Street tree point cloud progressive extraction method based on feature of ground object shape

Abstract

The application provides a street tree point cloud progressive extraction method based on ground object shape features, and relates to the technical field of vehicle-mounted LiDAR data processing. The method comprises the following steps: ground point clouds and non-ground point clouds are obtained by adopting cloth simulation filtering on the denoised point clouds; clustering and segmentation of discrete non-ground point clouds are realized by eight-neighbor clustering on a grid image, thereby obtaining N point cloud blocks as data processing units; features of each data processing unit are extracted; based on analysis of various ground object shape features, initial street tree point clouds are progressively extracted according to shape features in four aspects of area, height difference, projection and shape; and the initial street tree point clouds are optimized by using a binary algorithm to obtain final street tree results. Compared with the prior art, the method is easy to achieve high-precision ground object extraction effect, solves the problem that it is difficult to extract street trees with high precision due to dense vegetation and complex ground objects, and provides basic data support for subsequent street tree point cloud data statistics and three-dimensional modeling.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of vehicle-mounted LiDAR data processing, and particularly relates to a street tree point cloud progressive extraction method based on ground object morphological features. BACKGROUND

[0002] Vehicle-mounted LiDAR technology can efficiently and accurately obtain high-density three-dimensional discrete point cloud data of the ground surface and artificial objects such as buildings, and is widely used in three-dimensional reconstruction, urban planning and management, vehicle navigation, unmanned driving, disaster emergency and evaluation, etc., in which point cloud classification and extraction play a key role. The research on tree and attribute information extraction based on three-dimensional laser scanning data has become mature, but when scanning data in a large area, the vehicle-mounted LiDAR system has a relative disadvantage. Therefore, many studies on street trees have gradually shifted to vehicle-mounted laser scanning data processing.

[0003] Martin parameterizes the attributes of single street trees in the special case that the point clouds of the trees do not occlude each other, and finally models the street trees using these parameters; Yu et al. obtain the approximate position and geometric shape of the trees using a single point marking process, and the posterior distribution of the marked point cloud is obtained using a Bayesian model, while a reversible jump Markov Monte Carlo algorithm (R.JMCMC) is used for simulation experiments, and the best position of the trees is finally obtained, which has a certain practicality; Wang Dong et al. propose a method combining random forest and region segmentation to realize the extraction of street trees, which introduces a random forest algorithm to extract the point cloud of street trees and extracts the attribute information of single street trees and street trees through DBSCAN clustering and attribute information extraction method, which has adaptability, but its precision needs to be further improved; Zhang Yan et al. introduce a random forest algorithm, which accurately separates street tree point clouds from ground object point clouds through steps such as multi-dimensional feature vector construction, feature selection, random forest model establishment and random forest classification, but has relative limitations when facing complex environments. The above-mentioned extraction algorithms for street tree point clouds do not have good universality, and do not consider the case where low-height objects close to the tree trunk in some complex environments are difficult to remove, which affects the high-precision acquisition of street tree point clouds. SUMMARY

[0004] The technical problem to be solved by the present application is to solve the deficiencies of the prior art, and to provide a street tree point cloud progressive extraction method based on ground object morphological features.

[0005] The consideration of the extraction of the street tree point cloud environment can effectively improve the universality of the algorithm, and directly affects the extraction accuracy of the street tree point cloud. Therefore, through improving the universality and extraction accuracy of the extraction algorithm: the morphological characteristics of various ground objects are analyzed in depth, and the street trees are gradually extracted according to the morphological characteristics of area, height difference, projection and shape; the initial extracted street trees are optimized, the point cloud below the tree crown is optimized by using the street tree point cloud binary method, the small area object is extracted, and the final street tree point cloud data is obtained, which greatly improves the extraction accuracy of the street tree point cloud.

[0006] To solve the above technical problems, the application provides a street tree point cloud gradual extraction method based on ground object morphological characteristics, comprising the following steps:

[0007] Step 1: KD tree method (K-dimensional tree) is used to remove gross errors of original vehicle-mounted LiDAR point cloud data, and cloth simulation filtering (CSF) is used to obtain ground point cloud and non-ground point cloud after denoising.

[0008] Step 2: the clustering of eight-neighborhood of the grid image is realized, and the clustering segmentation of discrete non-ground point cloud is realized, so as to obtain N point cloud blocks as data processing units.

[0009] Step 2.1: non-ground point cloud rasterization;

[0010] Step 2.1.1: XOY two-dimensional plane grid is established according to the coverage range of ground object point cloud, and the formula is as follows, and the projection point cloud is projected to the grid plane;

[0011]

[0012]

[0013] In the formula, X max , X min , Y max , Y min are the maximum and minimum values of the point cloud coverage range in the X direction and the Y direction; Gridsize is the grid unit size, Gridsize=0.45 meters; m and n are the number of grid units in the X direction and the Y direction respectively; int(·) is the integer function.

[0014] Step 2.1.2: the lowest elevation value of the point cloud data in each grid is assigned to each grid, and the value of the empty grid is assigned to 0, so as to realize the rasterization of the ground object point cloud data;

[0015] Step 2.1.3: record and store the point cloud in each grid;

[0016] Step 2.2: Count the number of points in each grid N p , and determine whether it is greater than a given number threshold T N If N p > T N , the grid is retained and marked as a non-empty grid, otherwise, the grid containing points less than or equal to the number threshold T N is considered as an empty grid and assigned a value of 0.

[0017] Step 2.3: Assign the matrix value corresponding to the non-empty grid in the grid image to -1, set the index variable i, and set the initial value i = 1.

[0018] Step 2.4: Randomly select any non-empty grid in A as the initial center grid, and assign the value i to the center grid.

[0019] Step 2.5: Detect the grid with attribute value -1 in the eight-neighborhood of the center grid, and update its attribute value to i. The non-empty grids processed in steps 2.4 and this step will not be detected subsequently.

[0020] Step 2.6: For the newly detected non-zero grid, continue the eight-neighborhood detection, and repeat the process until no non-empty grid is found. At this point, a cluster unit is processed.

[0021] Step 2.7: Let i = i + 1, repeat steps 2.4 to 2.6, until all grid points are detected in the eight-neighborhood, and N independent cluster units are obtained.

[0022] Step 3: Extract the features of each cluster unit, including elevation-related features, projected area, projected span, projected area ratio, and other features.

[0023] For each cluster unit, the following features are extracted:

[0024] (1) Cluster unit elevation-related features include normalized elevation, maximum elevation, minimum elevation, and elevation difference of the gridded point cloud.

[0025] (2) Projected area S j is the area of the point cloud in the unit projected onto the two-dimensional XOY plane.

[0026] (3) Projected span K X and K Y are the maximum spans of each cluster unit in the grid X and Y directions, respectively.

[0027] Where the X-direction span is the difference between the maximum and minimum grid numbers in the X-direction corresponding to the projected unit, and the Y-direction span is the difference between the maximum and minimum grid numbers in the Y-direction corresponding to the projected unit.

[0028] (4) projection area ratio S r is the ratio of the projection area S1 of the overall point cloud in each clustering unit on the XOY plane and the projection area S2 of the partial point cloud on the XOY plane;

[0029] (5) area S of each clustering unit m and the number of point clouds contained, wherein the area S m is the number of plane grids occupied by each clustering unit.

[0030] Step 4: Based on the analysis of the morphological characteristics of various ground objects, the initial road tree point cloud is extracted according to the morphological characteristics of area, height difference, projection and shape in four aspects.

[0031] Step 4.1: Analyze the morphological characteristics of various ground objects;

[0032] Step 4.1.1: Ground information exists in the entire environment, and is basically flat in a local range, and the elevation of the ground point is the smallest in the entire environment;

[0033] Step 4.1.2: Buildings are distributed on both sides of the road, perpendicular to the ground, and the elevation of the building is the highest in the scene information, and the shape of the projection plane is a long strip shape, and the projection shape of the cross section is not much different from the projection shape of the entire building, and the projection area is larger than that of other ground objects;

[0034] Step 4.1.3: The road is distributed between the road and the building, and in terms of elevation, it is higher than the buildings, street lamps and power poles and other ground objects, and is higher than other ground objects. The projection shape of the road tree presents a cluster shape, the projection area is smaller than that of the building, and larger than that of other ground objects, and the projection area of the trunk part of the road tree is greatly different from the overall projection area;

[0035] Step 4.1.4: Vehicles are distributed on the road surface or on both sides of the road, and the elevation is lower than that of the road tree, building, street lamp and other ground objects, and higher than that of low vegetation, and the projection area is small;

[0036] Step 4.1.5: The road is distributed between the road trees or on both sides of the road, and the elevation is lower than that of the building, and higher than that of other ground objects, and is 5-6 meters, and the projection shape is irregular, and has good identification, and the projection area is small, and is usually not more than 1-2 square meters;

[0037] Step 4.1.6: Low vegetation is distributed around the road tree, and the elevation is higher than that of the ground and lower than that of other ground objects, and the maximum elevation is the same as that of the highest point of the trunk point. The projection shape is irregular, and the projection area is small.

[0038] Step 4.2: Secondary noise screening: count the area S of each unitm (m = 1…N), where the area is the number of grid cells occupied by the cluster unit, and an area threshold S is set y , and if there is a cluster unit whose area is less than the area threshold S y , the number of point clouds contained in the cluster unit also needs to be calculated, and when the number of points is less than the number threshold T M , it is determined to be secondary noise and is filtered out;

[0039] Step 4.3: Low-lying object screening: Calculate the maximum height difference value, and set a height difference threshold T H , if the maximum height difference value is less than the height difference threshold T H , the cluster unit is determined to be a low-lying object, and the cluster unit is deleted; if the maximum height difference value is greater than the set height difference threshold T H , the cluster unit is saved;

[0040] Step 4.4: Building screening: Set span thresholds T KX and T KY in the X and Y directions KX , if the span of the cluster unit in the X and Y directions is less than the given span threshold T KX and T KY , the cluster unit is considered to be a pole-shaped object, otherwise it is a building;

[0041] Step 4.5: Pole-shaped object screening: Set a threshold value, set an area threshold S according to the projection area of the cluster unit P , if the projection area of a cluster unit is greater than the set area threshold S P , the cluster unit is determined to be a tree; if the projection area is less than the threshold S P , it is determined to be other pole-shaped objects.

[0042] Step 5: Use a binary algorithm to optimize the initial street tree point cloud to obtain the final street tree result;

[0043] Step 5.1: Select point cloud data below 1.5 meters in the initial street tree point cloud data, grid it, and record the point cloud data in each grid with a grid size of 0.1 meters;

[0044] Step 5.2: Count and determine whether all grids are empty, if the grid is empty, assign a value of 0; if there is point cloud data in the grid, assign a value of 1 to the grid to generate a binary grid image;

[0045] Step 5.3: Cluster the binary image, set an index variable j and initialize j = 1;

[0046] Step 5.4: Randomly select any non-0 grid in the binary grid image as the initial center grid, and assign the value j to the center grid;

[0047] Step 5.5: the non-0 grid of the center grid is detected, and the attribute value of the non-0 grid is updated to j, and the non-empty grid processed in step 5.4 and this step is not detected subsequently;

[0048] Step 5.6: for the newly detected non-0 grid, the eight-neighborhood detection is continued, and the process is repeated until no non-empty grid is found, and thus a cluster unit is processed;

[0049] Step 5.7: j is set to j+1, and steps 5.4 to 5.6 are repeated until all grids are detected by eight-neighborhood detection, and M independent cluster units are obtained.

[0050] Step 5.8: it is judged whether the area of each cluster unit in the above result is less than a given area threshold T0, T0 is set to 0.01 square meters, and the number of grid units is 10, if the area of the cluster unit is less than T0, it is considered that the point cloud corresponding to the cluster unit is a tree trunk point cloud; if the area of the cluster unit is greater than T0, it is considered that the point cloud corresponding to the cluster unit is a non-road tree point cloud, and the non-road tree point cloud is removed.

[0051] The beneficial effects of the above technical method are:

[0052] (1) The present application provides a road tree point cloud progressive extraction method based on feature of ground objects, which provides technical support for high-precision extraction of road tree point cloud, and fully utilizes the feature of ground objects to quickly and efficiently extract road tree point cloud on the basis of in-depth analysis of the feature of common ground objects.

[0053] (2) After the road tree point cloud is progressively extracted based on the feature of ground objects, small area objects are extracted according to a point cloud binary algorithm, the road tree point cloud is optimized, and the extraction accuracy of the road tree point cloud is greatly improved. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The flowchart of the road tree point cloud progressive extraction method based on the feature of ground objects in the specific embodiment of the present application is shown;

[0055] Figure 2 The point cloud block effect diagram in the specific embodiment of the present application is shown;

[0056] Figure 3 The common ground object projection shape diagram in the specific embodiment of the present application is shown;

[0057] Figure 4 The experimental area diagram of experimental data in the specific embodiment of the present application is shown;

[0058] Figure 5 The experimental result diagram of experimental data in the specific embodiment of the present application is shown. DETAILED DESCRIPTION

[0059] The specific embodiments of the present application are described in further detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present application but are not intended to limit the scope of the present application.

[0060] The present application provides a street tree point cloud progressive extraction method based on the morphological characteristics of ground objects, as shown in Figure 1 The method comprises the following steps:

[0061] Step 1: KD tree method is used to remove gross errors of original vehicle-mounted LiDAR point cloud data, and cloth simulation filtering is used on the denoised point cloud to obtain ground point cloud and non-ground point cloud.

[0062] Step 2: clustering of eight-neighborhood of grid image is performed to realize clustering and segmentation of discrete non-ground point cloud, and N point cloud blocks are obtained as data processing units.

[0063] Step 2.1: non-ground point cloud gridding;

[0064] Step 2.1.1: XOY two-dimensional plane grid is established according to the coverage range of ground object point cloud, and the formula is as follows, and the point cloud is projected onto the grid plane;

[0065]

[0066]

[0067] In the formula, Xmax and Ymax are the maximum values in the X direction and the Y direction within the point cloud coverage range; Gridsize is the grid unit size, Gridsize = 0.45 meters; m and n are the number of grid units in the X direction and the Y direction, respectively; int(·) is the integer function. max min max min

[0068] Step 2.1.2: the lowest elevation value of point cloud data in each grid is assigned to each grid, and the value of empty grid is assigned to 0, to realize gridding of ground object point cloud data;

[0069] Step 2.1.3: record and store the point cloud in each grid;

[0070] Step 2.2: the number N of points in each grid is counted p , to determine whether it is greater than a given number threshold T N , which is set to 3 in the present application, if N p > T N ​​​​If the number of points contained in the grid is less than or equal to the number threshold T, the grid is regarded as a non-empty grid and is marked as 1, otherwise, the grid is regarded as an empty grid and is marked as 0. N If the number of points contained in the grid is less than or equal to the number threshold T, the grid is regarded as a non-empty grid and is marked as 1, otherwise, the grid is regarded as an empty grid and is marked as 0.

[0071] Step 2.3: Assign the matrix value at the non-empty grid in the corresponding grid image to -1, set an index variable i, and set the initial value i = 1.

[0072] Step 2.4: Randomly select any non-empty grid in A as the initial center grid, and assign i to the center grid.

[0073] Step 2.5: Detect the grid with attribute value -1 in the eight-neighborhood of the center grid, and update the attribute value to i. The non-empty grids processed in step 2.4 and this step will not be detected subsequently.

[0074] Step 2.6: For the newly detected non-0 grid, continue the eight-neighborhood detection, and repeat the process until no non-empty grid is found. At this point, one cluster unit is processed.

[0075] Step 2.7: Let i = i + 1, repeat steps 2.4 to 2.6, until all grid point clouds are detected in the eight-neighborhood, and N independent cluster units are obtained, as shown in Figure 2 .

[0076] Step 3: Extract the features of each cluster unit data processing unit, including elevation-related features, projected area, projected span, projected area ratio, and other features.

[0077] For each cluster unit, the following features are extracted:

[0078] (1) The cluster unit data processing unit elevation-related features include the normalized elevation of the gridded point cloud, the maximum elevation, the minimum elevation, and the elevation difference.

[0079] (2) The projected area S j is the area of the point cloud in the unit projected onto the two-dimensional XOY plane.

[0080] (3) The projected span K X and K Y are the maximum spans of each cluster unit in the grid X direction and Y direction, respectively.

[0081] Wherein, the X direction span is the difference between the maximum and minimum values of the grid number in the X direction corresponding to the projected unit; the Y direction span is the difference between the maximum and minimum values of the grid number in the Y direction corresponding to the projected unit.

[0082] (4) The projected area ratio S rThe ratio of the projection area S1 of the whole point cloud in each clustering unit on XOY plane and the projection area S2 of the partial point cloud on XOY plane;

[0083] (5) The area S of each clustering unit m and the number of point clouds contained, wherein the area S m is the number of plane grids occupied by each clustering unit.

[0084] Step 4: Based on the analysis of the morphological characteristics of various ground objects, the initial road tree point cloud is extracted gradually according to the morphological characteristics of area, height difference, projection and shape.

[0085] Step 4.1: The morphological characteristics of various ground objects are analyzed, and the projection shapes of various ground objects are shown in the following table: Figure 3

[0086] Step 4.1.1: The ground information exists in the whole environment, and in the local range, it is a plane, and in the whole environment, it is undulating, and the elevation of the ground point is the smallest;

[0087] Step 4.1.2: Buildings are distributed on both sides of the road, perpendicular to the ground, and the elevation of the building is the highest in the scene information, and the projection shape of the cross section is similar to the projection shape of the whole building, and the projection area is larger than that of other ground objects;

[0088] Step 4.1.3: The road trees are distributed between the road and the building, and in terms of elevation, the elevation is lower than that of the building, street lamp and power pole, and higher than that of other ground objects, the projection shape of the road tree presents a cluster shape, the projection area is smaller than that of the building, and larger than that of other ground objects, and the projection area of the trunk part of the road tree is greatly different from the whole projection area;

[0089] Step 4.1.4: Vehicles are distributed on the road surface or on both sides of the road, and the elevation is lower than that of the road tree, building, street lamp and other ground objects, and higher than that of low vegetation, and the projection area is small;

[0090] Step 4.1.5: The street lamp is distributed between the road trees or on both sides of the road, and the elevation is lower than that of the building and higher than that of other ground objects, which is 5-6 meters, the projection shape is irregular, and has good identification, the projection area is small, usually not more than 1-2 square meters;

[0091] Step 4.1.6: Low vegetation is distributed around the road tree, and the elevation is higher than that of the ground and lower than that of other ground objects, and the maximum elevation is the same as that of the highest point of the trunk point, the projection shape is irregular, and the projection area is small.

[0092] Step 4.2: Secondary noise screening: the area S of each clustering unit is counted​m (m = 1…N), where the area is the number of grid cells occupied by the cluster unit, and an area threshold S is set y , and if there is a cluster unit whose area is less than the area threshold S y , the number of point clouds contained in the cluster unit also needs to be calculated, and when the number of points is less than the number threshold T M , it is determined to be secondary noise and is filtered out, and in this embodiment T M = 10;

[0093] Step 4.3: Low-lying object screening: Calculate the maximum height difference value, set a height difference threshold T H , and in this embodiment it is set to 1.55 meters, if the maximum height difference value is less than the height difference threshold T H , the cluster unit is determined to be a low-lying object and is deleted; if the maximum height difference value is greater than the set height difference threshold T H , the cluster unit is saved;

[0094] Step 4.4: Building screening: Set the span thresholds T KX and T KY in the X and Y directions, and in this embodiment the span thresholds T KX and T KY are set to 20 meters and 40 meters, respectively, if the span of the cluster unit in the X and Y directions is less than the given span thresholds T KX and T KY , the cluster unit is considered to be a pole-shaped object, otherwise it is a building;

[0095] Step 4.5: Pole-shaped object screening: Set an area threshold S P according to the projection area of the cluster unit, and in this embodiment it is set to 1.5 square meters, if the projection area of a cluster unit is greater than the set area threshold S P , the cluster unit is determined to be a tree; if the projection area is less than the threshold S P , it is determined to be other pole-shaped objects.

[0096] Step 5: Use a binary algorithm to optimize the initial street tree point cloud to obtain the final street tree result;

[0097] Step 5.1: Select the point cloud data below 1.5 meters in the initial street tree point cloud data, grid it with a grid size of 0.1 meters, and record the point cloud data in each grid;

[0098] Step 5.2: Statistically determine whether all grids are empty, if the grid is empty, it is assigned a value of 0; if there is point cloud data in the grid, the grid is assigned a value of 1, and a binary raster image is generated;

[0099] Step 5.3: Clustering the binary image, setting an index variable j and initial j = 1;

[0100] Step 5.4: Randomly selecting any non-0 grid in the binary grid image as the current grid, and assigning i to the current grid;

[0101] Step 5.5: Detecting the non-0 grid in the eight-neighborhood attribute value of the center grid, updating the attribute value of the non-0 grid to j, and the non-empty grid processed in steps 5.4 and this step will not be detected subsequently;

[0102] Step 5.6: For the newly detected non-0 grid, continue the eight-neighborhood detection, and repeat the process until no non-empty grid is found, and thus a clustering unit is processed;

[0103] Step 5.7: Let j = j + 1, repeat steps 5.4 to 5.6 until all grids are detected in the eight-neighborhood, and M independent clustering units are obtained.

[0104] Step 5.8: Determine whether the area of each clustering unit in the above result is less than a given area threshold T0, the present application sets T0 = 0.01 square meters, corresponding to the number of grid units is 10, if the area of the clustering unit is less than T0, it is considered that the corresponding point cloud is a tree trunk point cloud; if the area of the clustering unit is greater than T0, it is considered that the corresponding point cloud is a non-road tree point cloud, and it is removed.

[0105] The experimental data set used in this embodiment is the vehicle-mounted LiDAR point cloud data measured in the campus of Jiaozuo, Henan Province, the data is located in the campus area, and the collection time is in the summer season, so the vegetation is lush. The experimental area contains right-angled buildings, roads, street trees, low objects, and aerial objects, and the terrain is relatively flat, containing 1101958 point cloud data, and the data is shown as Figure 4

[0106] The experiment was carried out on a computer with Intel(R) Core(TM) i5-6200U CPU @ 2.30 GHz 2.40 GHz, 64-bit, 16G memory and Win 10 operating system, and the experimental algorithm was edited and run with the help of MATLAB 2020a software platform, and the point cloud extraction accuracy of the present application was evaluated based on the standard data. At the same time, in order to verify the effectiveness of the algorithm of the present application, the extraction accuracy of the algorithm of the present application is analyzed and evaluated as follows: the first type of error is 3.6%, the second type of error is 8.7%, the total error is 5.0%, the Kappa coefficient is 0.8646, and the tree extraction accuracy is 95%; It is proved that in this embodiment, the street tree point cloud progressive extraction method based on the morphological features of the ground object of the present application has good effect. Figure 5 The final extraction effect diagram of the present example.

[0107] ​It should be noted that the above examples are only used to illustrate the technical solutions of the present application, and are not intended to limit the present application; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that the technical solutions recorded in the foregoing examples can be modified, or some or all of the technical features can be replaced by equivalents; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope defined by the claims of the present application.

Claims

1. A street tree point cloud progressive extraction method based on ground object shape features, characterized by comprising the following steps: Step 1: adopt KD tree method to remove gross errors of original vehicle-mounted LiDAR point cloud data, and adopt cloth simulation filtering to obtain ground point cloud and non-ground point cloud from the denoised point cloud; Step 2: realize clustering and segmentation of discrete non-ground point cloud by eight-neighbor clustering of raster images, and obtain N point cloud blocks as data processing units; Step 3: extract features of each clustering unit, including elevation-related features, projected area, projected span, projected area ratio, area of each clustering unit and number of point clouds contained; Step 4: based on analysis of morphological features of various ground objects, gradually extract initial road tree point cloud according to morphological features of area, height difference, projection and shape; Step 4.1: analyze morphological features of various ground objects; Step 4.2: Secondary noise screening: count the area S of each cell m (m = 1…N), while setting an area threshold S y , determine whether the area of the clustered cell is greater than the area threshold, if there is a clustered cell whose area is less than the area threshold S y , the number of point clouds contained in the clustered cell also needs to be calculated, when the number of points is less than the number threshold T M , it is determined as secondary noise and filtered out; Step 4.3: Low-height object screening: maximum height difference value is calculated, and a height difference threshold T is set H If the maximum height difference value is less than the height difference threshold T H , the clustering unit is determined as a low-height object, and the clustering unit is deleted; if the maximum height difference value is greater than the set height difference threshold T H , the clustering unit is saved. Step 4.4: Building screening: set span threshold T in X and Y directions KX and T KY If the span of the cluster unit in X and Y directions is less than the given span threshold T KX and T KY , then the cluster unit is considered as a pole-like feature, otherwise as a building. Step 4.5: Rod-like object screening: Set threshold value Set an area threshold value S according to the projected area of the cluster unit P If the projected area of a certain cluster unit is greater than the set area threshold value S P , then determine that the cluster unit is a tree; if the projected area is less than the threshold value S P , then determine it as other rod-like objects; Step 5: optimize the initial road tree point cloud by using a binary algorithm to obtain the final road tree result.

2. The method of claim 1, wherein the method comprises: The step 2 specifically comprises: Step 2.1: non-ground point cloud rasterization; Step 2.2: Count the number of points N in each grid p , and determine whether it is greater than a given number threshold T N If N p > T N , the grid is retained and marked as a non-empty grid. Otherwise, the grid is considered empty if the number of points is less than or equal to the number threshold T N , and assigned a value of 0. Step 2.3: assign the matrix value of the non-empty grid on the corresponding grid image to-1, set the index variable i, and set the initial value i=1; Step 2.4: randomly select any non-empty grid in A as the initial center grid, and assign i to the center grid; Step 2.5: detect the grid with attribute value-1 in the eight-neighborhood of the center grid, and update the attribute value to i. The non-empty grid processed in step 2.4 and this step will not be detected subsequently; Step 2.6: for the newly detected non-0 grid, continue the eight-neighborhood detection, and repeat the process until no non-empty grid is found. At this point, a clustering unit is completed; Step 2.7: let i=i+1, repeat steps 2.4 to 2.6 until all grid point clouds are completed eight-neighborhood detection, and N independent clustering units are obtained.

3. The method of claim 2, wherein the method further comprises: The step 2.1 specifically comprises: Step 2.1.1: establish an XOY two-dimensional plane grid according to the coverage range of the ground object point cloud, and the formula is as follows. Project the point cloud onto the grid plane; In the formula, X max , Y min , Y max , Y min are maximum and minimum values in the X direction and the Y direction within the point cloud coverage range; Gridsize is the grid cell size, Gridsize = 0.45 meters; m and n are respectively the number of grid cells in the X direction and the Y direction; int(·) is the integer function; Step 2.1.2: assign the lowest elevation value of the point cloud data in each grid to each grid, and assign 0 to the empty grid to realize rasterization of the ground object point cloud data; Step 2.1.3: record and store the point cloud in each grid.

4. The method of claim 1, wherein the method further comprises: The step 3 specifically comprises: The following features are extracted for each clustering unit: (1) the clustering unit elevation-related features include normalized elevation, maximum elevation, minimum elevation and elevation difference of the grid point cloud; (2) Projected area S j is the area of the point cloud in the unit projected onto the two-dimensional XOY plane; (3) Projected span K X and K Y are the maximum projected span of each cluster cell in the grid X and Y directions, respectively. The step 4.1 specifically comprises: (4) the ratio of the projection area S r is the ratio of the projection area S1 of the overall point cloud in the XOY plane and the projection area S2 of the partial point cloud in the XOY plane in each cluster unit. (5) Area S of each clustering unit m and the number of point clouds contained, wherein the area S m is the number of plane grids occupied by each clustering unit.

5. The method of claim 1, wherein the method further comprises: Step 4.1.1: ground information exists in the entire environment, appears as a plane in a local range, and is undulating in the entire environment, and the elevation of the ground point is the minimum. ​ Step 4.1.2: The buildings are distributed on both sides of the road, perpendicular to the ground, and the elevation of the buildings is the highest in the scene information, the shape of the buildings in the projection plane is long strip shape, and the projection shape of the cross section is not much different from the projection shape of the whole building, and the projection area is larger than that of other ground objects; Step 4.1.3: The street trees are distributed between the road and the buildings, and in terms of elevation, the elevation of the street trees is lower than that of the buildings, street lamps and power poles, and higher than that of other ground objects, the projection shape of the street trees is cluster-shaped, the projection area is smaller than that of the buildings and larger than that of other ground objects, and the projection area of the trunk part of the street trees is greatly different from the whole projection area; Step 4.1.4: The vehicles are distributed on the road surface or on both sides of the road, and the elevation is lower than that of the street trees, buildings, street lamps and other ground objects, and higher than that of low vegetation, and the projection area is small; Step 4.1.5: The street lamps are distributed between the street trees or on both sides of the road, and the elevation is lower than that of the buildings and higher than that of other ground objects, about 5-6 meters, the projection shape is irregular, has good recognition, and the projection area is small, not more than 1-2 square meters; Step 4.1.6: The low vegetation is distributed around the street trees, and the elevation is higher than that of the ground and lower than that of other ground objects, and the maximum elevation is the same as that of the highest point of the trunk point, the projection shape is irregular, and the projection area is small.

6. The method of claim 1, wherein the method further comprises: The step 5 specifically comprises: Step 5.1: Selecting point cloud data below 1.5 meters in the initial street tree point cloud data, gridding the point cloud data, the grid size is 0.1 meters, and recording the point cloud data in each grid; Step 5.2: Counting and judging whether all grids are empty, if the grid is empty, assigning a value of 0; if there is point cloud data in the grid, assigning a value of 1 to the grid, and generating a binary grid image; Step 5.3: Clustering the binary image, setting an index variable j and initializing j = 1; Step 5.4: Randomly selecting any non-0 grid in the binary grid image as an initial center grid, and assigning a value of j to the center grid; Step 5.5: Detecting the non-0 grid in the eight neighborhood of the center grid, updating the attribute value of the non-0 grid to j, and the non-empty grid processed in step 5.4 and this step will not be detected subsequently; Step 5.6: For the newly detected non-0 grid, continue the eight neighborhood detection, and repeat the process until no non-empty grid is found, and thus a clustering unit is completed; Step 5.7: Let j = j + 1, repeat steps 5.4 to 5.6 until all grids are completed eight neighborhood detection, and M independent clustering units are obtained; Step 5.8: Judging whether the area of each clustering unit in the above result is less than a given area threshold T0, setting T0 = 0.01 square meters, and the number of corresponding grid units is 10, if the area of the clustering unit is less than T0, it is considered that the corresponding point cloud is a trunk point cloud; if the area of the clustering unit is greater than T0, it is considered that the corresponding point cloud is a non-street tree point cloud, and it is removed.

Images