An airborne LiDAR point cloud filtering method and device based on supervoxel ground saliency

Abstract

The present invention provides an airborne LiDAR point cloud filtering method and device based on supervoxel-based saliency calculation. The method first denoises the input LiDAR point cloud, implements supervoxel segmentation using a boundary-preserving TBBP supervoxel algorithm, divides the supervoxels into scanning bands in N directions according to the centroid coordinates of the supervoxels; sorts and segments the scanning bands in each direction, and calculates the saliency values of each segment; uses the saliency values as coordination factors to construct an energy function of the optimal filtering plane and minimizes the energy function to solve for the optimal segmentation plane; classifies ground points and non-ground points according to the segmentation plane to determine the point cloud filtering result. The saliency calculated in units of supervoxels by the present invention can preliminarily filter the target point cloud in the vertical direction, improve the phenomenon that the point cloud of buildings with concave structures is easily misclassified as ground points, and can meet the filtering requirements in different terrain scenarios such as mountainous areas, forest areas, and complex buildings.

Description

Technical Field

[0001] The present invention belongs to the field of surveying and mapping science and technology, and relates to an airborne LiDAR point cloud filtering method and device based on supervoxel ground saliency, which are mainly applied to fields such as digital elevation model (DEM) generation, point cloud classification, and three-dimensional model reconstruction. Background Art

[0002] With the rapid development of unmanned aerial vehicle and lidar (LiDAR) technologies, airborne lidar has been widely used in the acquisition of three-dimensional spatial information. Lidar has a certain penetration ability, and LiDAR point clouds can penetrate sparse vegetation to obtain high-precision elevation information of the ground surface, which is an efficient and accurate data source for three-dimensional model reconstruction of the ground surface. The original LiDAR point clouds contain three-dimensional information of various ground objects such as the ground surface, vegetation, and buildings in the target scene. Point cloud filtering can screen out ground point clouds from the massive point clouds, laying a foundation for generating high-precision digital elevation models and subsequent further processing such as point cloud classification.

[0003] Existing point cloud filtering methods can be roughly divided into six categories according to their principles, namely slope-based, morphology-based, clustering-based, surface fitting-based, irregular triangulation network-based, and machine learning-based methods. Each of the above methods has its own advantages and disadvantages. In practical applications, it is necessary to select appropriate algorithms and prior thresholds according to the ground object characteristics of the application scenario to ensure the accuracy and speed of the filtering results. For areas with large differences in terrain features, there is still a lack of a general method to meet the point cloud filtering requirements in multiple scenarios. Summary of the Invention

[0004] The purpose of the present invention is to provide a LiDAR point cloud filtering method based on supervoxel ground saliency to solve the problem of incorrect filtering easily occurring in existing point cloud filtering methods based on regular grids in scenarios with large differences in terrain features. The present invention first constructs supervoxels for the input point clouds, divides multi-directional scan bands and calculates saliency for the point clouds in units of supervoxels, transforms the point cloud filtering problem into the problem of dividing the optimal filtering plane of the point clouds within the supervoxels, constructs an energy function in combination with saliency parameters, and uses the semi-global optimization idea to solve the optimal filtering plane, realizing the separation of ground points and non-ground points. Compared with existing methods, the present invention fully considers the point cloud distribution characteristics in different scenarios and can improve the universality of the filtering results as much as possible under the premise of minimizing the input of filtering parameters.

[0005] To achieve the above purpose, a LiDAR point cloud filtering method based on supervoxel ground saliency includes the following steps:

[0006] Step 1, load the point cloud data and perform preprocessing, and use a statistical filter to delete obvious outlier noise points in the LiDAR point clouds;

[0007] Step 2: Perform supervoxel segmentation on the preprocessed point cloud. The supervoxel resolution r is input externally to generate a supervoxel point cloud cluster that preserves the original local boundaries, and calculate the centroid coordinates of each supervoxel unit.

[0008] Step 3: Divide the supervoxels into scanning bands in N directions according to the centroid coordinates of the supervoxels.

[0009] Step 4: Sort and segment each scanning band, and calculate the computational significance value of each segment.

[0010] Step 5: Use the significance value as a coordination factor to construct an energy function of the optimal filtering plane and minimize this energy function to solve for the optimal segmentation plane. Classify the ground points and non-ground points according to the segmentation plane to determine the point cloud filtering result.

[0011] Further, in Step 2, a boundary-preserving TBBP supervoxel algorithm is used to implement supervoxel segmentation.

[0012] Further, the value of N is 16.

[0013] Further, in Step 3, the specific method for dividing the scanning bands is as follows:

[0014] Step 3.1: In the two-dimensional XOY plane, starting from the y-axis and rotating clockwise around the origin, with each 22.5° as a scanning direction, finally generate 16 equally divided scanning directions in the two-dimensional plane.

[0015] Step 3.2: In each scanning direction, using the supervoxel resolution r as the translation unit, generate multiple parallel lines passing through the target point cloud in the two-dimensional XOY plane, and divide the supervoxels into multiple scanning bands according to the supervoxel centroid coordinates.

[0016] Further, in Step 4, the specific method for segmenting the bands and calculating the significance is as follows:

[0017] Step 4.1: Within each scanning band, sort the supervoxel units in ascending order according to the x coordinate of the supervoxel center. For the bands in the scanning direction parallel to the y-axis, sort them in ascending order according to the y coordinate of the supervoxel center. Set the elevation value of each supervoxel to the elevation value of the lowest point within the supervoxel, and set the initial significance value of each supervoxel to 1.

[0018] Step 4.2: Calculate the absolute value of the elevation difference between two adjacent supervoxels in the sorting direction. If the absolute value is less than the segment elevation threshold ΔH, then classify this supervoxel into the current segment; otherwise, classify it into the next segment. Finally, complete the segmentation of all supervoxel units on the scanning line.

[0019] Step 4.3: After segmentation is complete, calculate the elevation difference between the last supervoxel in the current segment and the first supervoxel in the next adjacent segment, following the scanning order. If the elevation difference is greater than twice the segmentation threshold ΔH, subtract 1 / N from the significance value of all supervoxels in the segment. Repeat the above significance calculation steps for segments whose significance has not changed in this scan.

[0020] Step 4.4: perform steps 4.2 and 4.3 on N scanning directions respectively to obtain each supervoxel unit S i The significance value of .

[0021] Furthermore, the elevation threshold ΔH in step 4.2 is externally input and determined by the average error limit between the filtered ground points and the DEM generated from them.

[0022] Furthermore, the number of iterations in step 4.3 is determined by the actual terrain, and is set to 1 for areas with large terrain undulations and 3 for areas with flat terrain.

[0023] Furthermore, in step 5, the specific method for determining the energy function construction is as follows:

[0024] Step 5.1, calculate each supervoxel unit S i Minimum elevation of the point cloud within As the elevation value of the current supervoxel, H min To obtain the minimum elevation value of the entire input point cloud, the supervoxel point cloud is discretized in the elevation direction:

[0025]

[0026]

[0027] In the above formula, N z is the maximum value of the interval that the filter plane can take after the elevation discretization of the current supervoxel, d is the unit height of the elevation discretization; m is the elevation compensation coefficient; Supervoxel S i The elevation value of the candidate filtering plane, n is [0,N z ] Any integer in the interval; Supervoxel S i The elevation value of the candidate filtering plane The collection of supervoxels S on the t-th scan i The energy function is expressed as:

[0028] E t (S i )=E bata (S i )+E smooth (S i-1,S i )

[0029] Among them, data item E data Taking the significance value λ calculated in step 4 i as the coordination factor, which consists of the elevation difference between the elevation value of the candidate segmentation plane and the elevation value of the current supervoxel and the penalty term of the optimal elevation plane passing through the supervoxel, and is expressed as:

[0030]

[0031] Smoothing term E smooth is composed of the absolute value of the elevation difference between the current supervoxel S i and the elevation of the optimal filtering plane of the adjacent previous supervoxel S i-1 in the scanning direction, and is expressed as:

[0032]

[0033] Step 5.2, Accumulate the energy functions of N directions of each supervoxel unit and minimize this function to solve the optimal filtering plane. Based on the semi-global optimization idea, the solution result of this filtering plane is approximately the global optimal filtering plane, that is

[0034]

[0035] Step 5.3, According to the optimal filtering plane, make a point-by-point judgment on the supervoxel point cloud according to the following calculation method. If the elevation value of this point is less than the elevation value of the filtering plane of the supervoxel where it is located, or the absolute value of the difference between it and the filtering plane of the supervoxel where it is located is less than the discretization unit d, then this point is a ground point; otherwise, this point is a non-ground point.

[0036] Furthermore, d is set to be 0.2 times the segmentation threshold ΔH, and the value of m is set to 5.

[0037] The present invention also provides an airborne LiDAR point cloud filtering device based on supervoxel ground significance, including the following modules:

[0038] A preprocessing module, used to load point cloud data and perform preprocessing, and use a statistical filter to delete obvious outlier noise points in the LiDAR point cloud;

[0039] A supervoxel segmentation module, used to perform supervoxel segmentation on the preprocessed point cloud. The supervoxel resolution r is input externally, generate a supervoxel point cloud cluster that maintains the original local boundary, and calculate the centroid coordinates of each supervoxel unit;

[0040] A scanning strip division module, used to divide the supervoxels into N-direction scanning strips according to the centroid coordinates of the supervoxels;

[0041] A significance value calculation module, which is used to sort and segment the scan strips in each direction and calculate the calculated significance value of each segment;

[0042] A filtering module, which is used to use the significance value as a coordination factor, construct an energy function of an optimal filtering plane and minimize the energy function, solve the optimal segmentation plane, classify ground points and non-ground points according to the segmentation plane, and determine the point cloud filtering result.

[0043] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0044] 1) Based on the segmentation method, the present invention clusters the original point cloud into point cloud clusters with supervoxels as units. Compared with most current point cloud filtering methods based on regular grids, the supervoxel unit can preserve the boundary information of the original target and is applicable to complex environments such as steep slopes and woodlands. The significance calculated with supervoxels as units can preliminarily filter the target point cloud in the vertical direction and better distinguish ground points from non-ground points.

[0045] 2) In view of the different terrain characteristics of urban mountainous areas, the present invention introduces a multiple iteration filtering mechanism in the significance calculation stage, effectively preventing the phenomenon that the sunken structure on the top floor of a building is easily misclassified as a ground point, and enhancing the robustness of the filtering method. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] The present invention will be further described below in conjunction with the drawings and embodiments. In the drawings:

[0047] Figure 1 is a flowchart of the present invention.

[0048] Figure 2 is a schematic diagram of all scanning directions of the supervoxel of the present invention and the division of the scan strips in a single direction.

[0049] Figure 3 is a schematic diagram of the segmented ground significance calculation in the elevation direction. DETAILED DESCRIPTION OF THE INVENTION

[0050] In order to make the purpose, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0051] In order to solve the problem that the existing point cloud filtering method based on regular grids is prone to incorrect filtering in scenarios with large terrain feature differences, the technical solutions of the present invention will be specifically described below in conjunction with the drawings and embodiments.

[0052] Such as Figure 1As shown in the figure, an airborne LiDAR point cloud filtering method based on supervoxel ground saliency provided by an embodiment of the present invention includes the following steps:

[0053] Step 1: Load the point cloud data and perform preprocessing, and use a statistical filter to delete obvious outlier noise points in the LiDAR point cloud.

[0054] Step 2: For the preprocessed point cloud, the present invention uses a boundary-preserving TBBP supervoxel algorithm to implement supervoxel segmentation. The supervoxel resolution r is input externally, generating a supervoxel point cloud cluster that preserves the original local boundary, and calculating the centroid coordinates of each supervoxel unit.

[0055] Step 3: Divide the supervoxels into scan bands in N directions according to the centroid coordinates of the supervoxels. The division of the scan band directions is as Figure 2 shown. The specific method is as follows:

[0056] Step 3.1: In the two-dimensional XOY plane, starting from the y-axis and rotating clockwise around the origin, with every 22.5° as a scan direction, finally generating N equally divided scan directions in the two-dimensional plane.

[0057] The value of N is a constant. The larger the value, the more directions and the better the filtering smoothness, but the computational amount will also increase relatively. In this embodiment, N is taken as 16. Taking 16 scan bands can minimize the computational amount on the premise of ensuring the filtering accuracy.

[0058] Step 3.2: In each scan direction, using the supervoxel resolution r as the translation unit, generating multiple parallel lines passing through the target point cloud in the two-dimensional XOY plane, and dividing the supervoxels into multiple scan bands according to the centroid coordinates of the supervoxels.

[0059] Step 4: Sort, segment, and calculate the saliency value for each direction's scan band. The schematic diagram of the iterative scan saliency calculation is shown in Figure 3 , and the specific method is as follows:

[0060] Step 4.1: Within each scan band, sort the supervoxel units in ascending order according to the x coordinate of the supervoxel center. In particular, for the bands in the scan direction parallel to the y-axis, sort them in ascending order according to the y coordinate of the supervoxel center. Set the elevation value of each supervoxel to the elevation value of the lowest point within the supervoxel. The initial saliency value λ i of the supervoxel S i is set to 1 for all.

[0061] In step 4.2, the absolute value of the elevation difference between two adjacent supervoxels in the sorting direction is calculated. If the absolute value is less than the segment elevation threshold ΔH, the supervoxel is assigned to the current segment; otherwise, it is assigned to the next segment, ultimately completing the segmentation of all supervoxel units along the scan line. The elevation threshold ΔH is an external input parameter determined by the average error limit between the filtered ground points and the resulting DEM.

[0062] Step 4.3: After the segmentation is completed, the elevation difference between the last supervoxel in the current segment and the first supervoxel in the next adjacent segment is calculated according to the scanning order. If the elevation difference is greater than 2 times the segmentation threshold ΔH, the significance value λ of all supervoxel units in the segment is i For segments whose significance has not changed in this scan, the above significance calculation steps are repeated. The number of iterations is determined by the actual terrain. It is set to 1 for areas with large terrain undulations and 3 for areas with flat terrain.

[0063] Step 4.4: perform steps 4.2 and 4.3 on N scanning directions respectively to obtain each supervoxel unit S i The significance value λ i .

[0064] Step 5: Use the saliency value as a coordination factor to construct the energy function of the optimal filtering plane to solve the optimal segmentation plane and complete the point cloud filtering. The specific steps are as follows:

[0065] Step 5.1, calculate each supervoxel unit S i Minimum elevation of the point cloud within As the elevation value of the current supervoxel, H min To obtain the minimum elevation value of the entire input point cloud, the supervoxel point cloud is discretized in the elevation direction:

[0066]

[0067]

[0068] In the above formula, N z is the maximum value of the interval that the filter plane can take after the elevation discretization of the current supervoxel, d is the unit height of the elevation discretization, and its size is set to 0.2 times the segmentation threshold ΔH; m is the elevation compensation coefficient, which is generally set to 5; Supervoxel S i The elevation value of the candidate filtering plane, n is [0,N z ] Any integer in the interval; Supervoxel S i The elevation value of the candidate filtering plane The collection of supervoxels S on the t-th scan strip iThe energy function is expressed as:

[0069] E t (S i ) = E data (S i ) + E smooth (S i-1 , S i )

[0070] Among them, the data item E data uses the significance value λ calculated in step 4.4 i as a coordination factor, which consists of the elevation difference between the elevation value of the candidate filtering plane and the elevation value of the current supervoxel and the penalty term for the optimal elevation plane passing through the supervoxel, and is expressed as:

[0071]

[0072] The smooth term E smooth is composed of the absolute value of the elevation difference between the optimal filtering plane of the current supervoxel S i and the optimal filtering plane elevation of the adjacent previous supervoxel S i-1 in the scanning direction, and is expressed as:

[0073]

[0074] Step 5.2, accumulate the energy functions in N directions of each supervoxel unit and minimize this function to solve the optimal filtering plane. Based on the semi-global optimization idea, the solution result of this filtering plane is approximately the global optimal filtering plane, that is

[0075]

[0076] Step 5.3, make a point-by-point judgment on the supervoxel point cloud according to the following calculation method based on the optimal filtering plane. If the elevation value of this point is less than the elevation value of the filtering plane of the supervoxel where it is located, or the absolute value of the difference between it and the filtering plane of the supervoxel where it is located is less than the discretization unit d, then this point is a ground point; otherwise, this point is a non-ground point.

[0077] The present invention also provides an airborne LiDAR point cloud filtering device based on supervoxel ground significance, including the following modules:

[0078] A preprocessing module, which is used to load point cloud data and perform preprocessing, and use a statistical filter to delete obvious outlier noise points in the LiDAR point cloud;

[0079] A supervoxel segmentation module, which is used to perform supervoxel segmentation on the preprocessed point cloud. The supervoxel resolution r is input externally, generate a supervoxel point cloud cluster that maintains the original local boundary, and calculate the centroid coordinates of each supervoxel unit;

[0080] A scan band division module, configured to divide supervoxels into scan bands in N directions according to the centroid coordinates of the supervoxels;

[0081] A significance value calculation module, configured to sort and segment the scan bands in each direction, and calculate the calculated significance value of each segment;

[0082] A filtering module, configured to use the significance value as a coordination factor, construct an energy function of an optimal filtering plane and minimize the energy function, solve the optimal segmentation plane, classify ground points and non-ground points according to the segmentation plane, and determine the point cloud filtering result.

[0083] The specific implementation manners of the modules are the same as those of the steps, and the present invention will not describe them.

[0084] It should be understood that those of ordinary skill in the art can make improvements or transformations according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. An airborne LiDAR point cloud filtering method based on supervoxel ground saliency, characterized in that It includes the following steps: Step 1: Load the point cloud data and perform preprocessing. Use a statistical filter to remove obvious outlier noise points from the LiDAR point cloud. Step 2: Perform supervoxel segmentation on the preprocessed point cloud. The supervoxel resolution r is input externally. Generate supervoxel point cloud clusters that maintain the original local boundaries, and calculate the centroid coordinates of each supervoxel unit. Step 3: Divide the supervoxels into scan bands in N directions according to the centroid coordinates of the supervoxels. Step 4: Sort and segment the scan bands in each direction, and calculate the significance value of each segment. Step 5: Use the significance value as a coordination factor, construct an energy function for the optimal filtering plane and minimize this energy function to solve the optimal segmentation plane. Classify the ground points and non-ground points according to the segmentation plane to determine the point cloud filtering result. In the above Step 5, the specific method for constructing the energy function is as follows: Step 5.1, calculate the minimum elevation of the point cloud within each supervoxel unit as the elevation value of the current supervoxel, which is the minimum elevation of the entire input point cloud, discretize the supervoxel point cloud in the elevation direction: ​ In the above formula, is the maximum value that the filtered plane can take in the interval after the elevation discretization of the current supervoxel, is the unit height of the elevation discretization; m is the elevation compensation coefficient; is the supervoxel the elevation value of the candidate filtering plane, and n is any integer within the range; is the supervoxel the elevation value of the candidate filtering plane the set of, and the energy function of the supervoxel on the t-th scan strip is expressed as: Among them, the data item Use the significance value calculated in step 4 As the coordination factor, it is composed of the elevation value of the candidate segmentation plane And the elevation value of the current supervoxel The elevation difference and the penalty term of the optimal elevation plane passing through the supervoxel are expressed as: Smoothing term Composed of the absolute value of the elevation difference between the optimal filtering plane of the current supervoxel and the adjacent previous supervoxel in the scanning direction and is expressed as: Step 5.2: Accumulate the energy functions in N directions of each supervoxel unit and minimize this function to solve the optimal filtering plane. Based on the semi-global optimization idea, the solution result of this filtering plane is approximately the global optimal filtering plane, that is Step 5.3, according to the optimal filtering plane, each point in the supervoxel point cloud is judged point by point according to the following calculation method. If the elevation value of the point is less than the elevation value of the filtering plane of the supervoxel where it is located, or the absolute value of the difference between it and the filtering plane of the supervoxel where it is located is less than the discretization unit , then the point is a ground point; otherwise, the point is a non-ground point.

2. The airborne LiDAR point cloud filtering method based on supervoxel ground saliency according to claim 1, wherein: In Step 2, a boundary-preserving TBBP supervoxel algorithm is used to implement supervoxel segmentation.

3. The airborne LiDAR point cloud filtering method based on supervoxel ground saliency according to claim 1, characterized in that: The value of N is 16.

4. The airborne LiDAR point cloud filtering method based on supervoxel ground saliency according to claim 1, wherein: In the above Step 3, the specific method for scan band division is as follows: Step 3.1: In the two-dimensional XOY plane, starting from the y-axis, rotate clockwise around the origin. Every 22.5° is a scan direction, and finally 16 evenly divided scan directions are generated in the two-dimensional plane. Step 3.2: In each scan direction, using the supervoxel resolution r as the translation unit, generate multiple parallel lines passing through the target point cloud in the two-dimensional XOY plane, and divide the supervoxels into multiple scan bands according to the supervoxel centroid coordinates.

5. The airborne LiDAR point cloud filtering method based on supervoxel ground saliency according to claim 1, characterized in that: In the above Step 4, the specific method for strip segmentation and significance calculation is as follows: Step 4.1: Within each scan band, sort the supervoxel units in ascending order according to the x coordinate of the supervoxel center. For the bands in the scan direction parallel to the y-axis, sort them in ascending order according to the y coordinate of the supervoxel center. Set the elevation value of each supervoxel to the elevation value of the lowest point within this supervoxel, and set the initial significance value of each supervoxel to 1. Step 4.2, calculate the absolute value of the elevation difference between two adjacent supervoxels in the sorting direction. If the absolute value is less than the segmented elevation threshold , then classify this supervoxel into the current segment; otherwise, classify it into the next segment, and finally complete the segmentation of all supervoxel units on the scan line; Step 4.3, after the segmentation is completed, calculate the elevation difference between the last supervoxel in the current segment and the first supervoxel unit in the adjacent next segment in the scanning order. If the elevation difference is greater than twice the segmentation threshold , subtract 1 / N from the significance values of all supervoxel units in this segment. For the segments whose significance has not changed in this scan, iterate the above significance calculation steps; Step 4.4, perform the operations of Step 4.2 and Step 4.3 on N scanning directions respectively to obtain the significance value of each supervoxel unit .

6. The airborne LiDAR point cloud filtering method based on supervoxel ground saliency according to claim 5, characterized in that: Elevation threshold in Step 4.2 Input externally and determined by the average error limit between the filtered ground points and the DEM generated therefrom.

7. The airborne LiDAR point cloud filtering method based on supervoxel ground saliency according to claim 5, wherein: The number of iterations in Step 4.3 is determined by the actual terrain, set to 1 for areas with large terrain undulations and 3 for flat terrain areas.

8. The airborne LiDAR point cloud filtering method based on supervoxel ground saliency according to claim 1, wherein: is set to 0.2 times the segmentation threshold and the value of m is set to 5.

9. An airborne LiDAR point cloud filtering device based on supervoxel ground saliency, characterized in that, It includes the following modules: A preprocessing module, which is used to load the point cloud data and perform preprocessing, and use a statistical filter to remove obvious outlier noise points from the LiDAR point cloud. A supervoxel segmentation module, which is used to perform supervoxel segmentation on the preprocessed point cloud. The supervoxel resolution r is input externally. Generate supervoxel point cloud clusters that maintain the original local boundaries, and calculate the centroid coordinates of each supervoxel unit. A scan band division module, which is used to divide the supervoxels into scan bands in N directions according to the centroid coordinates of the supervoxels. A significance value calculation module, which is used to sort and segment the scan bands in each direction and calculate the significance value of each segment. A filtering module, which is used to take the saliency value as a coordination factor, construct an energy function of an optimal filtering plane and minimize the energy function, solve the optimal segmentation plane, classify ground points and non-ground points according to the segmentation plane, and determine the point cloud filtering result; The specific method for constructing the energy function is as follows: Step 5.1, calculate the minimum elevation of the point cloud within each supervoxel unit as the elevation value of the current supervoxel and, for the minimum elevation of the entire input point cloud, discretize the supervoxel point cloud in the elevation direction: ​ In the above formula, is the maximum value that the filtered plane can take in the interval after the elevation discretization of the current supervoxel, is the unit height of the elevation discretization; m is the elevation compensation coefficient; is the supervoxel elevation value of the candidate filtering plane, and n is any integer within the range; is the supervoxel elevation value of the candidate filtering plane set, and the energy function of the supervoxel on the t-th scan strip is expressed as: Among them, the data item Use the significance value calculated in step 4 As the coordination factor, it is composed of the elevation value of the candidate segmentation plane And the elevation value of the current supervoxel The elevation difference and the penalty term of the optimal elevation plane passing through the supervoxel are expressed as: Smoothing term Composed of the absolute value of the elevation difference between the optimal filtering plane of the current supervoxel and the adjacent previous supervoxel in the scanning direction and is expressed as: Step 5.2, accumulate the energy functions in N directions of each supervoxel unit and minimize the function to solve the optimal filtering plane. Based on the semi-global optimization idea, the solution result of this filtering plane is approximately the global optimal filtering plane, that is Step 5.3, according to the optimal filtering plane, make a point-by-point judgment on the supervoxel point cloud according to the following calculation method. If the elevation value of the point is less than the elevation value of the filtering plane of the supervoxel where it is located, or the absolute value of the difference between it and the filtering plane of the supervoxel where it is located is less than the discretization unit , then the point is a ground point; otherwise, the point is a non-ground point.

Images