A fitting-based ground target segmentation method

Abstract

The application discloses a fitting-based ground target segmentation method, comprising the following steps: preliminary screening, polar coordinate grid division, ground point classification, iterative optimization and post-processing of the preliminary classification result, removal of noise points and correction of misclassified points. The fitting-based ground target segmentation method of the application enhances the adaptability of the algorithm to noise and complex environment through a multi-step processing procedure; through pre-filtering and grid-based processing, the data processing amount is greatly reduced, thereby improving the overall efficiency of the algorithm.

Description

Technical Field

[0001] This invention belongs to the field of three-dimensional data processing and analysis technology, specifically relating to a fitting-based ground target segmentation method. Background Technology

[0002] Currently, there are various LiDAR-based methods for land target segmentation both domestically and internationally, mainly categorized into five types: grid-based, fitting-based, feature-based, graph-based, and learning-based.

[0003] The grid-based approach projects point cloud data onto a 2D elevation map and establishes surface attributes on each grid cell. While this method can quickly obtain segmentation results, it may incorrectly segment overhanging points and planar points on roads, leading to insufficient point cloud information for subsequent perception tasks.

[0004] Fitting-based methods typically divide a point cloud into multiple regions and fit the ground in each region using a planar or linear model. This algorithm is well-suited for roads with steep slopes, but its segmentation accuracy largely depends on the selection of seed points.

[0005] Feature-based methods classify land and non-land surfaces using the geometric relationships between neighboring points. These methods can accurately identify non-locations, but are prone to missegmentation due to noisy points.

[0006] Graph-based methods typically use graph models to segment points, while learning-based methods use neural networks. Both methods have excellent performance in ground segmentation, but they are both time-consuming. In addition, learning-based methods require a large amount of labeled data to train the model, and they are not well used for different types of sensors or scenarios that are different from the training dataset.

[0007] Ground segmentation, as a preprocessing task for many sensing tasks, should meet the following two requirements: First, the ground segmentation must be fast enough to operate the sensing system more effectively; second, the ground segmentation should be robust.

[0008] In summary, there is an urgent need for a new fitting-based ground target segmentation method to address the undersegmentation problem of existing fitting-based methods. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a fitting-based ground target segmentation method that can enhance the algorithm's adaptability to noise and complex environments, reduce data processing volume, and improve the overall efficiency of the algorithm.

[0010] The technical solution of this invention is: a ground target segmentation method based on fitting, comprising the following steps:

[0011] (1) Preliminary screening: Remove obvious non-ground points using a vertical point filter;

[0012] (2) Polar coordinate grid division: Point cloud is constructed using a polar coordinate grid. The origin of the lidar coordinate system is the coordinate origin. The coordinates are divided into K segments in the polar coordinate direction, and each segment is divided into L grids. The point cloud data is converted to a polar coordinate system using the following formula, and the data is mapped to a discrete polar coordinate grid:

[0013]

[0014] Where k represents the angular index in the polar coordinate grid, l represents the radial index in the polar coordinate grid, and r i,j θ represents the radial distance from point (i,j) to the origin. i,j r represents the azimuth angle in polar coordinates. max x represents the maximum radial distance of the point cloud. i,j and y i,j Let X and Y represent the coordinates of point (i,j) respectively;

[0015] (3) Ground point classification: Calculate the height value of each grid in the polar grid, and smooth the surface elevation grid using a weighted average filtering method. Set a threshold; if the height value of a grid is less than the threshold, the point is considered a ground point; otherwise, it is a non-ground point. The formula for calculating the height value of each grid in the polar grid is as follows:

[0016]

[0017] Where d represents the distance between two adjacent grid cells, and w represents the final weighted average. W represents the height value of each grid cell in the polar grid. u Z represents the weight of the u-th value. u Let u represent the ground height of the u-th value, and N represent the number of neighboring points;

[0018] (4) Iterative optimization, including the following steps:

[0019] S1. Set the height value of each grid cell in the polar grid. Update the height z to the minimum height value of the grid point, while leaving the height values ​​of other grids that do not contain the location empty;

[0020] S2. Fill the empty grid between two non-empty grids using linear interpolation;

[0021] S3, Set the height value of each grid cell in the polar grid. Perform smoothing;

[0022] S4. Extend the height value of each grid in the polar grid to adjacent grids;

[0023] S5. If the height value of each grid in the polar grid is less than the threshold in step (3), reclassify the non-ground point as a ground point;

[0024] (5) Post-processing: perform post-processing on the preliminary classification result, remove noise points, and correct misclassified points.

[0025] Further, the method for the vertical point filter to remove obvious non-ground points is: set a vertical threshold to filter out point cloud data related to the ground, that is, zp<threshold, where zp represents the coordinate value of point P on the Z axis, and threshold represents the vertical threshold.

[0026] Further, the method for processing boundaries in the step of smoothing the surface elevation value grid by the weighted mean filtering method is a mirror boundary processing method, and the mirror boundary processing method includes the following steps:

[0027] (1) Determine the boundary;

[0028] (2) Create virtual grids: for each boundary grid, create one or more virtual grids outside thereof according to the direction of the boundary;

[0029] (3) Assign values: the elevation value of the virtual grid is set to the same elevation value as the nearest boundary grid;

[0030] (4) Calculate weights: the weight of the virtual grid is determined according to the distance between the virtual grid and the boundary grid, the closer the distance, the larger the weight; the farther the distance, the smaller the weight;

[0031] (5) Weighted average: perform weighted average on the elevation values of the boundary grid and its neighbor grids to obtain the final elevation value of the boundary grid.

[0032] Still further, the calculation formula of the weighted average is:

[0033]

[0034] wherein, represents the final elevation value of the boundary grid, N represents the number of neighbor grids, z i represents the elevation value of the i-th neighbor grid, w i represents the weight.

[0035] Still further, the calculation formula of the distance is:

[0036]

[0037] Where d represents the Euclidean distance between the virtual grid point and the boundary grid point, and r v r represents the radial distance between virtual grid points. b θ represents the radial distance between boundary grid points. v θ represents the polar coordinate angle of a virtual grid point. b Represents the polar coordinate angle of the boundary grid points;

[0038] The formula for calculating the weight is:

[0039]

[0040] Among them, w i is the weight of the i-th grid cell, and d represents the Euclidean distance between the virtual grid point and the boundary grid point.

[0041] Furthermore, the formula for calculating the threshold is:

[0042] T rect (x,y)=μ(x,y)+k·σ(x,y),

[0043] Among them, T rect (x,y) represents the adaptive threshold at point (x,y), μ(x,y) represents the local mean at point (x,y), σ(x,y) represents the local standard deviation at point (x,y), and k represents a constant for adjusting the threshold sensitivity.

[0044] Furthermore, the step of filling the empty grid between two non-empty grids using the linear interpolation method is as follows:

[0045] (1) Determine the empty grid cells that need to be interpolated;

[0046] (2) Find two adjacent non-empty grids, and set these two non-empty grids as (k1, l1) and (k2, l2). Their heights are z respectively. k1,l1 and z k2,l2 ;

[0047] (3) Calculate the interpolation weights:

[0048]

[0049] Where l1 represents the position of the first non-empty grid, l2 represents the position of the second non-empty grid, l represents the interpolation position, w1 represents the interpolation weight of the first non-empty grid, and w2 represents the interpolation weight of the second non-empty grid.

[0050] (4) Calculate the interpolation height:

[0051]

[0052] Among them, z k,l The height of the empty grid is represented by z, w1 represents the interpolation weight of the first non-empty grid, w2 represents the interpolation weight of the second non-empty grid, and z represents the height of the empty grid. k1,l1 z represents the height value of the first non-empty grid cell. k2,l2 This represents the height value of the second non-empty grid.

[0053] (5) The calculated height value z k,l Assign the value to the empty grid;

[0054] (6) Repeat steps (1) to (5) above for all empty grids until all empty grids are filled.

[0055] Furthermore, the expression for the smoothing process in step S3 is:

[0056]

[0057] in, This represents the ground height value of the current grid cell. This indicates the ground height value of the grid cell to the left of the current grid cell. This indicates the ground height value of the grid cell adjacent to the right of the current grid cell.

[0058] Furthermore, in step S4, the height value of each grid in the polar grid is... The expression for extending to adjacent grids is:

[0059]

[0060] in, This represents the ground height value of the current grid cell. This represents the ground height value of the neighboring grid, and (k′,l′)∈N represents the grid coordinates in the neighborhood.

[0061] The beneficial effects of this invention are as follows: The ground target segmentation method based on fitting of this invention effectively reduces the impact of noise data by combining Vertical Point Filter (VPF) and Seed Point Extraction (SPE). Furthermore, it improves the quality of seed point selection, thereby enhancing the accuracy of surface fitting. Grid-based refinement (GBR) refines the coarse segmentation results obtained from the initial fitting, solving the problem of insufficient segmentation in some areas and ensuring the integrity of the final segmentation results. The overall framework design considers the complexity and variability of different scenarios, and the multi-step processing flow enhances the algorithm's adaptability to noise and complex environments. Pre-filtering and grid-based processing significantly reduce the amount of data processed, thereby improving the overall efficiency of the algorithm. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic flowchart of a fitting-based ground target segmentation method according to the present invention.

[0064] Figure 2 This is a schematic diagram of the preliminary screening steps of a fitting-based ground target segmentation method according to the present invention.

[0065] Figure 3 This is a polar coordinate system schematic diagram of a ground target segmentation method based on fitting according to the present invention.

[0066] Figure 4 This is a polar coordinate grid diagram of a fitting-based ground target segmentation method according to the present invention. Detailed Implementation

[0067] like Figure 1-4 As shown, a fitting-based ground target segmentation method includes the following steps:

[0068] (1) Preliminary screening: Remove obvious non-ground points using a vertical point filter;

[0069] (2) Polar coordinate grid division: Point cloud is constructed using a polar coordinate grid. The origin of the lidar coordinate system is the coordinate origin. The coordinates are divided into K segments in the polar coordinate direction, and each segment is divided into L grids. The point cloud data is converted to a polar coordinate system using the following formula, and the data is mapped to a discrete polar coordinate grid:

[0070]

[0071] Where k represents the angular index in the polar coordinate grid, l represents the radial index in the polar coordinate grid, and r i,j θ represents the radial distance from point (i,j) to the origin. i,j r represents the azimuth angle in polar coordinates. max x represents the maximum radial distance of the point cloud. i,j and y i,j Let X and Y represent the coordinates of point (i,j) respectively;

[0072] (3) Ground point classification: Calculate the height value of each grid in the polar grid, smooth the surface elevation value grid using a weighted mean filtering method, set a threshold, and if the height value of a grid is less than the threshold, the point is regarded as a ground point, otherwise it is a non-ground point; the calculation formula for calculating the height value of each grid in the polar grid is:

[0073]

[0074] wherein d represents the distance between two adjacent grids, w represents the final weighted average, represents the height value of each grid in the polar grid, W u represents the weight of the u-th value, Z u represents the ground height of the u-th value, and N represents the number of neighborhood points;

[0075] (4) Iterative optimization, including the following steps:

[0076] S1. Update the height value of each grid in the polar grid to the minimum height value z of the ground point at that grid point, and keep the height values of other grids that do not contain ground points as empty;

[0077] S2. Use a linear interpolation method to fill the empty grids between two non-empty grids;

[0078] S3. Perform smoothing processing on the height value of each grid in the polar grid ;

[0079] S4. Extend the height value of each grid in the polar grid to adjacent grids;

[0080] S5. If the height value of each grid in the polar grid is less than the threshold in step (3), reclassify the non-ground point as a ground point;

[0081] (5) Post-processing: Perform post-processing on the preliminary classification result, remove noise points, and correct misclassified points.

[0082] Preferably, the method for the vertical point filter to remove obvious non-ground points is: set a vertical threshold, and screen out point cloud data related to the ground, that is: zp<threshold, where zp represents the coordinate value of point P on the Z axis, and threshold represents the vertical threshold.

[0083] Preferably, the method for processing boundaries in the step of smoothing the surface elevation value grid by the weighted mean filtering method is a mirror boundary processing method, and the mirror boundary processing method includes the following steps:

[0084] (1) Determine the boundaries;

[0085] (2) Create virtual meshes: For each boundary mesh, create one or more virtual meshes outside of it according to the direction of the boundary;

[0086] (3) Assignment: Set the elevation value of the virtual grid to be the same as the elevation value of the nearest boundary grid;

[0087] (4) Calculate the weight: The weight of the virtual mesh is determined by its distance from the boundary mesh. The closer the distance, the greater the weight; the farther the distance, the smaller the weight.

[0088] (5) Weighted average: The elevation values ​​of the boundary grid and its neighboring grids are weighted and averaged to obtain the final elevation value of the boundary grid.

[0089] More preferably, the formula for calculating the weighted average is:

[0090]

[0091] in, This represents the final elevation value of the boundary grid, where N represents the number of neighboring grids, and z... i w represents the elevation value of the i-th neighboring grid. i Indicates the weight.

[0092] More preferably, the formula for calculating the distance is:

[0093]

[0094] Where d represents the Euclidean distance between the virtual grid point and the boundary grid point, and r v r represents the radial distance between virtual grid points. b θ represents the radial distance between boundary grid points. v θ represents the polar coordinate angle of a virtual grid point. b Represents the polar coordinate angle of the boundary grid points;

[0095] The formula for calculating the weight is:

[0096]

[0097] Among them, w i is the weight of the i-th grid cell, and d represents the Euclidean distance between the virtual grid point and the boundary grid point.

[0098] Preferably, the threshold is calculated using the following formula:

[0099] T rect (x,y)=μ(x,y)+k·σ(x,y),

[0100] Among them, T rect (x,y) represents the adaptive threshold at point (x,y), μ(x,y) represents the local mean at point (x,y), σ(x,y) represents the local standard deviation at point (x,y), and k represents a constant for adjusting the threshold sensitivity.

[0101] Preferably, the step of filling the empty grid between two non-empty grids using the linear interpolation method is as follows:

[0102] (1) Determine the empty grid cells that need to be interpolated;

[0103] (2) Find two adjacent non-empty grids, and set these two non-empty grids as (k1, l1) and (k2, l2). Their heights are z respectively. k1,l1 and z k2,l2 ;

[0104] (3) Calculate the interpolation weights:

[0105]

[0106] Where l1 represents the position of the first non-empty grid, l2 represents the position of the second non-empty grid, l represents the interpolation position, w1 represents the interpolation weight of the first non-empty grid, and w2 represents the interpolation weight of the second non-empty grid.

[0107] (4) Calculate the interpolation height:

[0108]

[0109] Among them, z k,l The height of the empty grid is represented by z, w1 represents the interpolation weight of the first non-empty grid, w2 represents the interpolation weight of the second non-empty grid, and z represents the height of the empty grid. k1,l1 z represents the height value of the first non-empty grid cell. k2,l2 This represents the height value of the second non-empty grid.

[0110] (5) The calculated height value z k,l Assign the value to the empty grid;

[0111] (6) Repeat steps (1) to (5) above for all empty grids until all empty grids are filled.

[0112] Preferably, the expression for the smoothing process in step S3 is:

[0113]

[0114] in, This represents the ground height value of the current grid cell. This indicates the ground height value of the grid cell to the left of the current grid cell. This indicates the ground height value of the grid cell adjacent to the right of the current grid cell.

[0115] Preferably, in step S4, the height value of each grid in the polar grid is... The expression for extending to adjacent grids is:

[0116]

[0117] in, This represents the ground height value of the current grid cell. This represents the ground height value of the neighboring grid, and (k′,l′)∈N represents the grid coordinates in the neighborhood.

[0118] In practical applications, during the Vertical Point Filter (VPF) stage, the point cloud is preprocessed by analyzing the vertical distribution characteristics in the point cloud data. VPF uses the vertical coordinate information of the point cloud to set a vertical threshold to filter out point cloud data that are related to the ground. Specifically, VPF calculates the vertical distribution characteristics of each point in the point cloud and filters out points that do not conform to the ground characteristics, such as floating objects or noise points, thereby obtaining a cleaner point cloud dataset. This step can significantly improve the quality of the data and reduce errors in subsequent processing.

[0119] Next, in the Seed Point Extraction (SPE) stage, preliminary seed points are automatically extracted from the point cloud data after VPF processing. The SPE algorithm selects representative seed points by analyzing the local geometric features of the point cloud, such as the density and spatial distribution of points. These seed points are usually located in important areas of the ground and can provide stable initial fitting values. The SPE algorithm ensures that these seed points can accurately represent the ground area, providing a solid foundation for the subsequent fitting process and reducing segmentation errors caused by improper selection of initial points.

[0120] Finally, mesh-based refinement (GBR) technology is used to further optimize the preliminary ground segmentation results. GBR first establishes a raster model of the point cloud, performs finite element analysis on it, and analyzes it. By refining the mesh boundary and optimizing the mesh interior, GBR can correct the undersegmentation problem in the preliminary segmentation. Specific methods include fine-tuning the ground region in the mesh to make the ground boundary smoother and more accurate, and handling outliers and discontinuous regions in the mesh, thereby improving the overall segmentation accuracy.

[0121] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that any non-creative modifications or equivalent substitutions to the technical solutions of the present invention, without departing from the scope of protection of the present invention, should be covered within the scope of the claims of the present invention.

Claims

1. A ground target segmentation method based on fitting, characterized in that, Includes the following steps: (1) Preliminary screening: Remove obvious non-ground points using a vertical point filter; (2) Polar coordinate grid division: Point cloud is constructed using a polar coordinate grid. The origin of the lidar coordinate system is the coordinate origin. The coordinates are divided into K segments in the polar coordinate direction, and each segment is divided into L grids. The point cloud data is converted into a polar coordinate system using the following formula, and these data are mapped to a discrete polar coordinate grid: ， in, k This represents the angle index in the polar coordinate grid. l Indicates the radial index in the polar coordinate grid. r i,j Representing a point ( i,j The radial distance from the origin. θ i,j This represents the azimuth angle in polar coordinates. r max This represents the maximum radial distance of the point cloud. x i,j and y i,j Representing points ( i,j The X and Y coordinates of ). (3) Ground point classification: Calculate the height value of each grid in the polar grid, and use a weighted average filtering method to smooth the surface elevation grid. Set a threshold; if the height value of a grid is less than the threshold, the point is considered a ground point; otherwise, it is a non-ground point. The formula for calculating the height value of each grid in the polar grid is: ， in, d Indicates the distance between two adjacent grid cells. w This represents the final weighted average. This represents the height value of each grid cell in the polar grid. W u Indicates the first u The weight of each value Z u Indicates the first u The ground height of each value, N Indicates the number of neighboring points; (4) Iterative optimization, including the following steps: S1. Set the height value of each grid cell in the polar grid. Updated to the minimum height value of ground points within this grid. z The height values ​​of other grids that do not include ground points are left empty; S2. Fill the empty grid between two non-empty grids using linear interpolation; S3, Set the height value of each grid cell in the polar grid. Perform smoothing; S4. Set the height value of each grid cell in the polar grid. Extend to adjacent grids; S5, the height value of each grid cell in the polar grid If the value is less than the threshold in step (3), the non-ground point is reclassified as a ground point; (5) Post-processing: Post-process the preliminary classification results to remove noise points and correct misclassified points.

2. The ground target segmentation method based on fitting according to claim 1, characterized in that, The method for the vertical point filter to remove obvious non-ground points is: set a vertical threshold to filter out point cloud data related to the ground, that is: zp <threshold, where zp represents the coordinate value of point P on the Z axis, and threshold represents the vertical threshold.

3. The ground target segmentation method based on fitting according to claim 1, characterized in that, The method for processing boundaries in the smoothing step of the weighted mean filtering method for surface elevation values ​​is the mirror boundary processing method, which includes the following steps: (1) Determine the boundaries; (2) Creating virtual meshes: For each boundary mesh, create one or more virtual meshes outside it according to the direction of the boundary; (3) Assignment: The elevation value of the virtual grid is set to the same as the elevation value of the nearest boundary grid; (4) Calculate the weight: The weight of the virtual mesh is determined based on its distance from the boundary mesh. The closer the distance, the greater the weight; the farther the distance, the smaller the weight. (5) Weighted average: The elevation values ​​of the boundary grid and its neighboring grids are weighted and averaged to obtain the final elevation value of the boundary grid.

4. The ground target segmentation method based on fitting according to claim 3, characterized in that, The formula for calculating the weighted average is: ， in, This represents the final elevation value of the boundary grid. N Indicates the number of neighboring grids. z i Indicates the first i Elevation values ​​of neighboring grids, w i Indicates the weight.

5. The ground target segmentation method based on fitting according to claim 3, characterized in that, The formula for calculating the distance is: ， in, d This represents the Euclidean distance between virtual grid points and boundary grid points. r v Represents the radial distance between virtual grid points. r b Indicates the radial distance between boundary grid points. θ v Represents the polar coordinate angle of the virtual grid point. θ b Represents the polar coordinate angle of the boundary grid points; The formula for calculating the weight is: in, w i It is the first i The weight of each grid cell, d This represents the Euclidean distance between virtual grid points and boundary grid points.

6. The ground target segmentation method based on fitting according to claim 1, characterized in that, The formula for calculating the threshold is: ， in, T rect ( x , y ) represents a point ( x,y Adaptive threshold at ) μ ( x , y ) represents a point ( x,y Local mean at point ( ) σ ( x , y ) represents a point ( x,y Local standard deviation at ) k It represents a constant that indicates the sensitivity of the threshold adjustment.

7. The ground target segmentation method based on fitting according to claim 1, characterized in that, The steps for filling the empty grid between two non-empty grids using the linear interpolation method are as follows: (1) Determine the empty grid cells that need to be interpolated; (2) Find two adjacent non-empty grids and set these two non-empty grids as ( k 1 ,l 1 )and( k 2 ,l 2 If their heights are respectively... z k1,l1 and z k2,l2 ; (3) Calculate the interpolation weights: ， in, l 1 Indicates the position of the first non-empty grid cell. l 2 Indicates the position of the second non-empty grid. l Indicates the position of the interpolation. w 1 represents the interpolation weight of the first non-empty grid. w 2 indicates the interpolation weight of the second non-empty grid; (4) Calculate the interpolation height: ， in, z k,l This represents the height value of an empty grid. w 1 represents the interpolation weight of the first non-empty grid. w 2 represents the interpolation weight of the second non-empty grid. z k1,l1 This represents the height value of the first non-empty grid cell. z k2,l2 This represents the height value of the second non-empty grid. (5) Calculate the height value z k,l Assign the value to the empty grid; (6) Repeat steps (1) to (5) above for all empty grids until all empty grids are filled.

8. The ground target segmentation method based on fitting according to claim 1, characterized in that, The expression for the smoothing process in step S3 is: ， in, This represents the ground height value of the current grid cell. This indicates the ground height value of the grid cell to the left of the current grid cell. This indicates the ground height value of the grid cell adjacent to the right of the current grid cell.

9. The ground target segmentation method based on fitting according to claim 1, characterized in that, In step S4, the height value of each grid in the polar grid is... The expression for extending to adjacent grids is: ， in, This represents the ground height value of the current grid cell. This represents the ground height value of neighboring grid cells. k′,l′ )∈N represents the grid coordinates in the neighborhood.

Images