Fractal growth model and refined tree modeling method

Abstract

The invention discloses a fractal growth model and refined tree modeling method, and belongs to the technical field of three-dimensional reconstruction, and the method comprises the steps: obtaining an original point cloud and an image set of a single tree in a reconstruction region, and the ground height information of each spatial position; data preprocessing and point cloud data segmentation; extracting a tree skeleton from the non-ground point cloud based on a graph theory, a Dijkstra algorithm and a cubic spline interpolation algorithm; layering reconstruction is carried out on the tree, a three-dimensional grid model, a branch cylinder model and a small branch model of a trunk are generated and fused into a tree model, texture mapping is carried out based on NeRF, and a comprehensive tree model is obtained. According to the method, a graph theory method is introduced as a key link in a skeleton extraction process, a tree skeleton is accurately extracted from a non-ground point cloud, a hierarchical reconstruction method is adopted, accurate extraction of the skeleton and tree modeling are realized, a NeRF technology is introduced during texture mapping, and modeling precision and texture quality are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of three-dimensional reconstruction, and in particular to a fractal growth model and a refined tree modeling method. Background Art

[0002] With the rapid development of 3D modeling and computer vision technologies, deep learning-based 3D tree modeling and texture mapping have shown tremendous potential in forestry management, ecological protection, and the development of digital twin forest areas. In forestry modeling and ecological assessment, ensuring the accuracy and detail of tree models is crucial for carbon sequestration assessments, green volume calculations, and planning conflict detection. Refined tree modeling often requires texture realism, representation of tree growth patterns, and terrain integration to meet the needs of forest resource management and ecological protection.

[0003] Current methods for tree modeling include: for example, patent application publication number CN112634432B. This method relies on manually drawing branch trajectories in a virtual environment, resulting in low modeling efficiency, regularized structural morphology, lack of real terrain constraints, and inability to achieve automatic reconstruction based on real-world data. For example, patent application publication number CN104392484B generates a new model by extracting components from existing tree models and connecting them with interpolation curves. Although this simplifies the three-dimensional data acquisition process, it relies heavily on a preset model library, lacks adaptability to real terrain, cannot reflect the actual tree structure, has limited morphological accuracy and naturalness, and does not support automated modeling.

[0004] Therefore, traditional modeling methods have difficulty in handling complex tree shapes and detail differences, and cannot guarantee the lighting consistency and natural transition of textures.

[0005] The RANSAC (Random Sample Consensus) algorithm calculates mathematical model parameters based on a sample dataset containing outliers to generate valid sample data. Point cloud data of a single tree contains a large number of ground points, which results in a large and cluttered point cloud and hinders subsequent point cloud classification and recognition. Therefore, these ground points must be removed first. The RANSAC algorithm is a commonly used method for removing these ground points.

[0006] NeRF (Neural Radiance Field) is a computer vision technology used to generate high-quality 3D reconstruction models. It uses a neural network (MLP) to model the implicit radiance field (density and color) of a scene. Its input is: multi-view 2D images and corresponding camera parameters. The rendering process is: sampling 3D points along the light ray, predicting the density and color of each 3D point through a neural network, and finally generating pixel color using volume rendering integration.

[0007] Dijkstra's algorithm, also known as Dijkstra's algorithm, is a typical shortest path algorithm that calculates the shortest path from a source node to other target nodes. It generates a tree structure consisting of the shortest paths from the source node to all other reachable nodes. The shortest path tree is a collection of the shortest paths from the source node to all nodes, organized in a tree structure. The tree satisfies the following conditions: the path from the source node to any node is the shortest path, and each node in the tree has only one parent node (no cycles).

[0008] Skeleton curves are widely used in real-time simulation animation synthesis, shape recognition, image segmentation, and 3D modeling. Skeleton curves typically consist of a set of interconnected line segments that connect important feature points of an object. Skeleton curves can capture an object's basic shape structure and topological relationships. In computer vision, skeleton curves are used to describe the characteristic contours of an object, enabling object detection, recognition, and reconstruction.

[0009] The Edge Split algorithm is a commonly used technique in mesh processing. It is mainly used to split the edges in the mesh. Specifically, a new vertex is inserted on the edge and the original triangle is split to form new triangles, thereby increasing the complexity and details of the mesh.

[0010] Fractal theory, proposed by mathematician B.B. Mandelbrot based on fractal geometry, studies the properties of fractals. Fractals originated from the coastline problem: the self-similarity of parts and the whole of a coastline. Mandelbrot defined such shapes as fractals, which resemble each other in some way. Because fractals are ubiquitous in nature, fractal theory has found numerous applications in biology and urban studies, such as in areas like organism growth, urban expansion, and digital image processing. Currently, established methods for modeling tree growth using fractals exist, such as the "Fractal-Based Tree Growth Modeling Method" proposed by Zhang Chengxin, Hu Xiaofang, and others. This method addresses the complexity and challenges of tree growth modeling by employing fractal methods. Based on the natural growth characteristics of tree branches and leaves, the leaf growth process is simulated, a fractal algorithm is used to generate the main branches, and a logistic model is used to simulate the tree's growth process. By varying parameters such as leaf size, tree depth, curvature, and tree size, an interactive dynamic growth model is implemented, which realistically depicts the tree's growth process. Summary of the Invention

[0011] The purpose of the present invention is to provide a fractal growth model and a refined tree modeling method that solves the above problems, which can not only realize refined modeling of trees, but also learn the lighting and color information in three-dimensional space, thereby generating a highly realistic texture effect.

[0012] In order to achieve the above-mentioned object, the technical solution adopted by the present invention is as follows: a fractal growth model and a refined tree modeling method, comprising the following steps; S1, obtain the original point cloud, image set, and ground height information h(x, y) of each spatial position in the reconstruction area for a single tree in the reconstruction area. The image set includes the original images from multiple perspectives used for NeRF reconstruction. h(x, y) is the ground height of each spatial position in the area where the tree is located. S2, data preprocessing and point cloud data segmentation; Preprocessing the original image and the original point cloud to obtain a two-dimensional image and a first point cloud; Fitting the ground plane based on the RANSAC algorithm, the first point cloud is segmented into ground point cloud and non-ground point cloud; Based on the RANSAC algorithm, the cylinder is fitted to divide the ground point cloud into the trunk point cloud and the branch point cloud; S3, extracting tree skeletons from non-ground point clouds, including S31~S33; S31, based on the Delaunay triangulation method, construct the adjacency graph G=(V,E) of the non-ground point cloud, where V is the vertex set and E is the edge set. For two vertices p i and p j , if there is an edge e between the two ij , then e ij The edge weight is w ij , , where 、 p i and p j Coordinates in the world coordinate system, ‖∙‖2 is the Euclidean distance; S32, select a point close to the ground as the root node, use Dijkstra algorithm to generate the shortest path tree of the root node, and use the set of nodes in the shortest path tree as the initial skeleton S initial ; S33, generate the initial skeleton S through the cubic spline interpolation algorithm initial The smoothed skeleton curve Q'(t); S4, hierarchical reconstruction of trees, including S41 to S44; S41, based on the ground height information constraint, use Poisson reconstruction to model the tree trunk point cloud and generate a 3D mesh model of the tree trunk M trunk ; S42, presetting a density threshold, dividing the points in the branch point cloud that are greater than the density threshold into the main branch point cloud P1, and the rest into the small branch point cloud P2; S43, reconstruct P1 using cylindrical fitting to generate a branch cylindrical model M branch ; S44, based on the fractal theory, simulates the growth of small branches in P2 and generates a small branch model M twig ; S5, for M trunk 、M branch and M twig Fusion is performed to obtain the tree model M tree ; S6, for tree model M tree Texture mapping is performed on each point to generate a comprehensive tree model .

[0013] Preferably, in S2, preprocessing the original image is converting the original image into a world coordinate system; The original point cloud is pre-processed as follows: the original point cloud is converted into a world coordinate system, the point cloud data is processed by a clustering algorithm, noise points are removed, and the first point cloud is obtained.

[0014] As a preference, in S33, an initial skeleton S is generated. initial The smoothed skeleton curve Q'(t) is specifically: Sb1, from S initial Extract discrete point sets from ; Sb2, use the cubic spline function Q(t) to interpolate and smooth the discrete point set to obtain the skeleton curve, where Q(t)=A⋅[t 3 ,t 2 ,t 1 ,1] T ,t∈[0,1], A is the coefficient matrix, t is the local normalization parameter; Sb3, optimize the coefficient matrix A so that Q(t) approaches S initial The final skeleton curve is marked as the smoothed skeleton curve Q'(t).

[0015] As a preference, in S4, the Poisson equation for Poisson reconstruction According to the following formula: , Where, is the volume field, ∇ 2 is the Laplace operator, ∇ is the divergence operator, is the normal vector field of the non-ground point cloud, λ is the control coefficient of the terrain fitting strength, and z is the height of each point in Q'(t).

[0016] As a preferred method, S43 reconstructs P1 using cylindrical fitting to generate a branch cylindrical model M branch ; Specifically; Sc1, the main branch point cloud P1 is divided into several sub-point clouds based on the clustering algorithm; Sc2, for each sub-point cloud, a cylindrical model is generated using a weighted regularized cylindrical fitting method, where the objective function F of the weighted regularized cylindrical fitting method is obtained according to the following formula; , Where n is the total number of points in the sub-point cloud, w i is the i-th point p in the sub-point cloud i The weight of dist(p i ,Cylinder(θ)) is point p i The distance to the cylinder, γ is the regularization parameter, θ is the cylinder parameter, θ prior is the branch prior parameter, is the square of the L2 norm; Sc3, the collection of cylindrical models of all sub-point clouds constitutes M branch。

[0017] As a preference, S5 to M trunk 、M branch and M twig Carry out integration, specifically: S51, splicing M trunk 、M branch and M twig and use weighted average fusion method to eliminate the seams; S52, using an edge segmentation algorithm to segment the fused edges, thereby achieving mesh encryption at the joints.

[0018] As a preferred embodiment, S6, for the tree model M tree Texture mapping is performed on each point to generate a comprehensive tree model , specifically; The illumination model of M tree Each point in generates light intensity at different viewing angles, and the light intensity of point p at viewing angle v is L(v,p); Pre-train the NeRF model with the image set to generate M tree The color of each point at different viewing angles, the color of point p at viewing angle v is C(r); Generate texture information of each point at different viewing angles and map it to M tree On the grid surface containing the point, the comprehensive tree model is obtained , where the texture information of point p at viewing angle v is T(v,p), T(v,p)=L(v,p)⋅C(r).

[0019] Compared with the prior art, the advantages of the present invention are: (1) Accurate skeleton extraction and tree modeling: The graph theory method in deep learning is introduced as a key link in the skeleton extraction process. The tree skeleton is accurately extracted from the non-ground point cloud. The layered reconstruction method is used to first model the trunk point cloud using Poisson reconstruction to generate a three-dimensional mesh model M of the trunk. trunk ; Then reconstruct the main branch point cloud P1 using cylindrical fitting to generate the branch cylindrical model M branch ; Then generate the small branch model M based on the small branch point cloud P2 based on fractal theory twig ; The three models are integrated to obtain the tree model M tree This method not only focuses on the accuracy of trunks and branches, but also can accurately process the fractal growth of small branches, ensuring the natural transition and accurate modeling of trunks, branches and small branches, and significantly improving the effect of tree modeling.

[0020] (2) Improving modeling accuracy and texture quality: The present invention introduces Neural Radiance Field (NeRF) technology during texture mapping, which not only improves the texture quality of the tree surface, but also ensures the lighting consistency and natural transition of the texture. Therefore, it effectively improves the details and lighting effects of the tree model, and can provide more realistic data support for forest management.

[0021] (3) Efficient application and ecological assessment: Through refined tree modeling, we can more accurately calculate the leaf area index (LAI) and biomass of trees and estimate the green volume of forest areas. These indicators have important application value in forestry management, ecological protection, and climate change research.

[0022] In summary, the present invention automatically completes the segmented modeling of tree trunks and branches based on the real-world collected laser point cloud and image data, introduces terrain elevation constraints to improve the structure's ground adhesion, and combines multimodal fusion and texture mapping to reconstruct high-precision, natural-feeling three-dimensional tree models, greatly improving the sense of reality and degree of automation. It solves the problems of lack of real-scene accuracy, poor versatility, and low degree of automation in traditional model splicing methods, and provides a new technical path for refined tree modeling and biomass assessment. It is more suitable for tree modeling in forestry digital scenes and complex natural environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0024] The present invention will be further described below with reference to the embodiments and accompanying drawings.

[0025] Example 1: See Figure 1 ,A fractal growth model and refined tree modeling method includes the following steps; S1, obtain the original point cloud, image set, and ground height information h(x, y) of each spatial position in the reconstruction area for a single tree in the reconstruction area. The image set includes the original images from multiple perspectives used for NeRF reconstruction. h(x, y) is the ground height of each spatial position in the area where the tree is located. S2, data preprocessing and point cloud data segmentation; Preprocessing the original image and the original point cloud to obtain a two-dimensional image and a first point cloud; Fitting the ground plane based on the RANSAC algorithm, the first point cloud is segmented into ground point cloud and non-ground point cloud; Based on the RANSAC algorithm, the cylinder is fitted to divide the ground point cloud into the trunk point cloud and the branch point cloud; S3, extracting tree skeletons from non-ground point clouds, including S31~S33; S31, based on the Delaunay triangulation method, construct the adjacency graph G=(V,E) of the non-ground point cloud, where V is the vertex set and E is the edge set. For two vertices p i and p j , if there is an edge e between the two ij , then e ij The edge weight is w ij , , where 、 p i and p j Coordinates in the world coordinate system, ‖∙‖2 is the Euclidean distance; S32, select a point close to the ground as the root node, use Dijkstra algorithm to generate the shortest path tree of the root node, and use the set of nodes in the shortest path tree as the initial skeleton S initial ; S33, generate the initial skeleton S through the cubic spline interpolation algorithm initial The smoothed skeleton curve Q'(t); S4, hierarchical reconstruction of trees, including S41–S44; S41, based on the ground height information constraint, use Poisson reconstruction to model the tree trunk point cloud and generate a 3D mesh model of the tree trunk M trunk ; S42, presetting a density threshold, dividing the points in the branch point cloud that are greater than the density threshold into the main branch point cloud P1, and the rest into the small branch point cloud P2; S43, reconstruct P1 using cylindrical fitting to generate a branch cylindrical model M branch ; S44, simulates the growth of small branches based on fractal theory for P2 and generates a small branch model M twig ; S5, for M trunk 、M branch and M twig Fusion is performed to obtain the tree model M tree ; S6, for tree model M tree Texture mapping is performed on each point to generate a comprehensive tree model .

[0026] Example 2: See Figure 1 , based on Example 1, more specifically: In S2, preprocessing the original image is to convert the original image into the world coordinate system; The original point cloud is pre-processed as follows: the original point cloud is converted into a world coordinate system, the point cloud data is processed by a clustering algorithm, noise points are removed, and the first point cloud is obtained.

[0027] In S33, the initial skeleton S is generated initial The smoothed skeleton curve Q'(t) is specifically: Sb1, from S initial Extract discrete point sets from ; Sb2, use the cubic spline function Q(t) to interpolate and smooth the discrete point set to obtain the skeleton curve, where Q(t)=A⋅[t 3 ,t 2 ,t 1 ,1] T ,t∈[0,1], A is the coefficient matrix, t is the local normalization parameter; Sb3, optimize the coefficient matrix A so that Q(t) approaches S initial The final skeleton curve is marked as the smoothed skeleton curve Q'(t).

[0028] In S4, Poisson's equation for Poisson reconstruction According to the following formula: , Where, is the volume field, ∇ 2 is the Laplace operator, ∇ is the divergence operator, is the normal vector field of the non-ground point cloud, λ is the control coefficient of the terrain fitting strength, and z is the height of each point in Q'(t).

[0029] S43 reconstructs P1 using cylindrical fitting to generate a branch cylindrical model M branch ; Specifically; Sc1, the main branch point cloud P1 is divided into several sub-point clouds based on the clustering algorithm; Sc2, for each sub-point cloud, a cylindrical model is generated using a weighted regularized cylindrical fitting method, where the objective function F of the weighted regularized cylindrical fitting method is obtained according to the following formula; , Where n is the total number of points in the sub-point cloud, w i is the i-th point p in the sub-point cloud i The weight of dist(p i ,Cylinder(θ)) is point p i The distance to the cylinder, γ is the regularization parameter, θ is the cylinder parameter, θ prior is the branch prior parameter, is the square of the L2 norm; Sc3, the collection of cylindrical models of all sub-point clouds constitutes M branch。

[0030] S5 vs. M trunk 、M branch and M twig Carry out integration, specifically: S51, splicing M trunk 、M branch and M twig and use weighted average fusion method to eliminate the seams; S52, using an edge segmentation algorithm to segment the fused edges, thereby achieving mesh encryption at the joints.

[0031] S6, for tree model M tree Texture mapping is performed on each point to generate a comprehensive tree model , specifically; The illumination model of M tree Each point in generates light intensity at different viewing angles, and the light intensity of point p at viewing angle v is L(v,p); Pre-train the NeRF model with the image set to generate M tree The color of each point at different viewing angles, the color of point p at viewing angle v is C(r); Generate texture information of each point at different viewing angles and map it to M tree On the grid surface containing the point, the comprehensive tree model is obtained , where the texture information of point p at viewing angle v is T(v,p), T(v,p)=L(v,p)⋅C(r).

[0032] Example 3: See Figure 1 , based on Example 1, the specific operations of each step are given.

[0033] Regarding step S1: In forestry real scene 3D modeling, point cloud data and image data are the most common basic data sources. Point cloud data is the original point cloud mentioned in the present invention, which can be represented as a 3D point set , where each point p i is a three-dimensional coordinate (x i ,y i ,z i Image data, i.e., the image set described in this invention, is used for NeRF reconstruction. Multiple, multi-view original images are collected using a drone according to the NeRF method. Ground height information h(x, y) represents the ground height at each spatial location and is used to align trees with terrain.

[0034] Regarding step S2: Both the original image and the original point cloud need to be converted to a unified world coordinate system. Taking the original point cloud as an example, first transform each point p i , through the rigid transformation matrix , realize the transformation from the laser scanning coordinate system to the world coordinate system, and obtain p i Corresponding to the point in the world coordinate system , , Including rotation and translation. The original image in the image set needs to be registered with the point cloud data so that the points on the original image match the point cloud at the corresponding position.

[0035] Clustering algorithms: Clustering algorithms are commonly used to remove noise points from point cloud data. Examples include the DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm, the RANSAC (Random Sample Consensus) algorithm, the region growing algorithm, and the K-Means clustering algorithm. Clustering algorithms not only optimize spatial density but also consider the spatial distribution characteristics of the point cloud, such as directionality and shape. Ultimately, they identify and remove noise points from the point cloud data, resulting in the first point cloud.

[0036] Regarding step S3: The goal of this step is to extract the skeleton of the tree from the non-ground point cloud. Since the structure of the tree is hierarchical and branched, it is necessary to extract the trunk and main branches through graph theory and the shortest path algorithm and smooth them. Therefore, the present invention uses the Delaunay triangulation method to construct the adjacency graph of the non-ground point cloud, and uses the Dijkstra algorithm to generate the shortest path tree of the root node to obtain the initial skeleton S initial。

[0037] Considering the spatial structure and morphology of trees and the connection between adjacent nodes, local smoothing is required to remove noise points and enhance the stability of the main skeleton. The smoothing process ensures the continuity of the skeleton by minimizing the distance between each point and its adjacent points. Therefore, the present invention uses the cubic spline interpolation algorithm for optimization. initial Extract discrete point sets from the y-axis, and use the cubic spline function Q(t) to interpolate and smooth the discrete point sets to generate a skeleton curve, and solve the coefficient matrix A in Q(t) to make Q(t) as close to S as possible. initial The midpoint is obtained by removing the irregular discrete error and finally obtaining a continuous and smooth Q'(t).

[0038] Regarding step S41: tree trunk modeling is one of the most critical steps in the terrain fusion process. At this stage, considering the influence of terrain, the present invention improves the Poisson equation and combines the tree trunk model with the terrain, that is, in the original Poisson equation Based on the above, we introduce the λ⋅(zh(x,y)) part. The default value of λ is 0.1 and can be adjusted according to the actual situation. The optimization goal of Poisson reconstruction is to make the trunk surface as smooth as possible while accurately fitting the terrain data. By solving the Poisson equation, we optimize the trunk surface to make it fit the terrain more accurately. After reconstruction, we obtain a model. Then, we use the marching cubes algorithm to extract the isosurface and generate the three-dimensional mesh model M of the trunk. trunk The improved Poisson reconstruction method of the present invention is the same as the Poisson reconstruction method of the prior art except for the difference in the equation.

[0039] Regarding step S42: the purpose is to divide the branch point cloud into two parts by density threshold, so as to be used for subsequent reconstruction.

[0040] Regarding step S43: In the tree branch modeling, a common method is cylindrical fitting. The main branches are modeled by the weighted regularized cylindrical fitting method, while ensuring that the connection between the branches and the trunk and other branches is natural and smooth. Since the tree has many main branches, the main branch point cloud P1 is also divided into multiple sub-point clouds, which are cylindrically fitted separately and then merged to form the branch cylindrical model M. branch When fitting the cylinder, P1 needs to be divided into multiple sub-point clouds through a clustering algorithm. The clustering algorithm is a spatial connectivity-based clustering algorithm such as the DBSCAN method.

[0041] Regarding step S44: Since the growth of small branches presents complex fractal characteristics, the present invention simulates the natural growth of small branches based on fractal theory and uses dynamic allometric rules to describe the growth of small branches during the growth process. In the rule, multiple small branches located at the same bifurcation point are clustered, where the diameter of the i-th small branch is Irregular changes as trees grow: , Among them, α(t) is the growth factor that changes with time t and is used to simulate the morphological changes of trees as they age, r s is the reference initial radius of branch growth, w i is the weight of the i-th small branch, which is used to indicate the relative importance of the small branch in the current growth stage, ∑w j is the sum of the weights of a cluster of small branches. After simulating all small branches, the Delaunay triangulation method is used to generate a triangular mesh model of the branches, and the small branch model M is obtained. twig。

[0042] Regarding step S5: The purpose is to get M trunk 、M branch and M twig Fusion is performed to generate a complete three-dimensional tree model M tree and ensure that the transitions between the parts are natural.

[0043] Regarding step S6: Borrowing the idea of NeRF model, deep neural network is used to learn the illumination and color information in three-dimensional space, thereby generating a texture effect with a strong sense of reality. According to the method of the present invention, the illumination intensity L(v,p) of point p under the viewing angle v is generated by the illumination model, and the color C(r) of point p under the viewing angle v is generated by the NeRF model. The texture information T(v,p) of point p under the viewing angle v is obtained by combining the two, and mapped to M tree Get , which is a complete tree model including color and density information, with realistic texture effects.

[0044] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A fractal growth model and a refined tree modeling method, characterized in that: The following steps are included: S1, obtain the original point cloud, image set, and ground height information h(x, y) of each spatial position in the reconstruction area for a single tree in the reconstruction area. The image set includes the original images from multiple perspectives used for NeRF reconstruction. h(x, y) is the ground height of each spatial position in the area where the tree is located. S2, data preprocessing and point cloud data segmentation; Preprocessing the original image and the original point cloud to obtain a two-dimensional image and a first point cloud; Fitting the ground plane based on the RANSAC algorithm, the first point cloud is segmented into ground point cloud and non-ground point cloud; Based on the RANSAC algorithm, the cylinder is fitted to divide the ground point cloud into the trunk point cloud and the branch point cloud; S3, extracting tree skeletons from non-ground point clouds, including S31~S33; S31, based on the Delaunay triangulation method, construct the adjacency graph G=(V,E) of the non-ground point cloud, where V is the vertex set and E is the edge set. For two vertices p i and p j , if there is an edge e between the two ij , then e ij The edge weight is w ij , , where 、 p i and p j Coordinates in the world coordinate system, ‖∙‖2 is the Euclidean distance; S32, select a point close to the ground as the root node, use Dijkstra algorithm to generate the shortest path tree of the root node, and use the set of nodes in the shortest path tree as the initial skeleton S initial ; S33, generate the initial skeleton S through the cubic spline interpolation algorithm initial The smoothed skeleton curve Q'(t); S4, hierarchical reconstruction of trees, including S41–S44; S41, based on the ground height information constraint, use Poisson reconstruction to model the tree trunk point cloud and generate a 3D mesh model of the tree trunk M trunk ; S42, presetting a density threshold, dividing the points in the branch point cloud that are greater than the density threshold into the main branch point cloud P1, and the rest into the small branch point cloud P2; S43, reconstruct P1 using cylindrical fitting to generate a branch cylindrical model M branch ; S44, simulates the growth of small branches based on fractal theory for P2 and generates a small branch model M twig ; S5, for M trunk 、M branch and M twig Fusion is performed to obtain the tree model M tree ; S6, for tree model M tree Texture mapping is performed on each point to generate a comprehensive tree model .

2. A fractal growth model and refined tree modeling method according to claim 1, characterized in that: In S2, preprocessing the original image is to convert the original image into the world coordinate system; The original point cloud is pre-processed as follows: the original point cloud is converted into a world coordinate system, the point cloud data is processed by a clustering algorithm, noise points are removed, and the first point cloud is obtained.

3. A fractal growth model and refined tree modeling method according to claim 1, characterized in that: In S33, the initial skeleton S is generated initial The smoothed skeleton curve Q'(t) is specifically: Sb1, from S initial Extract discrete point sets from Sb2, use the cubic spline function Q(t) to interpolate and smooth the discrete point set to obtain the skeleton curve, where Q(t)=A⋅[t 3 ,t 2 ,t 1 ,1] T ,t∈[0,1], A is the coefficient matrix, t is the local normalization parameter; Sb3, optimize the coefficient matrix A so that Q(t) approaches S initial The final skeleton curve is marked as the smoothed skeleton curve Q'(t).

4. A fractal growth model and refined tree modeling method according to claim 1, characterized in that: In S4, Poisson's equation for Poisson reconstruction According to the following formula: ， Where, is the volume field, ∇ 2 is the Laplace operator, ∇ is the divergence operator, is the normal vector field of the non-ground point cloud, λ is the control coefficient of the terrain fitting strength, and z is the height of each point in Q'(t).

5. A fractal growth model and refined tree modeling method according to claim 1, characterized in that: S43 reconstructs P1 using cylindrical fitting to generate a branch cylindrical model M branch ; Specifically; Sc1, the main branch point cloud P1 is divided into several sub-point clouds based on the clustering algorithm; Sc2, for each sub-point cloud, a cylindrical model is generated using a weighted regularized cylindrical fitting method, where the objective function F of the weighted regularized cylindrical fitting method is obtained according to the following formula; ， Where n is the total number of points in the sub-point cloud, w i is the i-th point p in the sub-point cloud i The weight of dist(p i ,Cylinder(θ)) is point p i The distance to the cylinder, γ is the regularization parameter, θ is the cylinder parameter, θ prior is the branch prior parameter, is the square of the L2 norm; Sc3, the collection of cylindrical models of all sub-point clouds constitutes M branch .

6. A fractal growth model and refined tree modeling method according to claim 1, characterized in that: S5 vs. M trunk 、M branch and M twig Carry out integration, specifically: S51, splicing M trunk 、M branch and M twig and use weighted average fusion method to eliminate the seams; S52, using an edge segmentation algorithm to segment the fused edges, thereby achieving mesh encryption at the joints.

7. The fractal growth model and refined tree modeling method according to claim 1, characterized in that: S6, for tree model M tree Texture mapping is performed on each point to generate a comprehensive tree model , specifically; The illumination model of M tree Each point in generates light intensity at different viewing angles, and the light intensity of point p at viewing angle v is L(v,p); Pre-train the NeRF model with the image set to generate M tree The color of each point at different viewing angles, the color of point p at viewing angle v is C(r); Generate texture information of each point at different viewing angles and map it to M tree On the grid surface containing the point, the comprehensive tree model is obtained , where the texture information of point p at viewing angle v is T(v,p), T(v,p)=L(v,p)⋅C(r).

Images