A method for extracting tree trunks from lidar point clouds based on machine learning

Through the machine learning-based lidar point cloud extraction method, using support vector machine and Euclidean distance grouping technology, the problems of slow extraction speed and low accuracy of trunk point cloud in the existing technology are solved, and fast and accurate trunk point clouds are achieved Data extraction, suitable for efficient backbone point cloud extraction of the entire sample site.

CN113205543BActive Publication Date: 2024-12-06NANJING FORESTRY UNIV
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202110584803.6
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-05-27
Publication Date
2024-12-06
Estimated Expiration
2041-05-27

AI Technical Summary

Technical Problem

The existing technology cannot quickly and accurately extract the trunk point cloud data. Traditional artificial methods are destructive and high workload, deep learning methods have a long calculation time, and clustering classification methods have poor point cloud extraction effect at the top of the trunk.

Method used

Using machine learning-based lidar point cloud extraction method, point cloud data is collected through ground-based lidar, feature parameters are calculated, support vector machines are used for classification training, and the main point cloud is finally extracted.

Benefits of technology

Fast and accurate trunk point cloud extraction is achieved. The support vector machine has good classification performance for small samples and has fast classification speed, which is suitable for efficient extraction of backbone point clouds in the entire sample site.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113205543B_ABST
    Figure CN113205543B_ABST
Patent Text Reader

Abstract

The present invention discloses a method for extracting trunks from laser radar point clouds based on machine learning, comprising the following steps: 1) using a ground-based laser radar to collect point cloud data of sample trees in a sample plot; 2) preprocessing the point cloud data to obtain sample point cloud data; 3) based on the sample point cloud data, calculating the characteristic parameters of any point in the sample point cloud data, and forming a characteristic matrix with the characteristic parameters; 4) selecting any sample tree in the sample plot, the sample point cloud data covered by the sample tree as a training sample, marking its trunk point cloud and non-trunk point cloud; 5) based on the training sample and the characteristic matrix, classifying and training the sample point cloud data of the entire sample plot by machine learning. Based on representative feature vectors, effective distinction can be made in discriminating trunk point clouds and non-trunk point clouds, and the marking training of a single tree is applicable to the entire sample plot.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to a method for accurately and quickly extracting a tree trunk, in particular to a method for extracting a tree trunk from a laser radar point cloud based on machine learning. Background Art

[0002] As an important place for tree nutrient absorption and transfer, the trunk of a tree has strong support and is the main part of the economic utilization of trees. The accurate extraction of the trunk is the basis for the accurate calculation of tree height and trunk volume, and is of great significance for the estimation of vegetation carbon fixation and biomass. Real-time and rapid extraction of the trunk is conducive to timely evaluation of tree yield, which can be applied to furniture manufacturing and estimation of building materials.

[0003] Traditional tree trunk resources are mainly obtained by manual measurement. This method requires cutting down trees to obtain the trunk part, which causes certain damage to standing trees and the damage is irreversible. At the same time, it also has disadvantages such as large workload and inability to quickly understand the basic situation of the trunk.

[0004] In recent years, with the continuous development of radar technology, laser radar detection has replaced the traditional manual sampling method. Laser radar detection can quickly obtain detailed structural information of trees and has great advantages in vegetation visualization. As an active remote sensing tool, laser radar can quickly collect high-density point cloud data, and the stable collection platform ensures the accuracy of data collection. By selecting appropriate laser radar scanning tools, detailed coverage information of the complete vegetation surface can be obtained without loss, and real and detailed three-dimensional tree data can be obtained. At present, domestic and foreign scholars often use tools such as terrestrial laser scanning (TLS) and mobile laser scanning (MLS) to efficiently obtain three-dimensional point cloud data of trees.

[0005] At present, there is no unified method for processing tree point clouds. Some scholars use deep learning methods for extraction, which usually requires a long processing time and a large amount of data as a basis in the early stage. Other scholars use clustering classification methods, which have poor continuity in the upper layer of the trunk and make it difficult to extract the point cloud at the top of the trunk. Therefore, how to use point cloud data to quickly and accurately extract the point cloud of the trunk is a technical problem that needs to be solved urgently. Summary of the invention

[0006] The purpose of the present invention is to provide a method for extracting tree trunks from lidar point clouds based on machine learning. Based on representative feature vectors, it can effectively distinguish between trunk point clouds and non-trunk point clouds, and the marking training of single trees is applicable to the entire sample plot. The tree point clouds are extracted using machine learning methods, and the support vector machine has good classification performance for small samples and fast classification speed, providing support for the efficient extraction of trunk point clouds.

[0007] To achieve the above object, according to one aspect of the present invention, the present invention provides the following technical solution:

[0008] A method for extracting tree trunks from lidar point clouds based on machine learning, comprising the following steps:

[0009] 1) Use ground-based LiDAR to collect point cloud data of sample trees in the sample plot;

[0010] 2) Preprocessing the point cloud data to obtain sample point cloud data;

[0011] 3) Based on the sample point cloud data, calculate the characteristic parameters of any point in the sample point cloud data, and form the characteristic parameters into a characteristic matrix;

[0012] 4) Select any sample tree in the sample plot. The sample point cloud data covered by the sample tree is the training sample, and mark its main point cloud and non-main point cloud;

[0013] 5) Based on the training samples and feature matrix, the sample point cloud data of the entire sample plot is classified and trained by machine learning;

[0014] 6) Perform Euclidean distance grouping and discrimination on all the marked main point cloud data in the sample plot, further extract the main point cloud, and obtain the final main point cloud data.

[0015] The present invention is further configured as follows: the step 1) uses a ground-based laser radar to collect point cloud data of sample trees in the sample plot, specifically,

[0016] Stations are set up at the four corners and the center of the sample plot. Ground-based lidar equipment is installed at each station for scanning. The point cloud data of the sample trees in the sample plot are obtained by splicing the cloud data of each station. Each point has a three-dimensional spatial coordinate (x, y, z).

[0017] The present invention is further configured as follows: the step 2) pre-processes the point cloud data to obtain sample point cloud data, wherein the pre-processing includes denoising, point cloud normalization, and cropping.

[0018] The present invention is further configured as follows: the step 3) calculates the characteristic parameters of any point in the sample point cloud data based on the sample point cloud data, and composes the characteristic parameters into a characteristic matrix, wherein the characteristic parameters include one or more of flatness, structure tensor and normal vector, the structure tensor includes one or more of structure tensor flatness, "characteristic entropy" and linear features, and the normal vector includes one or more of the distribution of adjacent point characteristic vectors and normal vectors, and the spherical point normal vector.

[0019] The present invention is further configured as follows: the step 3) calculates the characteristic parameters of any point in the sample point cloud data based on the sample point cloud data, and composes the characteristic parameters into a characteristic matrix, specifically,

[0020] Based on the sample point cloud data, the KNN algorithm is used to find the N nearest points around any point p0. These N points and any point p0 form a set P = {p i |p i =(x i ,y i ,z i ),i=0,1,2,...,N}, construct the covariance matrix D of the set P, calculate the eigenvalues ​​λ1,λ2,λ3 and eigenvectors e1,e2,e3 of the matrix D, and then, for each point p i The eigenvalue λ i Perform normalization to obtain

[0021]

[0022]

[0023] Among them, λ1>λ2>λ3;

[0024] Calculate the characteristic parameters,

[0025] Flatness

[0026]

[0027] The structural tensor flatness k1,

[0028]

[0029] "Characteristic entropy" k2,

[0030]

[0031] Linear feature k3,

[0032]

[0033] Nearby point feature vectors e1, e2, e3;

[0034] The normal vectors of the spherical points are eq1, eq2, eq3,

[0035] Point p i The point set is composed of all points in the sphere with as the center and n meters (n∈R) as the radius. The covariance matrix is ​​constructed and eigenvalue decomposition is performed to obtain it;

[0036] The distribution of point normal vectors c1, c2, c3,

[0037] The adjacent point feature vectors are combined into an N×3 matrix and the covariance matrix L is constructed p , perform eigenvalue decomposition to obtain,

[0038]

[0039] In the formula, e j represents the jth eigenvector,

[0040] Construct the point cloud feature parameter matrix Spi of point pi,

[0041]

[0042] The present invention is further configured as follows: in step 4), any sample tree in the sample plot is selected, and the sample point cloud data covered by the sample tree is used as a training sample, and its main point cloud and non-main point cloud are marked, specifically,

[0043] Manually crop all point clouds of any sample tree Pi in the sample plot, use the sample tree Pi as a training sample, and classify and mark it as a trunk point cloud or a non-trunk point cloud.

[0044] The present invention is further configured as follows: the step 5) performs classification training on the sample point cloud data of the entire sample plot by machine learning based on the training samples and the feature matrix, specifically,

[0045] Based on the training samples and the feature matrix Spi, the sample point cloud data is machine-learned, and the support vector machine classifier is used to train the sample point cloud data to determine whether each point cloud in the sample point cloud data is a trunk point cloud, and the trunk point cloud is preliminarily extracted.

[0046] The present invention is further configured as follows: Step 6) performs Euclidean distance grouping and discrimination on all the marked main point cloud data in the sample plot, further extracts the main point cloud, and obtains the final main point cloud data, specifically,

[0047] 6-1) Group the initially extracted trunk point cloud, set the minimum Euclidean distance between points in different groups to n meters, and the value of n should be adjusted according to the size of the trunk branches. If the distance from any point to any point in a group is less than n meters, the point will be assigned to the group, otherwise it will be divided into other groups;

[0048] 6-2) Count the number of point clouds in each group, retain all groups with point cloud numbers greater than the specified value, and form the final backbone point cloud data.

[0049] The present invention is further configured such that: the specified value in step 6-2) is 100.

[0050] Compared with the prior art, the advantages of the present invention are: based on multiple representative feature vectors, it can effectively distinguish between trunk point clouds and non-trunk point clouds, and the marking training of single trees is applicable to the entire sample plot; machine learning methods are used to extract tree point clouds, and support vector machines have good classification performance for small samples and fast classification speed, providing support for efficient extraction of trunk point clouds. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flow chart of a method for extracting tree trunks from lidar point clouds based on machine learning in the present invention;

[0052] Figure 2 The figure shows the number of trunk point clouds of three sample trees in any sampling area. The oblique column is the trunk point cloud of the hand button, which is regarded as the true value. The point column is the point cloud extracted by this method, including the trunk point cloud and a small number of branch point clouds. The solid column is the trunk point cloud with the branch point cloud removed on the basis of the point column.

[0053] Figure 3 The original point clouds and final trunk extraction results of three sample trees in any sampling area. (a), (c), and (e) are the original point clouds of the three sample trees, and (b), (d), and (f) are the final trunk extraction results, among which (a) and (b) are sample tree 1, (c) and (d) are sample tree 2, and (e) and (f) are sample tree 3. DETAILED DESCRIPTION

[0054] The present invention will be further described below in conjunction with the accompanying drawings.

[0055] The present invention provides a method for extracting tree trunks from lidar point clouds based on machine learning. Compared with the prior art, the method can effectively distinguish between tree trunk point clouds and non-trunk point clouds based on multiple representative feature vectors, and the marking training of a single tree is applicable to the entire sample plot. The tree point clouds are extracted using a machine learning method, and the support vector machine has good classification performance for small samples and a fast classification speed, providing support for the efficient extraction of trunk point clouds.

[0056] A method for extracting tree trunks from lidar point clouds based on machine learning, comprising the following steps:

[0057] 1) Use ground-based LiDAR to collect point cloud data of sample trees in the sample plot.

[0058] Specifically, stations are set up at the four corners and the center of the sample plot, and ground-based lidar equipment is installed at each station for scanning. The point cloud data of the sample trees in the sample plot is obtained by splicing the cloud data of each station. Each point has a three-dimensional spatial coordinate (x, y, z).

[0059] 2) Perform preprocessing such as denoising, point cloud normalization, and cropping on the point cloud data to obtain the sample point cloud data.

[0060] 3) Based on the sample point cloud data, calculate the characteristic parameters of any point in the sample point cloud data, and form the characteristic parameters into a characteristic matrix;

[0061] Based on the sample point cloud data, the KNN algorithm is used to find the N nearest points around any point p0 (x0, y0, z0). These N points and any point p0 form a set P = {p i |p i =(x i ,y i ,z i ),i=0,1,2,...,N}, construct the covariance matrix D of the set P, calculate the eigenvalues ​​λ1,λ2,λ3 and eigenvectors e1,e2,e3 of the matrix D, and then, for each point p i The eigenvalue λ i Perform normalization to obtain

[0062]

[0063]

[0064] Among them, λ1>λ2>λ3;

[0065] Calculate feature parameters, wherein the feature parameters include one or more of flatness, structure tensor and normal vector, the structure tensor includes one or more of structure tensor flatness, "feature entropy" and linear features, the normal vector includes one or more of distribution of neighboring point feature vectors and normal vectors, and spherical point normal vectors,

[0066] Flatness

[0067]

[0068] The structural tensor flatness k1,

[0069]

[0070] "Characteristic entropy" k2,

[0071]

[0072] Linear feature k3,

[0073]

[0074] Nearby point feature vectors e1, e2, e3;

[0075] The normal vectors of the spherical points are eq1, eq2, eq3,

[0076] Point p i The point set is composed of all points in the sphere with as the center and n meters (n∈R) as the radius. The covariance matrix is ​​constructed and eigenvalue decomposition is performed to obtain it;

[0077] The distribution of point normal vectors c1, c2, c3,

[0078] The adjacent point feature vectors are combined into an N×3 matrix and the covariance matrix L is constructed p , perform eigenvalue decomposition to obtain,

[0079]

[0080] In the formula, e j represents the jth eigenvector,

[0081] Construct the point cloud feature parameter matrix Spi of point pi,

[0082]

[0083] 4) Select any sample tree in the sample plot, and the sample point cloud data covered by the sample tree is used as the training sample, and its trunk point cloud and non-trunk point cloud are marked. Specifically, all point clouds of any sample tree in the sample plot are manually cropped, and the sample tree is used as a training sample. Depending on whether it is marked as a trunk point cloud, the trunk point cloud is marked as 1 and the non-trunk point cloud is marked as 0.

[0084] 5) Based on the training samples and the feature matrix, the sample point cloud data of the entire sample plot is classified and trained by machine learning. Specifically, based on the training samples and the feature matrix Spi, the sample point cloud data is machine learned, and the support vector machine classifier is used to train the sample point cloud data to determine whether each point cloud in the sample point cloud data is a trunk point cloud, and the trunk point cloud is preliminarily extracted.

[0085] 6) Perform Euclidean distance grouping and discrimination on all the marked main point cloud data in the sample plot, further extract the main point cloud, and obtain the final main point cloud data, specifically,

[0086] 6-1) Group the initially extracted trunk point cloud, set the minimum Euclidean distance between points in different groups to n meters, and the value of n should be adjusted according to the size of the trunk branches. If the distance from any point to any point in a group is less than n meters, the point will be assigned to the group, otherwise it will be divided into other groups;

[0087] 6-2) Count the number of point clouds in each group, retain all groups with more than 100 point clouds, and form the final backbone point cloud data.

[0088] The present invention adopts a method for extracting tree trunks from laser radar point clouds based on machine learning, and a simulation example is used to verify the optimization design method of the present invention. The process is as follows: Figure 1 shown.

[0089] The sample plot of this embodiment is located in the Yellow Sea National Forest Park in Dongtai City, Jiangsu Province (32°33′~32°57′N, 120°07′~120°53′E).

[0090] RIEGL VZ-400i terrestrial 3D laser scanner was used to collect data from the poplar plantation in the 20m×20m sample plot. The laser emission frequency was set to 100kHz, and stations were set up at the four corners and the center. The point cloud data of each station was matched with the corresponding panoramic photos, and the point cloud was rendered in RGB to color the point cloud. The overlapping point clouds between adjacent stations were used for splicing. Riscan Pro 64bit v2.6.1 software will automatically splice the data according to the error range allowed in the algorithm. For stations with long distances or a small number of overlapping point clouds, reference objects are manually found for splicing.

[0091] During implementation, N is set to 7, and n is set to 0.1; when counting the number of point clouds in each group, all groups with point cloud numbers greater than 100 are retained. The present invention randomly extracts point clouds of three poplar trees in the sampling area, and the total point cloud volumes are sample tree 1-137622, ​​sample tree 2-323873, and sample tree 3-197776, respectively. The manually deducted trunk point cloud volumes are sample tree 1-84704, sample tree 2-189335, and sample tree 3-126242, respectively. The feature parameters are calculated for each point to form a feature parameter matrix. Through machine learning, the final extracted trunk point cloud volumes are sample tree 1-81495, sample tree 2-178256, and sample tree 3-120220. The extraction results are as follows. Figure 2 As shown in the solid bars, the complete extraction rates were 96.21%, 94.15%, and 95.23%, respectively. Figure 3The original point clouds and final trunk extraction results of three randomly selected poplar tree samples, where (a), (c), and (e) are the original point clouds of the three sample trees, and (b), (d), and (f) are the final trunk extraction results, where (a) and (b) are sample tree 1, (c) and (d) are sample tree 2, and (e) and (f) are sample tree 3.

[0092] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments. The above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for extracting tree trunks from lidar point clouds based on machine learning, characterized in that: The following steps are involved: 1) Use ground-based LiDAR to collect point cloud data of sample trees in the sample plot; 2) Preprocess the point cloud data to obtain sample point cloud data; 3) Based on the sample point cloud data, calculate the characteristic parameters of any point in the sample point cloud data, and form the characteristic parameters into a characteristic matrix; 4) Select any sample tree in the sample plot. The sample point cloud data covered by the sample tree is the training sample, and mark its trunk point cloud and non-trunk point cloud; 5) Based on the training samples and feature matrix, the sample point cloud data of the entire sample plot is classified and trained by machine learning; 6) Perform Euclidean distance grouping and discrimination on all the marked main point cloud data in the sample plot, further extract the main point cloud, and obtain the final main point cloud data; The step 1) uses a ground-based laser radar to collect point cloud data of sample trees in the sample plot, specifically, Stations are set up at the four corners and the center of the sample plot. Ground-based laser radar equipment is set up at each station for scanning. The point cloud data of the sample trees in the sample plot is obtained by splicing the cloud data of each station. Each point has a three-dimensional spatial coordinate (x, y, z); The step 2) preprocesses the point cloud data to obtain sample point cloud data, wherein the preprocessing includes denoising, point cloud normalization, and cropping; The step 3) calculates the characteristic parameters of any point in the sample point cloud data based on the sample point cloud data, and composes the characteristic parameters into a characteristic matrix, wherein the characteristic parameters include one or more of flatness, structure tensor and normal vector, the structure tensor includes one or more of structure tensor flatness, "characteristic entropy" and linear features, and the normal vector includes one or more of the distribution of adjacent point characteristic vectors and normal vectors, and the spherical point normal vector; The step 3) calculates the characteristic parameters of any point in the sample point cloud data based on the sample point cloud data, and composes the characteristic parameters into a characteristic matrix, specifically, Based on the sample point cloud data, the KNN algorithm is used to find the N nearest points around any point p0. These N points and any point p0 form a set , construct the covariance matrix D of the set P, and calculate the eigenvalues ​​of the matrix D With the eigenvector , then, for each point p i The eigenvalue of Perform normalization to obtain , in, ; Calculate the characteristic parameters, Flatness , ; Structure Tensor Flatness , ; "Characteristic Entropy" , ; Linear Features , ; Nearby point feature vector ; Normal vector of spherical point , Point p i The point set is composed of all points in the sphere formed by the center of the sphere and the radius of n meters (n∈R). The covariance matrix is ​​constructed and the eigenvalue decomposition is performed to obtain the result; Distribution of point normal vectors , The adjacent point feature vectors are combined into an N×3 matrix and the covariance matrix L is constructed p , perform eigenvalue decomposition to obtain, In the formula, represents the jth eigenvector, Construct point p i The characteristic matrix Spi, 。 2. The method for extracting tree trunks from laser radar point clouds based on machine learning according to claim 1, characterized in that: In step 4), any sample tree in the sample plot is selected, and the sample point cloud data covered by the sample tree is used as a training sample, and its main point cloud and non-main point cloud are marked, specifically, Manually crop all point clouds of any sample tree Pi in the sample plot, use the sample tree Pi as a training sample, and classify and mark it as a trunk point cloud or a non-trunk point cloud.

3. The method for extracting tree trunks from laser radar point clouds based on machine learning according to claim 2, characterized in that: The step 5) performs classification training on the sample point cloud data of the entire sample plot by machine learning based on the training samples and the feature matrix, specifically, Based on the training samples and the feature matrix Spi, the sample point cloud data is machine-learned, and the support vector machine classifier is used to train the sample point cloud data to determine whether each point cloud in the sample point cloud data is a trunk point cloud, and the trunk point cloud is preliminarily extracted.

4. The method for extracting tree trunks from laser radar point clouds based on machine learning according to claim 3, characterized in that: The step 6) performs Euclidean distance grouping and discrimination on all the marked main point cloud data in the sample plot, further extracts the main point cloud, and obtains the final main point cloud data, specifically, 6-1) Group the initially extracted trunk point cloud, set the minimum Euclidean distance between points in different groups to n meters, and the value of n should be adjusted according to the size of the trunk branches. If the distance from any point to any point in a group is less than n meters, the point will be assigned to the group, otherwise it will be assigned to other groups; 6-2) Count the number of point clouds in each group, retain all groups with point cloud numbers greater than the specified value, and form the final backbone point cloud data.

5. The method for extracting tree trunks from laser radar point clouds based on machine learning according to claim 4, characterized in that: The specified value in step 6-2) is 100.

Citation Information

Patent Citations

  • Laser point cloud-oriented single tree segmentation method based on Faster R-CNN

    CN110378909A

  • Border tree trunk extraction method based on mobile laser scanning point cloud data

    CN110910407A

  • Forest site cloud data registration and segmentation method based on ground-based laser radar

    CN112381861A

  • Device and method for object detection in a vehicle environment

    DE102020000490A1