Urban green space scene laser radar point cloud single tree segmentation method

Abstract

The present application relates to a kind of urban green space scene laser radar point cloud single tree segmentation method, the method comprises: the point cloud of urban green space is carried out grid cubic processing;From the initial trunk position in grid is determined to fit DEM elevation plane, from DEM elevation plane upwards cutting tree bottom point to determine tree position;Using K-means clustering to carry out rough segmentation;Then according to single tree position, establish tree adjacent relationship, using ring extraction to carry out single tree optimization segmentation.The present application can make the trunk point cloud feature more obvious compared to tree crown point cloud, and the ground elevation information of the position where the trunk is located can be obtained, and compared with clustering time efficiency, it is based on the trunk point cloud and elevation information obtained to carry out a series of operations to complete rough single tree segmentation, and using adjacent relationship and ring extraction to optimize segmentation result, can make the accuracy of identification result higher.

Description

Technical Field

[0001] This invention relates to the field of point cloud data processing technology, and in particular to a method for segmenting individual trees in a LiDAR point cloud of an urban green space landscape scene. Background Technology

[0002] LiDAR (Light Detection and Ranging) is a remote sensing technology that uses laser beams to detect features such as the position and velocity of targets. It combines laser, GPS, and inertial measurement unit (IMU) technologies to rapidly acquire high-precision three-dimensional spatial data, playing a crucial role in multiple fields.

[0003] In the field of forestry resource monitoring and ecological research, individual tree segmentation is a key technology for quantifying urban green space structure and assessing tree growth status, and its accuracy directly affects the reliability of research conclusions. Therefore, individual tree segmentation of point cloud data has become an important research direction. However, due to the dense distribution and diverse morphologies of trees in a whole forest LiDAR point cloud data set, as well as the complex problems of noise, uneven density, and missing data in the point cloud data, the results of individual tree segmentation are often unsatisfactory. In the processing of whole forest point cloud data, tree extraction is mainly based on rule-based methods and machine learning methods. Among them, machine learning methods are often used for tree point cloud classification; rule-based methods extract the crown and trunk by designing specific rules, thereby enabling the measurement of detailed tree parameters.

[0004] Currently, there are various point cloud segmentation methods, ranging from simple clustering to advanced machine learning techniques that combine unsupervised and supervised methods, including K-means, DBSCAN, graph-based methods, and other machine learning methods. However, over-segmentation and under-segmentation between trees are still common problems in single tree segmentation. In terms of tree parameter extraction, existing methods have limitations in automatically distinguishing the size of individual tree crowns, and they also have insufficient applicability and poor extraction results in single tree segmentation and parameter extraction in urban green space scenarios. Summary of the Invention

[0005] To address the aforementioned problems, this invention provides a method for single-tree segmentation of LiDAR point clouds in urban green space landscape scenes, the method comprising:

[0006] The urban green space tree point cloud is divided into raster cubes, and the initial tree trunk is extracted by comparing neighboring cells. The positional parameters of the raster cube containing the initial tree trunk are fitted to obtain the DEM elevation plane. The point cloud at a certain height above the DEM elevation plane is obtained as the tree trunk point cloud. The tree trunk point cloud is merged and topologically checked. The processed tree trunk point cloud is projected onto the plane, and a fitted circle is calculated on the projected tree trunk point cloud. The center of the fitted circle is recorded as the tree center point for locating the tree position. K-means clustering is used to segment the tree center points to obtain multiple roughly segmented individual tree point clouds. The multiple roughly segmented individual tree point clouds are layered at a certain interval, and each layer is projected. In the projection of each layer, a ring is used to fit the tree edge. The points inside the ring are the core points of the individual tree, and the points outside the ring are the edge points of the individual tree. The tree center points are used as initial data to obtain the proximity relationship between trees. The tree proximity relationship is used to determine the affiliation of the edge points of the individual trees and optimize the segmentation results of easily confused individual trees.

[0007] Furthermore, the step of extracting the initial trunk through neighboring unit comparison includes: calculating the elevation difference between the highest and lowest points in each grid cube and recording it as... ,by To fill the pixel values ​​with a two-dimensional grid, let the side length of the grid cell be... The central grid is The distance from the center grid is less than or equal to The grid is If satisfied Then the central grid is considered The initial tree trunk is included. The corresponding grid cube is extracted, and then the extracted grid cube is divided into horizontal layers with small vertical spacing. If a blank layer appears in the layer, it is considered that the grid cube is affected by some noise points and is not included in the scope of the initial tree trunk points.

[0008] Furthermore, the step of fitting the position parameters of the grid cube containing the initial trunk to obtain the DEM elevation plane includes: statistically extracting the grid cubes containing the initial trunk, recording the position of each grid cube and the minimum elevation value within the cube, and fitting the grid cubes with the position of the grid cubes and the minimum elevation value within the cube to obtain the DEM elevation plane grid.

[0009] Furthermore, the merging and topology checking of the tree trunk point cloud includes: determining whether there are points that are attached to the boundary at the boundary of the tree trunk point cloud grid that can connect two grids; if there are attached points, the two grids should be merged; then determining whether the size of the merged point cloud block exceeds the preset size; if it does, checking whether the adjacent joint point cloud block intersects with or is too close to the boundary of this block; if so, it is considered that the bottom of the tree crown has been cut, and the two blocks are merged.

[0010] Furthermore, the method of fitting tree edges using a ring in the projection of each layer includes: for point clouds with low point density, using a ring to slide and expand from the center to the periphery to extract the point cloud of the tree layer projection until the ring expansion reaches a threshold.

[0011] For point clouds with high point density, a continuousness check is performed by shrinking the projection edges of individual tree layers towards the center point. When the ring meets the preset conditions, the number of ring layers at this point is saved. If the number of layers is found to be [missing information] in subsequent processes The ring, conforms to Furthermore, if it better fits the preset conditions, then it is considered a ring. The optimal segmentation position is determined by using... Overwrite records The search continued until the location was recorded. The tree is divided at its edge.

[0012] Furthermore, the step of using the tree center point as initial data to obtain the proximity relationship between trees includes: traversing the tree center points, establishing a Cartesian coordinate system with one tree center point as the origin each time, determining the quadrant in which other center points are located relative to the origin, calculating the distance relative to the origin, and retaining the index of the nearest tree in each quadrant after anomaly handling as the trees that may have intersecting canopies.

[0013] Furthermore, the step of using the tree proximity relationship to determine the attribution of the edge points of individual trees and optimizing the segmentation results of easily confused individual trees includes:

[0014] For point clouds with low point density, for each edge point, traverse the core points of the same tree and the core points of neighboring trees, find the core point of the tree closest to the edge point, and assign this edge point to the tree to which the found core point belongs. For point clouds with high point density, after finding the assignment of an edge point, record the edge point. If the distance between the next edge point and the current edge point is less than a preset value, then record the next edge point as well. Stop when the distance between a point and the recorded point group is greater than the preset value. Then all recorded edge points are assigned to the tree to which the core point found by the first edge point belongs.

[0015] Furthermore, the optimized easily confused single-tree segmentation results include anomaly handling. The anomaly judgment method is as follows: when the number of points in each layer of the point cloud generated after single-tree segmentation changes with height in an M-shape, it is determined that there is a small tree growing under a large tree and the tree canopy is interfering with the growth. The anomaly handling method is as follows: find the layer with the highest number of points in the lower layer after layer counting, and denot it as... The layer with the most points in the higher-level part is denoted as . Then look for something between and The layer with the fewest points between them is denoted as ,by The tree crowns are divided into large and small tree crowns, and then the points where the large trees are incorrectly divided are extracted and assigned to new trees.

[0016] The present invention also provides a computer storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, it implements the lidar point cloud single tree segmentation method for urban green space landscape scenes as described above.

[0017] This invention divides the 3D point cloud of urban green space into raster cubes, resulting in raster cubes of urban green space. This makes the features of the tree trunk point cloud more obvious compared to the tree crown point cloud, and it can also obtain the ground elevation information of the tree trunk location, while being more time-efficient than clustering. Then, based on the acquired tree trunk point cloud and elevation information, a series of operations are performed to complete coarse tree segmentation, and circular loop extraction is used to optimize the tree segmentation, extracting the core points and edge points of each tree. Combining the proximity relationships of individual trees within urban green space confirms the attribution of tree edge points, solving the problem of erroneous segmentation caused by intersecting tree crowns, thus improving the optimization effect of individual trees. Attached Figure Description

[0018] Figure 1 Overall Flowchart of LiDAR Point Cloud Single Tree Segmentation Method in Urban Green Space Landscape Scene

[0019] Figure 2 Flowchart for determining proximity

[0020] Figure 3 Topology check result diagram

[0021] Figure 4 Segmentation effect diagram Detailed Implementation

[0022] To better understand the above technical solutions, exemplary embodiments of this disclosure will be described in more detail below with reference to the accompanying drawings. While exemplary embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of this disclosure to those skilled in the art.

[0023] As one implementation scheme, Figure 1 This is a flowchart illustrating the overall process of the LiDAR point cloud single-tree segmentation method in an urban green space landscape scenario, as described in the embodiments of this application.

[0024] like Figure 1 As shown, the overall process of this method includes: cubicalization of individual trees in urban green spaces, establishment of the center point of each tree, coarse K-means segmentation, single-circle filtering of the segmented tree results, establishment of tree proximity relationships, optimization of the segmented tree results, and handling of abnormal situations. Each step specifically includes:

[0025] S1, Cubed treatment of urban green space with single wooden grid

[0026] The point cloud of urban green space trees is divided into raster cubes. Specifically, the overall point cloud is divided into raster cells according to the x and y coordinates, and the trees are transformed into raster cubes. The length and width of the raster cubes should be based on the actual spacing between trees.

[0027] S2. Establishment of the center point of a single log

[0028] S201 Extract the initial tree trunk

[0029] The initial raster cube containing the tree trunk is extracted by statistically analyzing its parameters. Based on tree growth characteristics, the elevation difference between the raster cube containing the tree trunk and its neighboring cubes should be at least 1 meter. The elevation difference between the highest and lowest points in each raster cube is then calculated and denoted as follows: ,by To fill the pixel values ​​with a two-dimensional grid, let the side length of the grid cell be... The central grid is The distance from the center grid is less than or equal to The grid is If satisfied Then the central grid is considered If the initial tree trunk is present, the corresponding raster cube is extracted; otherwise, it is considered not to be a raster cube. Center grid elevation difference For grid The elevation difference; in addition, in order to deal with complex situations, there are two other methods for anomaly judgment and handling.

[0030] Neighboring Grid Trunk Determination: This step aims to address the situation where a tree trunk is cut by multiple cubes. Since not every tree trunk will be exactly divided within the same grid cube, it's possible for a tree trunk to be divided into two or more grid cubes. Neighboring grid trunk determination is performed in this case. Assuming the trunk point of the same tree is divided into two neighboring grids, U1 and U2, that is, when U1 is the center, the Δz of U2 will be approximately the same as U1 and greater than that of surrounding grids centered at U1. For the Δz of other grid cubes within the length, it is still assumed that the U1 cube contains the trunk point.

[0031] Continuity assessment: A continuity assessment is performed on the grid cells containing tree trunks extracted from the above steps to eliminate interference from some ground noise points. Specifically, the selected grid cubes containing standard tree trunks are divided into horizontal layers with small vertical spacing. If a blank layer appears in the layering, the grid cube is considered to be affected by some noise points and is not included in the initial tree trunk point range.

[0032] S202 Fitting DEM Plane

[0033] The raster DEM data is fitted based on the parameters of the proposed standard tree trunk cube. The extracted raster cubes containing tree trunks are statistically analyzed, and the location of each raster containing a tree trunk and the minimum elevation value within the cube are recorded. These locations are used to fit a planar grid to record DEM values. The grid is traversed, and the nearest neighbor cube for each grid is found using the nearest neighbor algorithm. The value is used as the fitted DEM plane elevation value.

[0034] S203 Cut-off point

[0035] The base points of the trees are obtained from the fitted DEM elevation plane. Point clouds at a certain height are then obtained from the DEM elevation plane to form the tree trunk point cloud. Small grid cubes at the center points are then obtained.

[0036] S204 base points unite to form the trunk

[0037] Merging is performed on the extracted tree trunks within the grid-like point cloud to resolve situations where a trunk is cut into two grids. During the merging process, it's important to note that if the spacing between adjacent trees is too small, two adjacent grids may contain two tree trunks, leading to merging errors. Therefore, a adhesion check is performed when two adjacent grids contain tree trunks. The check determines if there are points at the grid boundary that adhere to the boundary and can connect the two grids. If such points exist, the two grids are considered to be merged; otherwise, they are not. This part includes a topology check.

[0038] Topology check: This checks for errors where tree trunks are too short, potentially truncating to the bottom of the canopy. The trunk being too short might be due to actual conditions or missing point clouds. When truncating to the bottom of the canopy, due to tree growth characteristics, different branches may create continuous point clouds in one direction. Because gaps exist between branches, the adhesion judgment is not suitable for this situation. This invention adds this module after point cloud merging, using the boundaries of different block-shaped point clouds for further merging. It checks if the size of the merged point cloud block exceeds a preset size. If it does, it checks if neighboring merged point cloud blocks intersect or are too close to the current block's boundary. If so, it's considered that the bottom of the canopy has been truncated, the two blocks are merged, and the check is repeated until no such situation exists before proceeding to the next step. The topology check effect is as follows: Figure 3 As shown, a is a schematic diagram of incorrect segmentation, and b is a diagram of the correction effect after topology checking.

[0039] S205 Calculate the center point of the tree

[0040] The tree trunk point cloud is used to calculate the center point of the tree. The tree trunk point cloud is projected onto a plane, and a fitted circle is calculated on the projected point cloud. The center of the fitted circle is recorded as the tree center point, which is used to locate the position of the tree.

[0041] There are two ways to calculate the center point:

[0042] The first method is to use the extracted tree trunk points as the initial points, project them to the Welzl algorithm to obtain the minimum covering circle, and use the center of the minimum covering circle as the center point of the tree trunk.

[0043] The second method is to calculate the average of the x, y, and z values ​​of the extracted tree trunk points and use them as the coordinates of the center point.

[0044] S3 K-means coarse segmentation

[0045] Using tree center points as input, K-means segmentation is performed to obtain coarsely segmented individual tree point clouds. Since K-means clustering requires setting the number of clusters and the approximate center point of each cluster, the number of tree center points is set to the number of clusters. K-means clustering is then performed using the tree center points as the center points of each cluster to coarsely segment the individual tree point clouds. Here, the tree size can be roughly fitted based on the known tree conditions. Adding the tree size value as a weight to the calculated distance during K-means segmentation can improve the segmentation. However, after the optimization algorithm is proposed later, the results without weighting are also good.

[0046] S4 Core Point Extraction

[0047] The segmentation is optimized using the coarsely segmented point clouds of individual trees obtained in S3 to better handle the situation of overlapping tree canopies.

[0048] S401 Single Wood Point Cloud Layered Projection

[0049] After dividing the point cloud of a single tree into multiple layers with a certain spacing, each layer is projected to fit the rough shape outline of the tree crown as closely as possible.

[0050] S402 circular fitting of tree edges

[0051] After fitting the edges of the ring, the core and center points are extracted. A ring of equal width is used to slide from the center outwards to extract the point cloud of the tree's layered projection. The number of points within the ring is used to determine the optimal point cloud representation. The point density of the point cloud in the annulus is calculated by combining the area S of the annulus. The N value reflects the size of the blank area within the ring. When the ring expands and reaches the set conditions, it is considered to have reached the boundary between the tree core point and the tree edge point. The points located within the ring are designated as the core points of a single tree, and the points outside the ring are designated as the edge points of a single tree. The core points and edge points are stored separately as single trees for further analysis.

[0052] There are two main ways to set the conditions here:

[0053] A method suitable for thinned point clouds, i.e., point clouds with low density: expand individual rings from the inside out and stop segmentation when a threshold is reached.

[0054] For point clouds with high point density, a continuous flow judgment is performed by shrinking a ring from the edge to the center. When the ring meets a preset condition, the number of ring layers at that time is saved. If the number of layers is found to be [missing information] in subsequent processes The ring, conforms to Furthermore, if it better fits the preset conditions, then it is considered a ring. The optimal segmentation position is determined by using... Overwrite records The search continued until the location was recorded. The segmentation is performed at the tree edges. This segmentation method offers higher accuracy and reduces the number of edge points. However, the calculation is relatively complex.

[0055] S5 builds tree proximity relationships

[0056] The proximity relationship determination uses tree center point data as initial data to obtain the proximity relationships between trees. The process is as follows: Figure 2As shown. Traverse the center points, establish a Cartesian coordinate system with one tree center point as the origin each time, determine the quadrant of other center points relative to the origin, and calculate the distance relative to the origin. After anomaly handling, retain the index of the nearest tree in each quadrant as the trees that may have intersecting canopies, which will be used for the next step of assigning edge points.

[0057] S6 Optimized Single-Tree Segmentation Results

[0058] The algorithm establishes connections between the edge points extracted in S4 and existing core points, assigning edge points to their respective cores to achieve optimized segmentation. Utilizing the idea of ​​finding the shortest edge from a point to the whole in Prim's minimum spanning tree algorithm, it traverses the edge points. For each edge point, it traverses the core points of the same index and the core points of neighboring trees obtained in S4, finding the core point of the tree closest to the edge point and assigning it to that core point. For high-density point clouds, a growth-based segmentation method is used: after finding the assignment of an edge point, it assigns all surrounding edge points within a certain range to the tree to which the found core point belongs. This edge point is recorded. If the distance between the next edge point and the current edge point is less than a preset value, the next edge point is also recorded, continuing until the distance between a point and the recorded point group is greater than the preset value. This method reduces the number of core point traversals, significantly reducing runtime. The optimized single-tree segmentation result is shown below. Figure 4 As shown.

[0059] S7 Abnormal Situation Handling

[0060] The main anomaly handling involves dealing with the situation where a small tree grows beneath a large tree. When a small tree grows beneath a large tree and the canopies interfere with each other, incorrect segmentation occurs, assigning part of the large tree's canopy to the small tree. Observation revealed that after layering the point cloud of the small tree resulting from the incorrect segmentation, the number of points in each layer changes with height in a roughly M-shaped pattern. Therefore, the layer with the highest number of points after layer counting is identified and denoted as [the layer with the highest number of points]. The layer with the most points in the higher-level part is denoted as . Then look for something between and The layer with the fewest points between them is denoted as ,by The tree crowns are divided into large and small tree crowns, and then the points where the large trees are incorrectly divided are extracted and assigned to new trees.

[0061] By extracting the incorrectly segmented points in the tree and re-assigning them, they can be correctly assigned to their respective locations.

[0062] In extreme cases, there might be two smaller trees under a large tree. In this case, the orientation relationship set by the center point is not applicable to the smaller trees on the edge and the large tree in the center, because although there is a tree between the two trees, their canopies still intersect. Therefore, this problem can be solved by exploring one more layer in the orientation relationship when confirming ownership.

[0063] This embodiment divides the 3D point cloud of urban green space into raster cubes, resulting in raster cubes of urban green space. This makes the features of the tree trunk point cloud more obvious compared to the tree crown point cloud, and it can also obtain the ground elevation information of the tree trunk location, which is more time-efficient than clustering. Then, based on the acquired tree trunk point cloud and elevation information, a series of operations are performed to complete coarse tree segmentation, and circular loop extraction is used to optimize the tree segmentation, extracting the core points and edge points of each tree. Combining the proximity relationships of individual trees within urban green space confirms the attribution of tree edge points, resolving the erroneous segmentation problem caused by intersecting tree crowns, thus improving the optimization effect of individual trees.

Claims

1. A method for single-tree segmentation of lidar point cloud in urban green space landscape scenes, characterized in that, The method includes: The point cloud of trees in urban green space is divided into raster cubes, and the initial tree trunk is extracted by comparing neighboring cells. The positional parameters of the initial trunk grid cube are fitted to obtain the DEM elevation plane, and the point cloud at a certain height is obtained from the DEM elevation plane as the trunk point cloud. The tree trunk point cloud is merged and topologically checked. The processed tree trunk point cloud is projected onto a plane. A fitted circle is calculated on the projected tree trunk point cloud. The center of the fitted circle is recorded as the tree center point, which is used to locate the tree position. K-means clustering was used to segment the center points of the trees, resulting in a coarsely segmented cloud of individual tree points. After dividing the roughly segmented point cloud of individual trees into layers with a certain spacing, each layer is projected. In the projection of each layer, a ring is used to fit the edge of the tree. The points inside the ring are the core points of the individual tree, and the points outside the ring are the edge points of the individual tree. The tree center point is used as initial data to obtain the proximity relationship between trees. The proximity relationship is used to determine the ownership of the edge point of the single tree and optimize the segmentation results of easily confused single trees. The initial trunk is extracted through neighbor-unit comparison, including: The elevation difference between the highest and lowest points in each grid cube is calculated and denoted as . ,by To fill the pixel values ​​with a two-dimensional grid, let the side length of the grid cell be... The central grid is The distance from the center grid is less than or equal to The grid is If satisfied Then the central grid is considered The initial tree trunk is included. The corresponding grid cube is extracted, and then the extracted grid cube is divided into horizontal layers with small vertical spacing. If a blank layer appears in the layer, it is considered that the grid cube is affected by some noise points and is not included in the scope of the initial tree trunk points.

2. The method for splitting a single piece of wood according to claim 1, characterized in that, The process of fitting the positional parameters of the initial trunk grid cube to obtain the DEM elevation plane includes: The grid cubes containing the initial tree trunk are statistically extracted. The position of each grid cube and the minimum elevation value within the cube are recorded. The grid cube positions and minimum elevation values ​​within the cubes are used for fitting to obtain the DEM elevation plane grid.

3. The method for splitting a single piece of wood according to claim 1, characterized in that, The merging and topology checking of the tree trunk point cloud includes: Determine if there are points adhering to the boundary of the tree trunk point cloud grid that connect two grids. If there are adhering points, the two grids should be merged. Then, determine if the size of the merged point cloud block exceeds the preset size. If it does, check if the adjacent joint point cloud block intersects with or is too close to the boundary of this block. If so, it is determined that the bottom of the tree crown has been cut, and then the two point cloud blocks are merged.

4. The method for splitting a single piece of wood according to claim 1, characterized in that, The method of fitting tree edges using a circular loop in the projection of each layer includes: For point clouds with low point density, use a ring to slide outwards from the center point of a single tree point cloud to extract the point cloud of the tree layered projection until the ring expansion reaches the threshold. For point clouds with high point density, a continuousness check is performed by shrinking the rings from the edge of the individual tree crowns in the layered projection towards the center point of the individual tree. When the rings meet the preset conditions, the number of ring layers at this time is saved. If the number of layers is found to be [missing information] in subsequent processes The ring, And the ring If the preset conditions are met, then the ring is considered to be circular. The optimal segmentation position is determined by using... Overwrite records The search continued until the location was recorded. The tree is divided at its edge.

5. The method for splitting a single piece of wood according to claim 1, characterized in that, The method of using the tree center point as initial data to obtain the proximity relationship between trees includes: traversing the tree center points, establishing a Cartesian coordinate system with one tree center point as the origin each time, determining the quadrant in which other center points are located relative to the origin, calculating the distance relative to the origin, and retaining the index of the nearest tree in each quadrant after anomaly handling.

6. The method for splitting a single piece of wood according to claim 1, characterized in that, The step of determining the affiliation of individual tree edge points based on the tree proximity relationships and optimizing easily confused individual tree segmentation results includes: For point clouds with low point density, for each edge point, traverse the core points of the same tree and the core points of neighboring trees, find the core point closest to the edge point, and assign this edge point to the tree to which the found core point belongs; For point clouds with high point density, after finding the affiliation of an edge point, the edge point is recorded. If the distance between the next edge point and the current edge point is less than a preset value, the next edge point is also recorded. This process continues until a point is found whose distance to the recorded point group is greater than the preset value. Then, all recorded edge points are assigned to the single tree to which the core point found by the first edge point belongs.

7. The method for splitting a single piece of wood according to claim 1, characterized in that, The optimized easily confused single-tree segmentation results include anomaly handling. The anomaly judgment method is as follows: when the number of points in each layer of the point cloud generated after single-tree segmentation changes with height in an M-shape, it is determined that there is a small tree growing under a large tree and the tree canopy is interfering with the growth. The anomaly handling method is: after layer counting, the layer with the most points in the lower layer is recorded as the highest-numbered layer. The layer with the most points in the higher-level section is denoted as the [layer name missing]. Then look for something between and The layer with the fewest points between them is denoted as ,by The tree crowns are divided into large and small tree crowns, and then the points where the large trees are incorrectly divided are extracted and assigned to new trees.

8. A computer storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the lidar point cloud single tree segmentation method for urban green space landscape scenes as described in any one of claims 1-7.

Images