CN115187979A

CN115187979A - Single-wood point cloud branch and leaf separation method based on graph theory

Info

Publication number: CN115187979A
Application number: CN202210642833.2A
Authority: CN
Inventors: 李世华; 田志林
Original assignee: University of Electronic Science and Technology of China
Current assignee: University of Electronic Science and Technology of China
Priority date: 2022-06-08
Filing date: 2022-06-08
Publication date: 2022-10-14

Abstract

The invention belongs to the technical field of ground laser radar point cloud data processing, and particularly relates to a method for separating branches and leaves of a single-wood point cloud based on a graph theory. The method is based on multi-station scanning single-tree point cloud data acquired by a ground laser radar, sequentially undergoes map construction, hierarchical clustering, initial branch extraction and final branch extraction, and establishes a single-tree point cloud branch and leaf separation method for distinguishing single-tree point cloud branch and leaf components. The method fully utilizes the shortest path information of the graph, does not rely on prior knowledge and machine learning, has the characteristics of simplification and high efficiency while ensuring the classification precision, provides a new thought for single-wood point cloud branch and leaf separation, and can be further applied to the fields of forest parameter extraction, intelligent forestry, vegetation radiation transmission modeling and the like.

Description

Single-wood point cloud branch and leaf separation method based on graph theory

Technical Field

The invention belongs to the technical field of ground laser radar point cloud data processing, and particularly relates to a method for separating branches and leaves of a single-wood point cloud based on a graph theory.

Background

Forests play a key role in the global carbon, water, energy cycle, and climate change. The forest parameters represent the spatial layout and growth condition of the forest, and have important significance for forest resource management and vegetation ecology research. The laser radar is an active remote sensing technology developed in recent decades, can depict three-dimensional structure information of a forest through massive space points (namely point clouds) or waveforms, and is a mainstream means for extracting forest parameters at present.

Laser radars can be broadly classified into satellite-borne laser radars, airborne laser radars, and foundation laser radars (TLS) according to the difference between laser-mounted platforms. Their differences in data acquisition capabilities determine the types and scales of forest parameters that they can extract. The ground-based laser radar performs panoramic scanning from the interior of the forest and can be used for describing the information of under-forest branches and low vegetation in detail through multi-station splicing. The method can directly extract single tree parameters such as tree height, crown width, tree crown volume, branch height, breast height and the like, and most importantly, the obtained branch information can be used for directly estimating the wood volume and the aboveground biomass and can also be used for separating wood components so as to improve the estimation precision of the leaf area index. In addition, the detailed forest point cloud can be used for the realistic three-dimensional reconstruction of a forest scene, and further used in the fields of intelligent forestry, vegetation radiation transmission modeling and the like.

The application usually needs to separate branches and leaves of forest point cloud in advance, and single wood is a basic unit of the forest, so that the research of the single wood point cloud branch and leaf separation method has great significance. The point cloud branch and leaf separation method is divided into a data driving method and a knowledge driving method according to whether training data are relied on or not.

The data-driven method mainly refers to a machine learning-based supervised classification method. Before classification, several to dozens of features are required to be generated for each point, then a classifier is trained through artificially marked training data, and finally the trained classifier is used for classifying target data. The machine learning algorithm is used for monitoring and classifying branches and leaves, and has the advantages of simplicity, convenience, no limitation of forest types and data sources, and the defect of needing to select labor-consuming and time-consuming training data according to data characteristics.

The knowledge-driven method is an unsupervised classification method specially used for separating branches and leaves of tree point cloud. The method is based on fully recognizing the morphological difference of branches and leaves and the growth rule of trees, so that training data is not needed. The current knowledge-driven method can be roughly divided into: 1. based on a model fitting method, the branch segments are detected by utilizing the characteristic that the branches are cylindrical through circle/ellipse/cylinder fitting; 2. based on a clustering method, by utilizing the characteristic that the branches are continuous and the leaves are scattered, the leaves are divided into a large number of small clusters, the branches are divided into a small number of large and slender clusters, and finally the large and slender clusters are distinguished through size and linear characteristics; 3. based on the method of graph theory, the characteristic that the branches and leaves are distributed in a layered mode is utilized, and branch and leaf division is directly carried out or the classification results obtained by other methods are optimized from the whole structure of the tree by means of the shortest path information of the graph.

However, the current graph theory-based branch and leaf separation method does not fully utilize the shortest path information of the graph. For example, vicari (2019) divides points in a certain distance from the tail end of a path into blade points, divides points of the path with the passing frequency larger than a set value into branch points, is simple in application, needs to adjust related parameters correspondingly according to the variety and the size of a tree, and cannot guarantee considerable classification precision, so that an author improves the robustness of the method by combining machine learning on the basis, and thus the method depends on prior parameters and is bloated.

Disclosure of Invention

Aiming at the problems or the defects, the invention provides a single-wood point cloud branch and leaf separation method based on graph theory, which fully utilizes the shortest path information of the graph, does not rely on prior knowledge and machine learning, and has the characteristics of simplification and high efficiency while ensuring the classification precision in order to solve the problems of dependence on prior parameters and flow bloat of the existing branch and leaf separation method based on graph theory.

A single-wood point cloud branch and leaf separation method based on graph theory comprises the following steps:

step 1, constructing a diagram: the method aims to connect discrete single-wood point clouds into an undirected weighted connected graph G, can overcome the influence of point cloud occlusion deletion to a certain extent, and accurately depict the surface morphology of trees with as few sides as possible.

1-1, inserting a root node: inserting a root node into the original single tree point cloud tree _ closed, selecting a section of point cloud with the height of 5-20cm at the bottom of the trunk, projecting the section of point cloud to the horizontal plane where the lowest point of the trunk is located, and performing least square circle fitting on the projected point cloud, wherein the circle center is the insertion position of the root node.

1-2, searching large-scale neighbor points: searching a large-range neighbor point neighbors _ K of each point in the tree _ group, wherein K is more than or equal to 50 and less than or equal to 300, and aiming at fully acquiring the proximity relation of each point; here, K neighbor search is adopted, and the more serious the data loss is, the larger the value of K is, but the memory consumption is increased.

1-3, constructing a small-scale neighborhood map:

1-3-1, creating a current set Q and an accessed set V, and putting root nodes into Q and V;

1-3-2. For each point in the current set Q, connecting it to the nearest 5 points not in V of its corresponding large-range neighbor points neighbors _ K, and adding these points to the temporary set temp _ Q, updating Q = temp _ Q, and then adding the points in the temporary set temp _ Q to V; this step is repeatedly executed until there is no point in the temporary set temp _ Q;

1-3-3. Initialize thr _ neighbor _ Dis = treeHight/30 (treeHight is tree height, thr _ neighbor _ Dis is neighbor distance threshold).

1-3-4. Define set un _ V = tree _ closed-V, and determine if un _ V is empty:

if un _ V is not empty, traversing each point in un _ V, if a point in V with a distance smaller than thr _ neighbor _ Dis exists in the corresponding large-range neighbor points neighbor _ K, connecting the point with the nearest 5 points in the points, and adding the point into temp _ Q; if the temp _ Q has a point, updating Q = temp _ Q, adding the point in the temp _ Q to V, and jumping to 1-3-2; and if the temp _ Q has no point, updating the thr _ neighbor _ Dis + = treeHight/60, and re-entering 1-3-4.

If un _ V is empty, the construction is completed, and a graph G = (V, E) is obtained, wherein G is composed of a vertex set V and an edge set E, and the weight of an edge is the Euclidean distance between two points.

Step 2, hierarchical clustering: and clustering the original single-wood point cloud tree _ closed in a hierarchical tangent graph mode.

2-1, shortest path extraction: the graph G obtained in the step 1 can accurately express the indication form and the topological structure of the single tree, so that the shortest path information (path sequence path _ list and path length path _ dis) from each point to the root node can be accurately extracted, and the shortest path is extracted by adopting a single-source shortest path algorithm (Dijkstra algorithm).

2-2, path length layering: layering the original single-tree point cloud tree _ group based on path _ dis, wherein the layering distance interval _ D = treeHight/30, and specifically, the method is realized by deleting edges located on the layering boundary in the graph G. And (5) performing hierarchical trimming operation, wherein the point cloud corresponding to the connected component of the graph G is the clustering result.

Step 3, extracting initial branches: and (4) extracting branch clusters with remarkable characteristics in the clustering result in the step (2).

3-1, coordinate transformation:

defining the axial direction of the cluster as the sum vector of direction vectors formed by each point in the cluster and the first precursor point on the path _ list, and then rotating the coordinate axis of the three-dimensional rectangular coordinate system to enable the coordinate z axis to be parallel to the axial direction of the cluster.

3-2, size filtering: if the absolute value of dimen _ z-interval _ D is greater than 0.25 integer \ D, the cluster is considered to be too small or too large, possibly noise, a blade or a blade cluster, and is filtered out firstly; and the medium _ z is the maximum z value-the minimum z value of the cluster along the z-axis direction after the coordinate transformation.

3-3, extracting the trunk and main branch clusters: and extracting trunk and main branch clusters with obvious cylindrical characteristics, and identifying through cylinder fitting, wherein the specific process is as follows.

3-3-1, cylinder fitting: through 3-1 coordinate transformation, the axial direction of the cluster is parallel to the coordinate z axis, so that the cluster can be projected to a two-dimensional plane along the z axis to carry out least square circle fitting, and the effect of cylinder fitting is achieved.

3-3-2, identifying based on relative fitting errors: the cylinder relative fit error formula is defined as follows.

In the formulaN is the number of points included in the cluster, d _i The distance from any point in the cluster to the axis of the fitted cylinder is defined, and r is the radius of the fitted cylinder; when rError<And when the time is 0.2, showing that the cluster-like cylindrical features are obvious, and extracting.

3-4, extracting the fine branch clusters: and extracting fine branch clusters with remarkable linear characteristics, and identifying the fine branch clusters through principal component analysis, wherein the specific process is as follows.

3-4-1, main component analysis:

let P be the cluster-like point set, and the covariance matrix of P is defined as follows:

wherein n is the number of points contained in P, P _i Is any point in P, P _c Is the centroid of P.

Cov _P Characteristic value (λ) of ₁ ≥λ ₂ ≥λ ₃ ) The degree of dispersion of P in three main component directions is characterized, and the spatial distribution characteristics (such as surface shape, linear shape and body shape) of P can be calculated by characteristic values.

3-4-2, identifying based on linear characteristics: defining Linearity = λ ₁ /(λ ₁ +λ ₂ +λ ₃ ) To represent linear characteristics of class clusters; when Linearity>When the time is 0.9, the linear characteristics of the cluster are obvious, and extraction is carried out.

3-5, classification correction: based on the tree growth rule that the tree trunk becomes thinner gradually from the tree trunk to the tree branches, a small number of leaf clusters which are divided into the tree trunk and the main branches by mistake and have the cylindrical characteristics are corrected, and the specific process is as follows.

3-5-1, traversing each cluster c which is divided into a trunk and a main branch in a 3-3 cylinder fitting mode, searching from any point in c to a root node along path _ list, and if a twig cluster obtained from 3-4 is encountered or a trunk and main branch cluster c 'obtained from 3-3 is encountered but the cylinder fitting radius of c is larger than that of c', indicating that the cluster c is wrongly divided, and removing the cluster c.

And 4, finally extracting the branches: and (4) taking the initial branch point obtained in the step (3) as a seed point, and extracting branch point cloud with non-significant features (such as bifurcations, bent branches, leaf winding, shading deletion and the like) in a region growing mode. The specific process is as follows:

4-1, seed point supplement: in order to speed up the region growing speed and make the region growing result more accurate, all the points of the initial branch point and the path _ list thereof are taken as seed points.

4-2, region growing:

4-2-1, creating a seed point set L and a branch point set F, and then putting the seed points obtained in the step 4-1 into the seed points L and the branch point set F.

4-2-2 if D _pp′ <0.25 star interval _Dand D _rp′ <D _rp Then p' is added to the temporary set temp _ L. Where p is a point in L, p' is a point where p is adjacent to and not in F on the original graph G (constructed in step 1, without edges deleted), D _pp′ Denotes the distance p to p', D _rp And D _rp′ Respectively representing the shortest path distances path _ dis of p and p' to the root node.

If the temp _ L has a point, updating L = temp _ L, putting the point in temp _ L into F, and executing 4-2-2 again;

if there is no point in temp _ L, the growth ends.

And 4-3, wherein points in the set F are branch point clouds, and a complementary set (tree _ closed-F) of the F is a leaf point cloud.

Step 1 the principle of constructing a graph:

for the unordered point cloud construction map, a neighborhood map is generally constructed by connecting each point with k neighbors of the point. However, when there is a missing occlusion in the point cloud data, k needs to be increased to ensure that the potential surfaces are connected together, which results in too dense edges making up the graph on the one hand and a wrong connection on the other hand. The method adopts a composition mode of large-scale neighbor point search and small-scale neighborhood graph construction, and can connect neighbor points with as few edges as possible under the condition of slight loss of single-tree point cloud, and simultaneously ensure that the connection relation is consistent with the tree topological structure.

Step 2, hierarchical clustering principle:

it is essentially a graph cut based clustering process, as compared to cutting the tree layer by layer with a knife, the last complete tree is cut into a number of separate clumps, except that here the graph expressing the surface morphology of the tree is cut, the position of the cut is a hierarchical interface determined by the shortest path length, and a number of connected components (clusters) are formed as a result of the cut.

Step 3, the principle of extracting the initial branches:

after hierarchical clustering, trunk/main branches with consistent length and presenting cylindrical characteristics and fine branch clusters with presenting linear characteristics have good distinguishability with small and discrete single leaf or dense leaf clusters, and can be separated out first, and size filtering, cylinder fitting (as shown in fig. 2) and principal component analysis (as shown in fig. 3) are used for detecting the characteristics. The initial branches are extracted from two aspects of cylindrical fitting and linear characteristics, so that the robustness of the method is improved, and the initial branches can be extracted as much as possible for trees of different varieties and sizes.

Step 4, final branch extraction principle:

the initial branch extraction can only extract branch clusters with remarkable features, and branch clusters at irregular shapes (bifurcations, bent branches, leaf winding, shading deletion and the like) are difficult to identify. And extracting branch point clouds at irregular positions by using the global features of the branches, taking the initial branch points as seed points and combining neighborhood and shortest path information of the map in a region growing mode. All the region growing is performed by taking the point as the unit, but not by taking the cluster as the unit, so that the region growing boundary is more accurate.

Firstly, constructing a graph for single-tree point cloud, and then extracting the information of the shortest path from each point to a root node (located in the center of a base of a trunk); layering the point cloud according to the path length, deleting edges on layered boundary lines, wherein the connected components of the graph are layered clustering results; extracting trunk and main branch clusters by using cylindrical characteristics, and extracting fine branch clusters by using linear characteristics so as to obtain initial branch points; taking the initial branch point as a seed point, extracting branch points at the irregular positions such as bifurcations and the like through region growing based on neighborhood and shortest path information of the graph; and obtaining final branch points through initial branch extraction and region growth, wherein the complementary set of the final branch points is leaf point cloud. According to the method, through point cloud layering, axial estimation, classification correction and region growth, the shortest path information of the graph is fully utilized, the method is simplified and efficient in calculation, the branches and leaves of the single-tree point cloud can be quickly and accurately separated, and the process is as shown in fig. 1.

In summary, the single-tree point cloud branch and leaf separation method for distinguishing the single-tree point cloud branch and leaf components is established based on multi-station scanning single-tree point cloud data acquired by a ground laser radar through drawing construction, hierarchical clustering, initial branch extraction and final branch extraction. The method is simple and efficient in calculation, can quickly and accurately separate branches and leaves of the single-wood point cloud, provides a new idea for the branch and leaf separation of the single-wood point cloud, and can be further applied to the fields of forest parameter extraction, intelligent forestry, vegetation radiation transmission modeling and the like.

Drawings

FIG. 1 is a schematic flow diagram of the present invention;

FIG. 2 is a schematic diagram of cylinder fitting;

FIG. 3 is a principal component analysis diagram;

FIG. 4 is a graph showing the separation results of branches and leaves; a. blade point cloud; b. and (4) branch point cloud.

Detailed Description

The invention is described in further detail below with reference to an example of 11 singles in conjunction with the accompanying drawings: the development environment is Pycharm, the programming language is Python 3.7, and the used software packages mainly comprise numpy, skearn and networkx.

Step 1, data acquisition and pretreatment: 11 actual measured multi-station scanning single-tree point cloud data containing multiple tree species and sizes are used for verifying the method, and relevant parameters are shown in a table 1. In addition to 1 magnolia sieboldii datum collected from the university campus of electronics, other single-tree data are all from publicly measured data sets in other related studies. The collecting equipment of the magnolia sieboldii is Leica Scanstation C10, and the collecting equipment of the rest trees is RIEGL VZ-400. The above data were voxel down-sampled (sampling interval: magnolia sieboldii 0.01m; rest 0.05 m) using CloudCompare point cloud processing software and branch and leaf separation was performed by manual cropping. The following is a detailed procedure for the branch and leaf separation of the 11 single-wood point cloud data by the method of the present invention.

TABLE 1 Single-wood point cloud data

And 2, constructing a diagram according to the step 1 in the invention content.

And 3, performing hierarchical clustering according to the step 2 of the invention content.

And 4, extracting the initial branches according to the step 3 of the invention content.

And 5, performing final branch extraction according to the content of the invention and the step 4.

Step 6, precision evaluation:

through the steps 2-5, the point cloud branch and leaf separation result of each tree is finally obtained, and the point cloud branch and leaf separation result of the magnolia sieboldii is shown in fig. 4. The separation result is quantitatively evaluated by four indexes of Precision, recall, F1-score and Accuracy: wherein, precision, recall and F1-score three indexes are specific to a certain class (branch or leaf), and Precision represents the ratio of correctly classified points to the points classified into the class; recall represents the ratio of the number of correctly classified points to the number of true points; f1-score is the harmonic mean of Precision and Recall; accuracy represents the ratio of the number of correctly classified points to the total number of points for all classes. The classification precision of each tree is shown in table 2; the average value of Accuracy is 0.971, the standard deviation is 0.013, the maximum value is 0.989, the corresponding novraguesH 20, the minimum value is 0.945, the corresponding tree 2; the F1-score of the branches is lower than that of the leaves, and the difference is 10 percentage points on average.

This example deals with 11 actual measured multi-station scanned single-tree point cloud data containing multiple tree species and sizes according to the method of the present invention. According to the method steps, the point cloud branches and leaves of each tree are separated in a parameter self-adaptive setting mode, and quantitative Precision evaluation is carried out on the separation result through four indexes of Precision, recall, F1-score and Accuracy. The result shows that the research method can obtain higher and stable branch and leaf separation precision aiming at trees of different varieties and sizes, but the classification precision of the branches is obviously lower than that of the leaves, and the analysis reasons include the following points: 1. the quantity difference of the branch and leaf point clouds is huge; 2. the extraction of the twigs is not sufficient; 3. some leaves of trees adhere to the trunk and thick branches, and are easily divided into branches by mistake. In addition, the operating environment of the present embodiment is as follows: windows 7, intel (R) Core (TM) i5-7500,8GB RAM; the treatment time per tree is 1-2 minutes, indicating that the method of the invention is computationally efficient.

TABLE 2 precision of branch and leaf separation

As can be seen by the above examples: firstly, constructing a graph for the point cloud of the single tree, and then extracting the information of the shortest path from each point to a root node; layering the point cloud according to the path length, deleting edges on layered boundary lines, wherein the connected components of the graph are layered clustering results; extracting trunk and main branch clusters by using cylindrical characteristics, and extracting fine branch clusters by using linear characteristics so as to obtain initial branch points; taking the initial branch point as a seed point, extracting branch points at the irregular positions such as bifurcations and the like through region growing based on neighborhood and shortest path information of the graph; and (4) obtaining final branch points through initial branch extraction and region growing, wherein the complementary set of the final branch points is leaf point cloud. According to the invention, through point cloud layering, axial estimation, classification correction and region growth, the shortest path information of the graph is fully utilized, prior knowledge and machine learning are not relied on, the classification precision is ensured, and the method has the characteristics of simplification and high efficiency, provides a new thought for branch and leaf separation of single-wood point cloud, and can be further applied to the fields of forest parameter extraction, intelligent forestry, vegetation radiation transmission modeling and the like.

Claims

1. A method for separating branches and leaves of a single-wood point cloud based on graph theory is characterized by comprising the following steps:

step 1, constructing a graph: connecting the discrete single-wood point clouds into an undirected weighted connected graph G;

step 2, hierarchical clustering: clustering the original single-wood point cloud tree _ cluster in a hierarchical tangent graph mode;

2-1, shortest path extraction: extracting a shortest path information path from each point to a root node, a path sequence path _ list and a path length path _ dis from the graph G obtained in the step 1 by adopting a single-source shortest path algorithm;

2-2, path length layering: layering an original single-tree point cloud tree _ close based on a path length path _ dis, wherein the layering interval _ D = treeHight/30, the method is realized by deleting edges positioned on a layering boundary in a graph G, and the point cloud corresponding to the connected components of the graph G is a clustering result through layering edge cutting operation;

step 3, extracting initial branches: extracting branch clusters with obvious characteristics in the clustering result in the step 2;

and 4, finally extracting the branches: taking the initial branch point obtained in the step 3 as a seed point, and extracting a branch point cloud with an insignificant characteristic in a region growing mode, wherein the method specifically comprises the following steps:

4-1, seed point supplement: taking all the points of the initial branch point and the path sequence path _ list thereof as seed points;

4-2, region growing:

4-2-1, creating a seed point set L and a branch point set F, and then putting the seed points obtained in the step 4-1 into the seed point set L and the branch point set F;

4-2-2 if D _pp′ <0.25 space \uD and D _rp′ <D _rp Adding p' into the temporary set temp _ L; where p is a point in L, p' is a point where p is adjacent to the original graph G without the deleted edge and is not in F, D _pp′ Denotes the distance p to p', D _rp And D _rp′ Respectively representing the shortest path distances path _ dis from p and p' to the root node;

if the temp _ L has no point, the growth is finished;

and 4-3, points in the set F are the branch point cloud, and the complementary set tree _ closed-F of the F is the leaf point cloud.

2. The graph theory-based method for separating branches and leaves of single-wood point cloud, according to claim 1, wherein the step 1 specifically comprises the following steps:

step 1-1, inserting a root node: inserting a root node into the original single tree point cloud tree _ closed, selecting a section of point cloud with the height of 5-20cm at the bottom of a trunk, projecting the section of point cloud to a horizontal plane where the lowest point of the trunk is located, and performing least square circle fitting on the projected point cloud, wherein the circle center is the insertion position of the root node;

1-2, searching large-scale neighbor points: searching large-range neighbor points neighbor _ K of each point in the original single-wood point cloud tree _ closed, wherein K is more than or equal to 50 and less than or equal to 300, so as to fully obtain the proximity relation of each point, and the more serious the data loss is, the larger the value of K is, but the memory consumption is increased;

1-3, constructing a small-scale neighborhood graph:

1-3-3. Initializing thr _ neighbor _ Dis = treeHight/30, treeHight being tree height, thr _ neighbor _ Dis being a neighbor distance threshold;

1-3-4. Define un _ V = tree _ closed-V, and judge if un _ V is null:

if un _ V is not empty, traversing each point in un _ V, if a point in V with a distance smaller than thr _ neighbor _ Dis exists in the corresponding large-range neighbor points neighbor _ K, connecting the point with the nearest 5 points in the points, and adding the point into temp _ Q; if the temp _ Q has a point, updating Q = temp _ Q, adding the point in the temp _ Q to V, and jumping to 1-3-2; if the temp _ Q has no point, updating thr _ neighbor _ Dis + = treeHight/60, and re-entering 1-3-4;

3. The graph theory-based method for separating branches and leaves of single-wood point cloud as claimed in claim 1, wherein: the single-source shortest path algorithm in the step 2-1 is a Dijkstra algorithm.

4. The graph theory-based method for separating branches and leaves of single-wood point cloud, according to claim 1, wherein the step 3 specifically comprises the following steps:

3-1, coordinate transformation: defining the axial direction of the class cluster as the sum vector of direction vectors formed by each point in the class cluster and the first precursor point on the path sequence path _ list of the point, and then rotating the coordinate axes of the three-dimensional rectangular coordinate system to ensure that the coordinate z axis is axially parallel to the class cluster;

3-2, size filtering: if the absolute value of dimen _ z-interval _ D >0.25 × interval _D, the cluster is considered to be too small or too large, and the cluster is filtered; the dimension _ z is the maximum z value-the minimum z value of the cluster along the z-axis direction after the coordinate transformation;

3-3, extracting the trunk and main branch clusters: the trunk and main branch clusters with obvious cylindrical characteristics are extracted and are identified through cylinder fitting, and the specific process is as follows:

3-3-1, cylinder fitting: through 3-1 coordinate transformation, the axial direction of the cluster is parallel to the coordinate z axis, so that the cluster can be projected to a two-dimensional plane along the z axis to carry out least square circle fitting, and the effect of cylinder fitting is achieved;

3-3-2, identifying based on relative fitting errors: the cylinder relative fit error formula is defined as follows:

wherein n is the number of points included in the cluster, d _i The distance from any point in the cluster to the axis of the fitted cylinder is shown, and r is the radius of the fitted cylinder; when rError<When 0.2 hour, the cluster-like cylindrical features are obvious, and extraction is carried out;

3-4, extracting the fine branch clusters: extracting fine branch clusters with obvious linear characteristics, and identifying the fine branch clusters through principal component analysis, wherein the specific process is as follows:

3-4-1, main component analysis: let P be a cluster-like point set, and the covariance matrix of P is defined as follows:

wherein n is the number of points contained in P, P _i Is any point in P, P _c Is the centroid of P;

Cov _P characteristic value λ of ₁ ≥λ ₂ ≥λ ₃ Representing the dispersion degree of P in three principal component directions, wherein the spatial distribution characteristic of P is obtained by calculating characteristic values;

3-4-2, identifying based on linear characteristics: definition Linearity = λ ₁ /(λ ₁ +λ ₂ +λ ₃ ) To represent linear characteristics of the class clusters; when Linearity>When 0.9, indicating that the linear characteristics of the cluster are obvious, extracting;

3-5, classification correction: based on the tree growth rule that the trunk becomes thinner gradually from the tree branches, the tree leaf cluster which is mistakenly divided into the trunk and the main branches and has the cylindrical characteristic is corrected, and the specific process is as follows:

3-5-1, traversing each cluster c which is divided into a trunk and a main branch in a cylinder fitting mode in the step 3-3, searching from any point in the cluster c to a root node along a path sequence path _ list, and if the branch cluster obtained in the step 3-4 is met or the cluster c 'of the trunk and the main branch obtained in the step 3-3 is met but the cylinder fitting radius of the c is larger than that of the c', indicating that the cluster c is wrongly divided and removed.