Large-range forest point cloud individual tree segmentation method based on progressive greedy algorithm

Abstract

The invention discloses a large-range forest point cloud individual tree segmentation method based on a progressive greedy algorithm, and the method comprises the steps: obtaining airborne LiDAR point cloud data, and filtering ground points in the point cloud through employing cloth simulation filtering; voxelization: calculating a voxel key value corresponding to each point cloud and creating a corresponding voxel by taking an AABB bounding box of the tree point cloud as a boundary, encoding the voxel key value into a hash index by adopting a spatial hash function, and storing a pointer of the corresponding voxel on an index address; performing progressive greedy segmentation; carrying out post-treatment; segmenting the point clouds according to the point clouds contained in each voxel and the single tree clusters to which the voxels belong; and outputting the point cloud of each tree. According to the method, a bottom-up segmentation strategy is adopted, inaccurate bottom-layer information is gradually transmitted to accurate top-layer information and then segmentation is carried out, starting from canopy overall distribution characteristics instead of local highest points is carried out for the top-layer information, difficult seed point identification and crown breadth detection steps are avoided, and rapid greedy individual tree segmentation is realized.

Description

Technical Field

[0001] The invention belongs to the technical field of point cloud data processing, and in particular relates to a large-scale forest point cloud single tree segmentation method based on a progressive greedy algorithm. Background Art

[0002] In the past two decades, the rapid development of LiDAR technology can describe the 3D structure of trees and forests in detail with unprecedented accuracy. Compared with time-consuming and labor-intensive manual inventory, airborne platforms equipped with LiDAR sensors can efficiently complete large-scale forest surveys without being restricted by harsh environments such as terrain; at the same time, compared with traditional optical remote sensing methods that can only obtain forest surface information, it can obtain the complete canopy 3D structure and can penetrate the forest canopy to a certain extent to accurately obtain some understory information. Therefore, airborne LiDAR surveys have become a research hotspot in recent years. However, in order to obtain the attribute data of each tree, it is first necessary to be able to accurately and automatically separate individual trees in the vertical and horizontal directions.

[0003] The single tree segmentation algorithm can identify and segment independent single trees from point cloud data, but it is a huge challenge to segment all single trees completely and accurately. To this end, many scholars have proposed a large number of single tree segmentation algorithms, which can be divided into raster-based segmentation, point cloud-based segmentation and joint segmentation according to the type of data they process. However, the raster-based segmentation and point cloud-based segmentation methods rely on the selection of seed points. The accuracy of seed point recognition largely determines the accuracy of the final segmentation result, and it is difficult to guarantee the segmentation accuracy; while the joint segmentation method often sacrifices the segmentation efficiency and has high algorithm complexity. In addition, the point cloud elevation normalization preprocessing is required before executing the joint segmentation method, but it is difficult to obtain accurate understory terrain in dense forest areas, which leads to deformation of the canopy after normalization, affecting the segmentation accuracy. Therefore, how to balance segmentation accuracy and efficiency for single tree segmentation is the current problem. Summary of the invention

[0004] The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art and propose a large-scale forest point cloud single tree segmentation method based on a progressive greedy algorithm.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] The single tree segmentation method of large-scale forest point cloud based on progressive greedy algorithm includes the following steps:

[0007] Obtain airborne LiDAR point cloud data and use cloth simulation filtering to filter out ground points in the point cloud;

[0008] Voxelization, using the AABB bounding box of the tree point cloud as the boundary, calculate the voxel key value corresponding to each point cloud and create the corresponding voxel, use the spatial hash function to encode the voxel key value into a hash index and store the pointer of the corresponding voxel at the index address;

[0009] Gradual greedy segmentation, obtain candidate single tree clusters, and determine whether each candidate single tree cluster is the correct single tree according to the established single tree judgment rules;

[0010] Post-processing: calculate the minor axis length of each single tree cluster and determine whether it is an over-segmented single tree cluster based on the preset minimum single tree radius. If it is over-segmented, try to merge it into the adjacent single tree cluster;

[0011] Segment the point cloud according to the point cloud contained in each voxel and the single tree cluster to which the voxel belongs;

[0012] Output a point cloud for each tree.

[0013] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0014] 1. The present invention adopts a bottom-up segmentation strategy to avoid dependence on seed points; the inaccurate bottom-level information is gradually transferred to the accurate top-level information before segmentation, and the top-level information is based on the overall distribution characteristics of the canopy rather than the local highest point, avoiding difficult seed point identification and crown width detection steps, and realizing fast greedy single tree segmentation.

[0015] 2. The present invention applies the greedy idea to the field of single tree segmentation, regards each point as a computing node, and actively searches for the single tree vertex that is most likely to belong to without relying on the seed point.

[0016] 3. In order to take into account the complex forest stand conditions such as mixed forests, the present invention adopts a progressive segmentation strategy from large to small crown width to reduce over-segmentation and under-segmentation.

[0017] 4. The present invention adopts a sparse voxel tissue segmentation process to avoid constructing invalid voxels in pointless areas, and minimizes memory and time consumption while making full use of 3D information. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 is a flow chart of the method of the present invention;

[0019] Figure 2 It is a schematic diagram for constructing sparse voxels;

[0020] Figure 3 This is the greedy segmentation implementation flow chart;

[0021] Figure 4 It is the principle diagram of fitting residual calculation;

[0022] Figure 5 This is a schematic diagram of the calculation principle of the major and minor axis ratio and the relative position of the tree vertices. DETAILED DESCRIPTION

[0023] The present invention is further described in detail below in conjunction with embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0024] Example

[0025] like Figure 1 As shown, the present invention, a large-scale forest point cloud single tree segmentation method based on a progressive greedy algorithm, comprises the following steps:

[0026] S1, obtain airborne LiDAR point cloud data, and use cloth simulation filtering to filter out ground points in the point cloud;

[0027] S2. Traditional voxelization methods require the entire point cloud bounding box to be divided into voxels, which results in a large waste of memory and computing resources. To solve this problem, a sparse voxel division method is used to avoid constructing invalid voxels in pointless areas. Specifically, sparse voxels only store voxels with points to save memory; and a hash function is used to encode the spatial index of the voxel to ensure that the voxel information can be quickly retrieved through the spatial index. In this implementation, specifically:

[0028] Voxelization, using the AABB bounding box (Axis-Aligned Bounding Box) of the tree point cloud as the boundary, calculates the voxel key value corresponding to each point cloud and creates the corresponding voxel; Figure 2 As shown, a spatial hash function is used to encode the voxel key value into a hash index and a pointer to the corresponding voxel is stored at the index address.

[0029] Due to the influence of the regular arrangement of voxels, when performing greedy segmentation, any voxel with the same elevation will be randomly selected as the attribute, which will lead to greater uncertainty and error in the segmentation result. Therefore, this embodiment calculates the centroid of each voxel according to the coordinates of the point cloud contained in each voxel, and defines the voxel position by the centroid of the voxel. The calculation formula is:

[0030]

[0031] Where c is the centroid of the body, n is the number of points contained in the voxel; (x i ,y i , z i ) is the coordinate of the ith point in the voxel. The voxelized resolution V is related to the point density and forest characteristics. When the point density is low, the resolution V should be appropriately reduced to ensure that the voxels of the same tree are interconnected. However, too small a resolution may smooth out the features between adjacent trees, resulting in under-segmentation problems in subsequent segmentation. For point densities greater than 400 pts / m 2For scenes with a canopy width greater than 1m, the resolution V can be 0.1m.

[0032] S3: Perform progressive greedy segmentation to obtain candidate single tree clusters and determine whether each candidate single tree cluster is a correct single tree according to the established single tree judgment rule.

[0033] Greedy algorithm is an algorithm that takes the best or optimal (i.e. most favorable) choice in the current state at each step, so as to seek the best or optimal final result. The problem of single tree segmentation in airborne LiDAR point cloud can be abstracted as the optimization problem of finding the most likely single tree for each LiDAR point. Assuming that each tree has only one highest point, the above problem can be transformed into the problem of finding the most likely point for each point. Considering each LiDAR point as a graph node, taking into account the connectivity characteristics of the branches and leaves of a single tree, each node can find a shortest path to reach its highest point, and for the highest point that is not the node's, it is either unreachable because it is not connected, or the graph distance is much greater than the highest point that is correctly attributed. However, since the highest point is unknown, it is impossible to give an exact three-dimensional target, and the shortest graph path method such as Dijkstra is used for global solution. To this end, assuming that the highest point of each tree is at an infinite height, the direction vector of each point and its highest point should be closest to (0, 0, 1) (zenith direction). The above problem can be deduced as each node expects to find the highest point in the direction of the minimum zenith angle under the connectivity constraint. If the greedy idea is used for local solution, the node with the smallest zenith angle connected to each node is found in the local range. When all nodes are calculated, the entire scene will generate multiple directed graphs. The root node of each directed graph is the highest point of a tree, and the directed graph is all the nodes included in a single tree. Figure 3 As shown in the figure, a node P0 located at the junction of a single tree usually has a smaller zenith angle than the line connecting the node P2 on the tree Tree3 to which it belongs, compared with the line connecting the node P1 on the adjacent tree Tree2. Therefore, this method can ensure that the adjacent single trees are separated from the crown depression, while taking into account the connectivity characteristics of the sparse nodes under the forest and retaining more information below the crown of the single tree.

[0034] In this embodiment, the progressive greedy segmentation step includes:

[0035] For any voxel i, first find its corresponding neighboring voxel j through the voxel index according to the maximum greedy range R; then calculate the zenith angle of the line connecting voxel i and each domain voxel j, and point voxel i to the domain voxel j with the smallest zenith angle; after all voxels are calculated, construct a directed graph according to the direction of each voxel, and find all the root nodes (root voxels) in the directed graph; each root node (root voxel) corresponds to a single tree vertex, and all the voxels it connects constitute a candidate single tree cluster;

[0036] Among them, the zenith angle α between any two voxelsi,j It is the angle between the line connecting the two voxels and the zenith direction, which is calculated by the following formula:

[0037]

[0038] in, is the zenith direction, c i and c j are the centroids of the two voxels respectively.

[0039] The parent voxel of each voxel is the neighborhood range R i The neighboring voxel with the smallest inner zenith angle is voxel c. i The parent voxel c father Expressed as:

[0040] c father =c j =argmin(α i,j )wherec j ∈R i

[0041] For any voxel c i , first according to the neighborhood range R i Determine the neighboring voxels that need to be calculated for the zenith angle, and then calculate c i and the zenith angle of each neighboring voxel, and finally c i Point to the voxel c with the smallest zenith angle j , which is c i The parent voxel of

[0042] After the parent voxel search is completed, all voxels have two situations: the parent voxel is found and the parent voxel is not found; the voxel that has not found the parent voxel is the root voxel, that is, the voxel that may be the highest point of a single tree;

[0043] The voxel that finds its parent voxel is the leaf voxel. Recursively searching for the parent voxel along the parent voxel can eventually belong to a root voxel. All voxels belonging to the same root voxel are components of the same single tree. All voxels are divided into different clusters according to different root voxels, and each cluster corresponds to a single tree.

[0044] The surface of a tree crown is usually not a strictly convex surface. The concave surrounding the locally raised branches will cause the branch vertices to be mistakenly identified as root nodes, which will cause the over-segmentation problem of individual trees. Increasing the greedy range R when searching for the parent voxel allows the branch vertices to be merged into the adjacent root voxels, which will be able to avoid this problem. However, a large greedy range may ignore the concave features between adjacent trees, which will cause smaller trees to be mistakenly merged into adjacent trees, causing the under-segmentation problem of individual trees. For multi-layer forests with large variations in crowns and tree heights, it is difficult to simultaneously take into account the over-segmentation and under-segmentation problems of individual trees using a single greedy range for greedy segmentation. To this end, the method of the present invention adopts a progressive segmentation strategy, that is, a series of gradually decreasing greedy range values ​​R are used to perform greedy segmentation. Large greedy range values ​​are used to segment individual trees with larger crown widths, and small greedy ranges are used to segment individual trees with smaller crown widths.

[0045] In this embodiment, the progressive greedy segmentation adopts a series of gradually decreasing greedy range values ​​R i Perform progressive iterative greedy segmentation respectively;

[0046] In the initial iteration, the preset maximum greedy range value R is used to perform greedy segmentation on all voxels, and the characteristics of the single tree cluster after segmentation are calculated. If the single tree judgment rule is met, it is marked as a real single tree and the corresponding voxel is deleted; then the maximum greedy range value R2=R1-V is reduced, and greedy segmentation and single tree judgment are performed again on the remaining voxels. The above process is iterated until the greedy range is the same as the voxel resolution V.

[0047] Accurately identifying the real single trees in the segmentation results during the progressive segmentation process is the key to ensuring the segmentation accuracy of the algorithm. Based on a full analysis of the morphological characteristics of single trees, the present invention integrates and constructs multiple single tree morphological judgment rules. In this embodiment, the formulated single tree judgment rules specifically include surface point fitting residuals, major and minor axis ratios, and relative positions of tree vertices; when the morphological characteristics of the segmented cluster meet the three rules at the same time, the single tree cluster is judged to be a real single tree.

[0048] The residual of surface point fitting is the residual of fitting the surface voxels of each single tree to the Pollock crown model; the Pollock model can describe a variety of crown shapes from cones to ellipsoids and is widely used to express the crown shape of single trees. The objective function of the Pollock crown model is:

[0049]

[0050] Among them, (x j ,y j , z j ) is the coordinate of the crown center of the current single tree, a is the radius along the z-axis, b is the radius in the xy plane, and c is the crown shape parameter;

[0051] The surface point fitting residual is expressed as:

[0052]

[0053] Among them, (x i ,y i , z i ) is the coordinate of the i-th surface point in the current tree; r ij is the fitting residual of this point; Ir j is the internal point rate of the current single tree whose fitting residual is less than the threshold; n is the number of surface points contained in the current single tree. j Less than the specified threshold I th When , the single tree cluster meets the surface point interior point rate requirement.

[0054] The surface fitting residual inlier rate of a real single tree is usually large, while the surface fitting residual inlier rate of two or more single trees that are incorrectly segmented into a cluster is usually small, such as Figure 4 shown.

[0055] The ratio of the major and minor axes is used to determine whether the shape of a single tree is close to a circle; generally speaking, the ratio of the major and minor axes of the projection of a single tree crown point on the xy plane should be close to 1. Considering that the calculation of the OBB bounding box of the single tree projection point is easily affected by noise, which interferes with the calculation results of the major and minor axes, the ratio of the first and second eigenvalues ​​of the projection point is used to approximate the ratio of the major and minor axes. First, the covariance matrix C of the single tree projection point is calculated. 2×2 :

[0056]

[0057] in, is the centroid of the single tree projection point, m is the number of all voxels contained in the cluster; c i Represents the centroid of a voxel.

[0058] Then singular value decomposition (SVD) is used to decompose the covariance matrix to obtain the λ1 and λ2 eigenvalues, and the major-minor axis ratio ls is defined as:

[0059]

[0060] The threshold of ls is set to 2. If it is less than 2, the tree is considered to meet the major and minor axis constraints; otherwise, the tree is considered to be segmented incorrectly.

[0061] The relative position of the tree vertex is defined as whether the tree vertex is located at Figure 5 If the tree is in the blue box, the tree is considered to have a reasonable shape. Otherwise, the cluster segmentation is wrong.

[0062] Calculate the v1 and v2 eigenvectors of the single tree projection point, and calculate the projection coordinates proj of all points in the cluster on v1 and v2 i =[v1·(x i ,y i ),v2·(x i ,y i )], and obtain the maximum and minimum values ​​v after projection 1_min , v 1_max , v 2_min and v 2_max , then the coordinates of the highest point of the cluster (tree vertex) after projection [z_max v1 ,z_max v2 ]The following conditions should be met:

[0063]

[0064] The three judgment rules strictly judge individual trees from three aspects: surface point distribution, overall crown shape, and highest point position, which can avoid the problem of under-segmentation of individual trees as much as possible.

[0065] S4, post-processing, calculating the minor axis length of each single tree cluster, and judging whether it is an over-segmented single tree cluster according to the preset minimum single tree radius.

[0066] Affected by noise and extremely protruding branches, even if a progressive strategy is used in the segmentation process, there may still be over-segmentation problems for a few branches. For this reason, after the greedy segmentation is completed, a post-processing step for over-segmentation is introduced, specifically:

[0067] Calculate the short axis length of each single tree cluster according to the set minimum single tree radius R min , determine whether the minor axis length is less than R min Identify clusters of single trees that may be over-segmented;

[0068] If so, find the adjacent single tree cluster, calculate the single tree morphological characteristics after merging the over-segmented single tree cluster into the adjacent single tree cluster, and determine the internal point rate Ir after merging m The original interior point rate Ir of the adjacent single tree cluster o Does the relationship between them satisfy Ir o -Ir m <0.2×Ir o , whether the ratio of the major and minor axes and the relative position of the tree vertices meet the single tree judgment rule;

[0069] If all the conditions are met, the single tree cluster is merged into the adjacent single tree cluster, otherwise it is classified as a noise cluster; if the requirements are met with multiple adjacent single tree clusters, it is merged into the single tree cluster with the largest inlier rate.

[0070] S5, segmenting the point cloud according to the point cloud contained in each voxel and the single tree cluster to which the voxel belongs;

[0071] S6. Finally, output the point cloud of each tree.

[0072] It should also be noted that in this specification, terms such as "comprises", "includes" or any other variations thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of more restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0073] The above description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A large-scale forest point cloud single tree segmentation method based on a progressive greedy algorithm, characterized by: The following steps are involved: Obtain airborne LiDAR point cloud data and use cloth simulation filtering to filter out ground points in the point cloud; Voxelization, using the AABB bounding box of the tree point cloud as the boundary, calculate the voxel key value corresponding to each point cloud and create the corresponding voxel, use the spatial hash function to encode the voxel key value into a hash index and store the pointer of the corresponding voxel at the index address; Gradual greedy segmentation, obtain candidate single tree clusters, and determine whether each candidate single tree cluster is the correct single tree according to the established single tree judgment rules; Post-processing: calculate the minor axis length of each single tree cluster and determine whether it is an over-segmented single tree cluster based on the preset minimum single tree radius. If it is over-segmented, try to merge it into the adjacent single tree cluster; Segment the point cloud according to the point cloud contained in each voxel and the single tree cluster to which the voxel belongs; Output a point cloud for each tree.

2. The large-scale forest point cloud single tree segmentation method based on the progressive greedy algorithm according to claim 1 is characterized in that: Voxelization also includes calculating the centroid of each voxel based on the coordinates of the point cloud contained in each voxel, and defining the voxel position by the voxel centroid. The calculation formula is: Where c is the centroid of the body, n is the number of points contained in the voxel; (x i ,y i , z i ) is the coordinate of the i-th point in the voxel.

3. The large-scale forest point cloud single tree segmentation method based on the progressive greedy algorithm according to claim 2 is characterized in that: Progressive greedy segmentation includes: For any voxel i, first find its corresponding neighboring voxel j through the voxel index according to the maximum greedy range R; then calculate the zenith angle of the line connecting voxel i and each domain voxel j, and point voxel i to the domain voxel j with the smallest zenith angle; after all voxels are calculated, construct a directed graph according to the direction of each voxel, and find all the root nodes in the directed graph; each root node corresponds to a single tree vertex, and all the voxels it connects constitute a candidate single tree cluster; Among them, the zenith angle α between any two voxels i,j It is the angle between the line connecting the two voxels and the zenith direction, which is calculated by the following formula: in, is the zenith direction, c i and c j are the centroids of the two voxels respectively.

4. The large-scale forest point cloud single tree segmentation method based on the progressive greedy algorithm according to claim 3 is characterized in that: The parent voxel of each voxel is the neighborhood range R i The neighboring voxel with the smallest inner zenith angle is voxel c. i The parent voxel c father Expressed as: c father =c j =argmin(α i,j )wherec j ∈R i For any voxel c i , first according to the neighborhood range R i Determine the neighboring voxels that need to be calculated for the zenith angle, and then calculate c i and the zenith angle of each neighboring voxel, and finally c i Point to the voxel c with the smallest zenith angle j , which is c i The parent voxel of After the parent voxel search is completed, all voxels have two situations: the parent voxel is found and the parent voxel is not found; the voxel that has not found the parent voxel is the root voxel, that is, the voxel that may be the highest point of a single tree; The voxel that finds its parent voxel is the leaf voxel. Recursively searching for the parent voxel along the parent voxel can eventually belong to a root voxel. All voxels belonging to the same root voxel are components of the same single tree. All voxels are divided into different clusters according to different root voxels, and each cluster corresponds to a single tree.

5. The large-scale forest point cloud single tree segmentation method based on progressive greedy algorithm according to claim 1 is characterized in that: Progressive greedy partitioning uses a series of gradually decreasing greedy range values ​​R i Perform progressive iterative greedy segmentation respectively; In the initial iteration, the preset maximum greedy range R is used to perform greedy segmentation on all voxels, and the characteristics of the single tree cluster after segmentation are calculated. If the single tree judgment rule is met, it is marked as a real single tree and the corresponding voxel is deleted; then the maximum greedy range value R2=R1-V is reduced, and greedy segmentation and single tree judgment are performed again on the remaining voxels. The process is iterated until the greedy range is the same as the voxel resolution V.

6. The large-scale forest point cloud single tree segmentation method based on progressive greedy algorithm according to claim 1 is characterized in that: The single tree judgment rules formulated specifically include the surface point inside point ratio, the major and minor axis ratio, and the relative position of the tree vertices; when the morphological characteristics of the segmented cluster meet the three rules at the same time, the single tree cluster is judged to be a real single tree.

7. The large-scale forest point cloud single tree segmentation method based on the progressive greedy algorithm according to claim 6 is characterized in that: The surface point fitting residual is the fitting residual of the surface voxel of each single tree as the Pollock crown model. The objective function of the Pollock crown model is: Among them, (x j ,y j , z j ) is the coordinate of the crown center of the current single tree, a is the radius along the z-axis, b is the radius in the xy plane, and c is the crown shape parameter; The surface point fitting residual is expressed as: Among them, (x i ,y i , z i ) is the coordinate of the i-th surface point in the current tree; r ij is the fitting residual of this point; Ir j is the inlier rate of the current single tree whose fitting residual is less than the threshold; n is the number of surface points contained in the current single tree. j Less than the specified threshold I th When , the single tree cluster meets the surface point interior point rate requirement.

8. The large-scale forest point cloud single tree segmentation method based on progressive greedy algorithm according to claim 6 is characterized in that: The ratio of the major and minor axes is used to determine whether the shape of a single tree is close to a circle. The ratio of the first and second eigenvalues ​​of the projection point is used to approximate the ratio of the major and minor axes. First, the covariance matrix C of the single tree projection point is calculated. 2×2 : in, is the centroid of the single tree projection point, m is the number of all voxels contained in the cluster; c i represents the center of mass of a voxel; Then singular value decomposition SVD is used to decompose the covariance matrix to obtain the λ1 and λ2 eigenvalues, and the major-minor axis ratio ls is defined as: The threshold of ls is set to 2. If it is less than 2, the tree is considered to meet the major and minor axis constraints; otherwise, the tree is considered to be segmented incorrectly.

9. The large-scale forest point cloud single tree segmentation method based on progressive greedy algorithm according to claim 6, characterized in that: The relative positions of the tree vertices are as follows: Calculate the v1 and v2 eigenvectors of the single tree projection point, and calculate the projection coordinates proj of all points in the cluster on v1 and v2 i =[v1·(x i ,y i ),v2·(x i ,y i )], and obtain the maximum and minimum values ​​v after projection 1_min , v 1_max , v 2_min and v 2_max , then the coordinates of the highest point of the cluster after projection [z_max v1 ,z_max v2 ]The following conditions should be met:

10. The large-scale forest point cloud single tree segmentation method based on progressive greedy algorithm according to claim 1, characterized in that: The post-processing is as follows: Calculate the short axis length of each single tree cluster according to the set minimum single tree radius R min , determine whether the minor axis length is less than R min Identify clusters of single trees that may be over-segmented; If so, find the adjacent single tree cluster, calculate the single tree morphological characteristics after merging the over-segmented single tree cluster into the adjacent single tree cluster, and determine the internal point rate Ir after merging m The original interior point rate Ir of the adjacent single tree cluster o Does the relationship between them satisfy Ir o -Ir m <0.2×Ir o , whether the ratio of the major and minor axes and the relative position of the tree vertices meet the single tree judgment rule; If all the conditions are met, the single tree cluster is merged into the adjacent single tree cluster, otherwise it is classified as a noise cluster; if the requirements are met with multiple adjacent single tree clusters, it is merged into the single tree cluster with the largest inlier rate.

Images