Figure 5 Different categories in the image have different colors; gray represents the rendering result.

[0096] p2. Seed point selection and growth rule setting: In the building, seed points are selected based on the best neighborhood, dimensional features, normal vector and point-to-surface distance, and growth rules are set. The specific method is as follows;

[0097] 1) Select seed points. In the building class, the planar point is used as the starting seed point. During the growth process, all growth points are used as new seed points, and any point is only subjected to a growth operation once.

[0098] 2) Set growth rules. First, the growth point must be a planar point as a prerequisite. Then, the growth point must meet the distance requirements, normal vector requirements, and point-to-plane distance requirements:

[0099] ;

[0100] Where D is the Euclidean distance between the growth point and the seed point. φ is the optimal neighborhood radius of the seed point, and φ is the angle between the normal vectors of the growth point and the seed point. th It is the threshold of the angle between the normal vectors, d ’ It is the larger value of the point-to-surface distance, d ’ th It is the point-to-surface distance threshold.

[0101] 3) Assigning line points. Line points are also part of the plane. After all plane points have grown, the line points are assigned to the nearest neighbor plane to obtain the initial segmentation result, such as... Figure 6 (b)

[0102] p3. Over-segmented plane merging: Due to the transmissivity of laser light, windows appear as holes in the point cloud. This spatial discontinuity causes the initial plane obtained from segmentation to be over-segmented, such as... Figure 7 (a) To address the over-segmentation problem, the initial result is optimized through planar merging. Planar merging needs to meet the following requirements:

[0103] (1) Requirement for the angle between the normal vectors of two planes. Calculate the angle between the normal vectors of the two planes; the angle must be less than the threshold φ. th .

[0104] (2) Point-to-plane distance requirement. For two planes P1 and P2, the geometric centers C1 and C2 are obtained respectively. The point-to-plane distance is calculated as follows:

[0105] ;

[0106] Where D c Let C1 and C2 be the spatial distance. , The point-to-plane distance must be less than a threshold normal vector between the two planes. .

[0107] (3) Planar distance requirement. Oversegmentation is caused by missing windows. The maximum height of the windows in the scene can be used as the distance threshold. First, a balanced binary tree is built with a point in P1 as the center, and the search range is constructed using the distance threshold. Then, P2 is queried for points within this range. If a point exists, the planar distance requirement is met. Figure 7 This is a merged image of facets, where Figure 7 (a) shows the initial segmentation result. Figure 7 (b) is a distance threshold of 1.8m; Figure 7 (c) represents a distance threshold of 0.5m. Figure 7 (b) represents the patch merging effect when the distance threshold is set to 1.8m, which effectively solves the problem of over-segmentation. Figure 7 (c) represents the patch merging effect when the distance threshold is 0.5m, and oversegmentation still exists.

[0108] Furthermore, any content not described in detail in this specification is existing technology known to those skilled in the art.

[0109] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0110] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for long-range ground laser point cloud plane segmentation, characterized in that, Specifically, the following steps are included: T1. Determine the optimal neighborhood based on the dynamic search range. The specific determination method includes the following steps: e1. Ground filtering: Use cloth filtering algorithm to separate non-ground points from ground points and remove non-ground point clouds; e2. Sampling Interval Estimation: Based on the acquisition principle, point V is randomly selected from the ground point cloud, and the k nearest neighbor algorithm is used to find k nearest neighbor points. The formula is: , m = 1, 2,..., V; In the formula, p m For random points, p mn For p m The nearest neighbor, S(p) m () represents the nearest neighbor set; In the scanner coordinate system, according to p m and p mn Calculate the horizontal angle Δα between the three-dimensional coordinates. mn and the vertical angle Δβ mn If Δα mn For angles ≥180°, use 360° - Δα. mn Indicates the horizontal angle, even if Δα mn Δβ mn ∈[0°,180°]; ； e3. Establishing the Dynamic Search Range: The point spacing of the point cloud acquired by the terrestrial laser scanner is related to the scanning distance and sampling interval. The point spacing directly affects the point cloud density. Utilizing this characteristic, based on the sampling interval, the scanning distance is combined with... Calculate the theoretical point spacing, p m Theoretical distance between points for ; Based on existing experiments, neighborhoods are selected within a range of point spacings from 1 to 10 to maximize local features. Therefore, p... m Dynamic search range of points Determined as: ; e4. Optimal neighborhood search using joint clustering analysis of internal and external indices and Shannon entropy: Selecting int from As the clustering radius, the KNN method is used to cluster the point cloud, and the clustering results are evaluated using internal and external indices. The best clustering result is selected. The internal indices evaluate the results based on the spatial relationship of each cluster. Considering the impact of the massive amount of point cloud on the algorithm efficiency, the DB index is selected among the internal indices, and the Rand index is selected among the external indices to verify the rationality of the DB index. If clustering divides the point cloud into m classes M={M1,M2,…,Mm}, and manual classification divides it into n classes N={N1,N2,…,Nn}, the DB index and Rand index are calculated as follows: ； Among them, avg(M i ), avg(M j ) represents M i M j The average point spacing, dcen(M) i M j ) is M i With M j The distance between the centroids, n ij For M i With N j The number of commonalities , M respectively i N j The number of points, where N is the total number of points in the point cloud; At this time, the optimal clustering result corresponds to the neighborhood radius corresponding to the integer int, and the neighborhood radius corresponding to the integer int is extract the effective subset in the dynamic search range as the left and right spans Shannon entropy represents the amount of information in the neighborhood of a point. The smaller the value, the more obvious the local features. Shannon entropy is calculated as follows: Ef=- ln( )- ln( )- ln( ) ; Where Ef is the Shannon entropy, δ1, δ2, and δ3 are the plane fitting residuals of the neighborhood of the point, calculated by principal component analysis, and δ1 ≥ δ2 ≥ δ3. Within the effective subset of the dynamic search range, 0.1... The interval values ​​are taken as the neighborhood radius. The covariance matrix of the neighborhood fitting plane is calculated using PCA. The eigenvalues ​​of the matrix are squared to obtain the plane fitting residuals δ1, δ2, and δ3. The Shannon entropy is calculated according to the formula for Shannon entropy. The neighborhood radius that minimizes Ef is selected. That is, the optimal neighborhood is determined as follows: ; At this point, the local features are most prominent. The dimensional features (Dim) of the point are analyzed based on the fitting residuals. ； Wherein, Dim∈[0, 2], respectively corresponding linear point, plane point and discrete point, and ≥ ≥ ; T2. Design of a region growing algorithm based on dimensional features and optimal neighborhood, the specific method of which includes the following steps: p1. Building extraction: Select building categories based on the proportion of planar points in each category; p2. Seed point selection and growth rule setting: In the building, seed points are selected based on the best neighborhood, dimensional features, normal vector, and point-to-surface distance; p3. Oversegmented plane merging: Due to the transmissivity of laser, windows appear as holes in the point cloud. The spatial discontinuity causes the initial plane obtained by segmentation to be oversegmented.

2. The long-range ground laser point cloud plane segmentation method according to claim 1, characterized in that: In step e2, the sampling interval is the horizontal or vertical angle between adjacent scan lines of the scanner, and the radians corresponding to the horizontal or vertical angle are the angular resolution.

3. The long-range ground laser point cloud plane segmentation method according to claim 1, characterized in that: In step e2, based on the acquisition principle, within the spherical neighborhood S(p) determined by KNN... m In ), p m The nearest neighbor points are most frequent on adjacent scan lines. Simultaneously, the distribution of horizontal angles is statistically analyzed across several intervals, and a histogram is constructed with an interval width of Δ. The interval with the most points is selected from the histogram, and the average horizontal angle α within that interval is calculated. avg .

4. The long-range ground laser point cloud plane segmentation method according to claim 3, characterized in that: The appropriateness of the parameter Δ setting affects the estimation quality of the sampling interval. Δ should be set to values ​​in 0.001° intervals within the range [0.005°, 0.015°], and α should be calculated accordingly. avg Finally, take all α avg The median value med(α) avg ) represents the final sampling interval.

5. The long-range ground laser point cloud plane segmentation method according to claim 1, characterized in that: In step e4, the internal indicators are Dunn's index, silhouette coefficient, and Davidson-Bolding index, which are based on the spatial relationships of each cluster. The external indicators use the manual classification results as reference data and compare them with the clustering results to analyze the effectiveness of the clustering results. They are divided into standard mutual information index, F-Measure index, Rand index, and Jacarrd coefficient.

6. The long-range ground laser point cloud plane segmentation method according to claim 1, characterized in that: The specific method for setting the growth rules in step p2 is as follows: U1. Selecting Seed Points: In the building category, the planar point is used as the starting seed point. During the growth process, all growth points are used as new seed points. Any point is only subjected to a growth operation once. U2. Set growth rules: First, the growth point must be a planar point as a prerequisite. Then, the growth point must meet the distance requirements, normal vector requirements, and point-to-plane distance requirements. ； Where D is the Euclidean distance between the growth point and the seed point. It is the optimal neighborhood radius of the seed point. It is the angle between the normal vectors of the growth point and the seed point. It is the threshold of the angle between the normal vectors, d ’ It is the value of the point-to-plane distance. It is the point-to-surface distance threshold; U3. Assigning Linear Points: Linear points are also part of the plane. After all plane points have grown, the linear points are assigned to the nearest plane to obtain the initial segmentation result.

7. The long-range ground laser point cloud plane segmentation method according to claim 1, characterized in that: The over-segmentation problem in step p3 is addressed by optimizing the initial result through planar merging. Planar merging needs to meet the following requirements: (1) Plane normal vector angle requirement: calculate the angle between the normal vectors of two planes, and the angle needs to be less than a threshold ; (2) Point-to-plane distance requirement: For two planes P1 and P2, find the geometric centers C1 and C2 respectively. The point-to-plane distance is calculated as follows: ； where D c is the spatial distance between C1, C2, , is the normal vector of the two planes, the point-plane distance needs to be less than a threshold ; (3) Planar distance requirement: Oversegmentation is caused by missing windows. The maximum height of the windows in the scene is used as the distance threshold.

8. The long-range ground laser point cloud plane segmentation method according to claim 7, characterized in that: First, construct a balanced binary tree centered on a point in P1. Construct a search range using a distance threshold. Then, query points in P2 within this range. If a point exists, the planar distance requirement is satisfied.

Images