Single tree skeleton line extraction method based on ground laser radar point cloud

Abstract

The invention discloses a ground laser radar point cloud-based individual tree skeleton line extraction method. The method comprises the following steps of S1, refining a point cloud by adopting an improved Laplacian-based shrinkage method; s2, acquiring an initial skeleton point by using a self-adaptive farthest sphere sampling method; firstly, an improved Laplacian contraction method is adopted to contract a point cloud to a middle axis in a branch point cloud, and self-adaptive farthest ball sampling is utilized to preliminarily extract skeleton points; wherein each skeleton point corresponds to one contraction point set. In order to extract a skeleton line more accurately, an angle-based adaptive adjustment strategy is carried out to adjust an initial skeleton point, and a final centralized skeleton point is obtained. Compared with a simple neighborhood determination method, the method for determining the neighborhood based on the hybrid model can obtain more complete neighborhood information while reducing the calculation complexity. Through the square of the distance between the points and the new edge weight combination of the geodesic distance, a single-tree skeleton which better accords with the real shape of the tree can be extracted.

Description

Technical Field

[0001] This invention relates to the technical field of single-tree skeleton line extraction methods, and in particular to a single-tree skeleton line extraction method based on ground-based lidar point clouds. Background Technology

[0002] Ground-based lidar, as an important active remote sensing technology, is widely used in various fields due to its high precision and efficiency. In forest survey missions, ground-based lidar can quickly acquire target data.

[0003] The three-dimensional coordinate information of the target scene helps in obtaining forest structure parameters and conducting carbon cycle analysis. Through dense point clouds, ground-based lidar technology provides crucial data support for forest resource management, playing a key role, particularly in individual tree modeling and forest parameter estimation. In forestry applications, individual tree modeling is a critical component for achieving accurate results. By modeling individual trees, parameters such as ground biomass and stand volume can be estimated, and the accurate calculation of diameter at breast height (DBH) also depends on the precision of the modeling process.

[0004] In many single-tree modeling methods, skeleton line extraction is a crucial step, playing a key role in building accurate models of individual trees. By extracting the tree's skeleton lines, the shape and structure of the tree can be described more accurately, enabling precise calculation and estimation of various tree parameters. This method not only provides a better understanding of forest structure and characteristics but can also be used to build 3D tree models and analyze the impact of environmental factors, providing important data support for forest resource management and protection.

[0005] Current skeleton extraction methods can be categorized into cluster-based segmentation methods and point cloud refinement-based methods. Cluster-based segmentation methods directly divide the point cloud into different clusters and extract skeleton points based on the cluster information. Point cloud refinement-based methods move the point cloud closer to its central axis before extracting skeleton points. Current skeleton extraction methods struggle to balance fast and high-quality extraction. Most methods rely on single neighborhood determination methods, which fail to address issues such as noise, outliers, and data gaps, and cannot obtain complete neighborhood information.

[0006] Tree growth environment, data scanning methods, and even self-occlusion of fine branches can significantly increase the difficulty of refining the skeleton, while currently used weighting methods can also lead to topological or morphological errors. Summary of the Invention

[0007] The purpose of this invention is to propose a method for extracting single-tree skeleton lines based on ground-based lidar point clouds, so as to solve the problems of large computational load and low efficiency in existing technologies.

[0008] A method for extracting single-tree skeleton lines based on ground-based lidar point clouds includes the following steps:

[0009] S1, a point cloud is refined using an improved Laplacian-based shrinkage method;

[0010] S2, the initial skeleton points are obtained using the adaptive farthest sphere sampling method and then centered.

[0011] S3 uses a hybrid model to determine neighborhood vertex information;

[0012] S4 extracts the single-wood skeleton by combining the weighted combination of the square of the Euclidean distance and the geodesic distance.

[0013] The single-tree skeleton line extraction method based on ground-based lidar point clouds provided by the present invention firstly uses a Laplacian contraction method improved based on weight adjustment factors and an adaptive farthest sphere based on geometric characteristics.

[0014] Sampling is used to initially extract skeleton points. Then, angle constraints are used to iteratively adjust the skeleton points, achieving centering, especially when branch points are incomplete. To extract skeleton lines, a hybrid neighborhood determination model is proposed, which maintains graph structure simplicity while accurately connecting branch nodes. Furthermore, a new edge weight matrix definition is introduced, a weighted combination of the squared Euclidean distance and the geodesic distance. This method ensures the construction of correctly connected skeleton lines, especially for branches within the canopy. This invention achieves automated extraction...

[0015] The process of obtaining a more realistic tree skeleton and improving the representation of topological details is simple and easy to implement, and can provide a good foundation for the subsequent application of terrestrial LiDAR in forest areas.

[0016] In addition, the single-wood skeleton line extraction method based on ground-based lidar point clouds according to the present invention may also have the following additional technical features:

[0017] Furthermore, in step S1, the improved Laplace equation is calculated using the following formula:

[0018]

[0019] In the formula, κ represents the weight adjustment factor, P is the shrinkage result of the point set P for a 3D point cloud, L is the Laplacian matrix based on the curvature flow operator, and W... L and W H These are the diagonal matrices representing the equilibrium contraction and attraction constraints, respectively.

[0020] Furthermore, in step S1, the weight adjustment factor K takes values ​​in the range of [0,1], and is calculated using the following formula:

[0021] K = (λ1 - λ2) / λ1

[0022] In the formula, λ1 and λ2 are two eigenvalues ​​obtained by calculating the covariance matrix CPi from the neighboring points of point Pi (i = 1, 2, ..., n).

[0023] Furthermore, in step S2, the sampling radius is calculated using the following formula for adaptive farthest sphere sampling:

[0024]

[0025] In the formula, S is the shape characteristic function constructed from eigenvalues, i.e., S = λ1 / (λ1 + λ2 + λ3), where λ1 ≥ λ2 ≥ λ3, r max It is the maximum sampling radius.

[0026] Furthermore, in step S2, reliable skeleton points and undetermined skeleton points are distinguished by whether the point cloud of the contracted point set is complete.

[0027] Furthermore, in step S2, reliable skeleton points are centered using circle fitting. Based on angle constraints, the undetermined skeleton points are adjusted to the point and its projection point (projected along the line connecting two adjacent reliable points).

[0028] At the center of 10, the process iterates until the condition is met and then stops.

[0029] Furthermore, in step S3, the R-nearest neighbor algorithm is used to select a larger radius parameter r to determine the r-neighborhood information. Based on this, Delaunay triangulation is performed, and the ring neighborhood of a vertex is taken as the adjacent vertex.

[0030] Furthermore, in step S4, the edge weights are calculated using the following formula for shortest path analysis:

[0031]

[0032] Where, d ij It is p i and p j The Euclidean distance between two skeleton points, D ij It is p i and p j Two skeletal points

[0033] The geodesic distance between them, that is, in In the graph structure created as edge weights, p is calculated. i and p j Two skeleton points

[0034] The path distance between them. Attached Figure Description

[0035] The above and / or additional aspects and advantages of the embodiments of the present invention will become apparent and readily understood from the description of the embodiments in conjunction with the following drawings, wherein:

[0036] Figure 1 This is a flowchart illustrating the method for extracting single-tree skeleton lines based on ground-based lidar point clouds according to an embodiment of the present invention.

[0037] Figure 2 The diagrams illustrate Laplace contraction based on different operators: (a) branch point cloud colored according to elevation values; (b) branch point cloud colored according to linear eigenvalues; (c) result of Laplace contraction of branch point cloud using cotangent weights; and (d) result of Laplace contraction of branch point cloud using a linear weight adjustment factor.

[0038] Figure 3 A schematic diagram of the process of obtaining the initial skeleton points for adaptive farthest sphere sampling: (a) different tree point cloud densities at branches and forks after shrinkage; (b) different sampling scale settings at branches and forks; (c) adaptive farthest sphere sampling; (d) initial skeleton points for adaptive farthest sphere sampling.

[0039] Figure 4 A schematic diagram of the process of obtaining the initial skeleton points for adaptive farthest sphere sampling: (a) Graph construction result after setting r to three times the average point spacing; (b) Graph construction result after setting r to 10 times the average point spacing; (c) Hybrid neighborhood determination model combining r nearest neighbors and single-ring neighborhoods; (d) Graph construction result of the hybrid neighborhood determination model.

[0040] Figure 5 The diagrams illustrate the skeleton extraction for different edge weight matrices. (a) The edge weight value is defined as the skeleton line extracted by the Euclidean distance between nodes; (b) The edge weight value is defined as the skeleton line extracted by the square of the Euclidean distance between nodes; (c) The edge weight value is defined as the skeleton line extracted by the square of the Euclidean distance between nodes and the geodesic distance. Detailed Implementation

[0041] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0042] In view of the problems existing in the prior art, this invention provides a method for extracting the skeleton line of a single tree based on ground-based lidar point clouds. In the process of obtaining the single tree model using this method, the determination of the neighborhood by a hybrid model and the definition of novel weights are the core of the method. First, an improved Laplace shrinkage method based on the introduction of a linear weight adjustment factor is used to enhance the robustness of shrinkage to incomplete branch points. Then, the sampling radius is automatically calculated adaptively using the different density distributions of points corresponding to geometric features at the bifurcation and branch sections.

[0043] The farthest sphere sampling is used to extract initial skeleton points, reducing the dependence on the sampling radius parameter. Simultaneously, centering adjustments are necessary to ensure that each extracted skeleton point is located at the center of its corresponding set of contraction points, especially when branch points are incomplete. To extract skeleton lines, a hybrid neighborhood determination model is proposed, which involves Delaunay triangulation based on the R-neighborhood to obtain the one-ring neighborhood of the vertex and...

[0044] Adjacent vertices are treated as adjacent vertices. Furthermore, a novel edge weight matrix definition is introduced, which is a weighted combination of the squared Euclidean distance and the geodesic distance. This skeleton graph optimization method ensures the construction of correctly connected skeleton lines, improving the robustness of obtaining branch topology and skeleton line extraction.

[0045] Please combine Figure 1 The single-tree skeleton line extraction method based on ground lidar point cloud provided in this embodiment of the invention includes the following steps S1 to S4.

[0046] S1 uses an improved Laplacian-based shrinkage method to refine the point cloud.

[0047] As mentioned earlier, skeleton points play a crucial role in acquiring individual tree models. Using skeleton points instead of point cloud construction to form skeleton lines can significantly reduce computational burden and improve efficiency. The Laplacian shrinkage method based on curvature flow operators is a normal vector-based shrinkage method. Compared to traditional Laplacian shrinkage methods based on cotangent weights, the curvature flow operator-based Laplacian shrinkage method reduces the influence of point cloud mesh shape, such as... Figure 2 (c) and Figure 2 As shown in (d), the Laplacian shrinkage method based on the curvature flow operator can generally be applied to point cloud shrinkage with different point density distributions and different degrees of missing data. The Laplacian shrinkage method based on the curvature flow operator is an iterative method. In each iteration, the Laplacian matrix based on the curvature flow operator is first calculated. After each iteration, the diagonal matrix of the shrinkage and attraction constraints for the next iteration needs to be updated. Therefore, after several iterations, these points will be moved to the corresponding shrinkage positions, resulting in a shrinkage point set with approximately zero volume.

[0048] The improved Laplace equation is calculated according to formula (1).

[0049]

[0050] In the formula, κ represents the weight adjustment factor, and for a 3D point cloud, P is the result of shrinking the point cloud P.

[0051] L is the Laplacian matrix based on the curvature flow operator, W L and W H These are the diagonal matrices representing the balanced shrinkage and attraction constraints, respectively. The distribution of branch points directly affects the shrinkage result, especially in incomplete point clouds with a single scan viewpoint or branch self-occlusion. Thicker regions are subject to stronger shrinkage constraints, causing vertices in thinner regions near their neighborhoods to deviate from their central positions. This invention aims to shrink tree point clouds to approximately the central axis using an improved Laplacian shrinkage method, facilitating the extraction of linear skeletons. Figure 2 (b) shows the branch point cloud of a single station. Edge points have larger linear eigenvalues ​​than non-edge points. It is hoped that points located far from the central axis...

[0052] Larger contraction constraints are set at edge points. Therefore, the weight adjustment factor κ is determined by the linear eigenvalues ​​of each point. Generally, κ∈[0,1], and the weight adjustment factor is set using formula (2).

[0053] K=(λ1-λ2) / λ1(2)

[0054] In the formula, λ1 and λ2 pass through point P. i Calculate the covariance matrix of adjacent points (i = 1, 2, ..., n). The two eigenvalues ​​obtained.

[0055] S2 uses an adaptive farthest sphere sampling method to obtain the initial skeleton points and then performs centering adjustments.

[0056] After the point cloud is shrunk, the result is still not a one-dimensional skeleton. The point densities of branches and forks are different, and using a single sampling scale cannot guarantee that the final one-dimensional skeleton can accurately capture the details of the branch structure. Therefore, different sampling scales are used for branches and forks. According to formula (3), the sampling radius is calculated by utilizing the differences in geometric features between forks and branches, and adaptive farthest sphere sampling is used to obtain the initial skeleton points.

[0057]

[0058] In the formula, S is the shape characteristic function constructed from eigenvalues, i.e., S = λ1 / (λ1 + λ2 + λ3), where λ1 ≥ λ2 ≥ λ3, r max It is the maximum sampling radius.

[0059] Figure 3 This describes the process of extracting skeleton points based on adaptive farthest sphere sampling. Figure 3 (a) shows the different tree point cloud densities at branches and forks after contraction. Figure 3 (b) shows the different sampling scale settings at the branches and forks. It can be seen that a smaller sampling scale is used at the forks, and a larger sampling scale is used at the branches. After automatically calculating the radius of the sampling sphere, the point within the sampling sphere farthest from the previous skeleton point is taken as the new skeleton point, such as... Figure 3 As shown in (c) Figure 3 (d) is the skeleton point after the final adaptive farthest sphere sampling.

[0060] After obtaining the initial skeleton points, the improved Laplacian shrinkage method cannot guarantee that the extracted skeleton points are located at the center of their corresponding shrunken point set, especially when processing point clouds at bifurcation points. This is because during the shrinkage process, regions with high point cloud density are subject to strong shrinkage constraints, causing vertices in thinner regions near their neighborhoods to deviate from their center positions. Therefore, it is necessary to adjust the center of the extracted skeleton points to ensure they are located at the center of their respective local point sets.

[0061] S3 uses a hybrid model to determine neighborhood vertex information.

[0062] The commonly used R-nearest neighbor algorithm for determining neighborhoods significantly affects node connectivity when the distance parameter r is chosen. For example... Figure 4 As shown in (a), setting r to 3 times the average node spacing will cause some distant nodes to be unable to connect, resulting in an incomplete branch topology. Conversely, if r is set to 10 times the average node spacing... Figure 4 (b) Although it connects distant nodes, it makes the graph structure more complex and causes a large computational burden.

[0063] Therefore, a hybrid neighborhood model is used to determine node connectivity. For example... Figure 4 As shown in (c), Delaunay triangulation is performed on the R neighborhood to obtain the ring neighborhood of the vertex as the adjacent vertex.

[0064] By employing this hybrid neighborhood model, the constructed graph accurately captures the branching topology while maintaining structural simplicity.

[0065] S4 extracts the single-wood skeleton by combining the weighted combination of the square of the Euclidean distance and the geodesic distance.

[0066] Figure 4 This represents the result of shortest path analysis performed on skeleton points using different edge weights. Figure 5(a) defines the weight as the Euclidean distance between connected nodes. This method can capture the basic topology of the branches, but it may connect the spatially closest nodes even if they belong to different branches, thus forming incorrect branches. Figure 5 (b) defines the weight as the square of the Euclidean distance between connected nodes. Although it can effectively express the branch topology, it is easily affected by outliers, causing the skeleton line to fail to show the characteristics of smooth branch growth.

[0067] Since using Euclidean distance as edge weights leads to a "coarse" branch skeleton, and using the square of Euclidean distance as edge weights leads to an "overfitted" branch skeleton, neither fully reflects the true branch topology. Therefore, a new method for defining the edge weight matrix is ​​proposed. First, a graph structure is created using the square of Euclidean distance as the edge weights. Based on this graph, Dijkstra's shortest path algorithm is used to calculate the geodesic distance D between each pair of nodes. ij The edge weights between nodes are calculated according to formula (4).

[0068]

[0069] Where, d ij It is p i and p j The Euclidean distance between two skeleton points, D ij It is p i and p j The geodesic distance between two skeleton points, that is, in In the graph structure created as edge weights, p is calculated. i and p j The path distance between two skeleton points.

[0070] The evolutionary steps of skeleton extraction involve obtaining the shortest weighted path for each point by combining the weighted sum of the squared Euclidean distance and the geodesic distance. This approach can extract relatively accurate skeleton lines, such as... Figure 4 As shown in (c).

[0071] Evolutionary steps of skeleton extraction

[0072] Input: Adaptively adjusted skeleton point p i i = 1, 2, ..., n

[0073] The ring of the skeleton point p

[0074] Step 1: Based on the adjacency information of the skeleton points, use the square of the Euclidean distance. As edge weight, creation

[0075] Build the graph structure and perform shortest path analysis;

[0076] Based on the calculation of the geodesic distance D between connected skeleton points ij That is, the shortest path distance between two points;

[0077] Step 2: Update the edge weights between the two points according to formula (4).

[0078] Step 3: Perform the shortest path analysis again based on the new edge weight information.

[0079] Output: Single-wood frame wire

[0080] In summary, the single-tree skeleton line extraction method based on terrestrial lidar point clouds provided by this invention first uses an improved Laplacian shrinkage method based on weight adjustment factors and adaptive farthest sphere sampling to initially extract skeleton points. Simultaneously, a centering adjustment is performed to ensure that each extracted skeleton point is located within its corresponding shrinkage point.

[0081] The method focuses on the center of the point set, especially when branch points are incomplete. To extract skeleton lines, a hybrid neighborhood model is proposed that maintains the simplicity of the graph structure while accurately connecting branch nodes. Furthermore, a new edge weight matrix definition is introduced, a weighted combination of the squared Euclidean distance and the geodesic distance. This method ensures the construction of correctly connected skeleton lines, particularly for branches within the tree canopy. This approach automates the extraction of more realistic tree skeletons and improves topological detail representation. The process is simple and easy to implement, providing a solid foundation for subsequent applications of terrestrial LiDAR in forest areas.

[0082] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0083] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A single tree skeleton line extraction method based on ground laser radar point cloud, characterized in that, The method comprises the following steps: S1, using a shrinkage method based on improved Laplace to refine the point cloud; S2, using an adaptive farthest ball sampling method to obtain initial skeleton points and performing centralization adjustment; S3, using a hybrid model to determine neighborhood vertex information; S4, using a weight combination of the square sum of Euclidean distance and geodesic distance to extract a single tree skeleton.

2. The method of claim 1, wherein, In step S1, the point cloud is shrunk by using the following improved Laplace equation: wherein, κ denotes a weight adjustment factor, for a three-dimensional point cloud, is a set of points P resulting from a shrink operation, L is a Laplacian matrix based on a curvature flow operator, and are diagonal matrices of equilibrium shrink and attraction constraints, respectively.

3. The method of claim 2, wherein, In step S1, the range of the weight adjustment factor is: κ ∈ [0, 1], the weight adjustment factor is calculated using the following formula K=(λ1−λ2) In the formula λ, λ 1 and λ 2 is through points Pi ( i =1,2,..., n Calculate the covariance matrix of adjacent points. The two eigenvalues ​​obtained.

4. The method of claim 1, wherein, In step S2, the geometric feature difference of the bifurcation point and the branch point is used to calculate the shape feature value, and the sampling radius is calculated according to the following formula to realize adaptive farthest ball sampling to obtain initial skeleton points: wherein S is a shape feature function constructed from eigenvalues, i.e. S = λ1 (λ1+ λ2+ λ3), where λ1≥ λ2≥ λ3, is the maximum sampling radius.

5. The method of claim 4, wherein, In step S2, the complete skeleton points of the shrunk point set are adjusted to the center by using circle fitting, and for the skeleton points of the incomplete shrunk point set, they are adjusted to the midpoint of the projection point in the direction of the adjacent two Reliable point connection, and the process is iterated until it meets the angle constraint, and the adjustment is completed.

6. The method of claim 1, wherein, In step S3, a larger radius parameter r is selected to determine the r neighborhood information, and on this basis, Delaunay triangulation is performed, and the one-ring neighborhood of the vertex is taken as the adjacent vertex.

7. The method of claim 1, wherein, In step S4, the weight is calculated according to the following formula to perform shortest path analysis: where, is pi and pj the Euclidean distance between two skeleton points, is pi and pj the geodesic distance between two skeleton points, i.e., in the graph structure created as edge weights, the path distance between two skeleton points is computed. pi and pj the path distance between two skeleton points.

Images