A tree individualization modeling method for urban twin scene construction

Abstract

The application provides a tree individualization modeling method for urban twin scene construction, color point cloud and real orthophoto are respectively generated based on tilt model data, ground point cloud and tree point cloud are segmented from the color point cloud, terrain grid data and tree grid data are respectively generated by interpolation, single tree crown vector data is extracted from the real orthophoto based on a network model, the tree point cloud is instantiated, single tree point cloud is extracted, mesh network is performed on the single tree point cloud, a single tree reconstruction model is generated, a single tree fine model object is matched and generated by adopting a three-dimensional shape similarity measurement mode of the single tree individualization reconstruction and the tree three-dimensional precision model, and the model instance is scaled and the scene position is placed in combination with the height and crown width information of the single tree, so that the requirement of efficient individualization modeling and rendering of the tree model in a large-scale urban twin scene is met.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of three-dimensional scene modeling, and more particularly, to a tree individualization modeling method for city twin scene construction. BACKGROUND

[0002] As an important component of city three-dimensional scene, the fine modeling and large-scale efficient rendering of trees are key technologies for the construction of simulation-level city twin scene. Currently, the mainstream technical solutions for city tree modeling include fine individual modeling based on ground laser radar point cloud and random point generation planting modeling based on image vegetation area.

[0003] Laser radar equipment cannot directly collect color and texture information of tree branches and leaves, and can only reconstruct the three-dimensional geometry of trees at present, and cannot extract the texture of branches and leaves. The segmentation of individual tree point cloud also mainly relies on manual work. In addition, the cost of laser radar point cloud data acquisition is higher than that of oblique photography and remote sensing technology, and it is mainly suitable for fine modeling of small-scale trees.

[0004] The random planting modeling based on image vegetation area usually extracts the vector surface of the vegetation coverage area based on DOM, and then randomly plants and places individual tree models in the vegetation coverage area according to a certain grid density. This grid random planting model method has low reconstruction accuracy for the position, height and tree species of trees, and is difficult to meet the requirements of fine restoration of simulation-level city scene construction. SUMMARY

[0005] The present application provides a tree individualization modeling method for city twin scene construction, which can solve the problems of low efficiency of current city tree individualization modeling and slow rendering speed of large-scale scene. First, a tree individualization reconstruction method based on oblique model is realized, which reduces the time-consuming and cost of individual modeling of trees when constructing city scene.

[0006] The present application provides a tree individualization modeling method for city twin scene construction, which includes:

[0007] Step S1, three-dimensional node sampling of the oblique model data of the specified area to generate color point cloud data, and vertical sampling to generate true orthographic image TDOM;

[0008] Step S2, semantic segmentation of the color point cloud data based on a deep learning model to obtain ground point cloud, grass point cloud and tree point cloud;

[0009] Step S3, merging the ground point cloud and the grass point cloud, and generating DEM terrain grid data based on Kriging interpolation;

[0010] Step S4, merging the ground point cloud and the tree point cloud, and generating DTSM tree surface grid data expressing the ground tree relief based on Kriging interpolation;

[0011] Step 5, extracting single tree crown circumscribed rectangle vector data from the true orthographic image TDOM based on a preset network model;

[0012] Step S6, instance segmentation of the tree point cloud in step S2 based on the single tree crown circumscribed rectangle vector data, each segmented point cloud cluster being a single tree point cloud;

[0013] Step S7, Mesh network processing of each single tree point cloud to obtain a single tree simplified single model structure;

[0014] Step S8, similarity matching of the single tree simplified single model structure with each tree fine model in a tree three-dimensional model asset library, and matching the tree fine model with the highest similarity as a single tree model expressing a specified area;

[0015] Step S9, spatial overlay analysis of the single tree crown circumscribed rectangle vector data extracted in step S5, the DEM terrain grid data generated in step S3, and the DTSM tree surface grid data generated in step S4, to calculate the bottom elevation, top elevation and crown width of each single tree, and obtain a single tree bounding box;

[0016] Step S10, scaling and translation of the single tree model matched in step S8 according to the single tree bounding box to the center point position of the single tree, to complete the urban twin scene tree individualization modeling and instantiation rendering.

[0017] The tree individualization modeling method for urban twin scene construction provided by the present application can automatically construct a single tree model, and uses instance scene rendering to reduce the cost of urban scene tree individualization modeling, shorten the modeling cycle, and realize efficient rendering of a large number of urban scene tree models, thereby providing a basis for urban twin scene construction. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 The tree individualization modeling method for urban twin scene construction provided by the present application is shown in the flowchart.

[0019] Figure 2 The single tree height and crown width parameter calculation diagram in the embodiment of the present application is shown. DETAILED DESCRIPTION

[0020] In order to make the purposes, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are some embodiments but not all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application. In addition, the technical features in each embodiment or in a single embodiment provided by the present application can be combined with each other at will to form a feasible technical solution, and such combination is not restricted by the order of steps and / or structure composition mode, but should be based on the fact that a person of ordinary skill in the art can realize it. When the combination of technical solutions appears contradictory or unfeasible, it should be considered that the combination of technical solutions does not exist and is not within the protection scope of the present application.

[0021] In view of the problems of large construction range, high model simulation requirement, difficulty in single tree modeling, long modeling cycle and high cost in the current city twin scene construction, the present application provides a method for automatically modeling single trees and instantiating and rendering high-simulation tree models in a city twin scene based on inclined model data (OSGB). The method of the present application first derives color point cloud, true orthographic image TDOM, terrain grid DEM and other multi-source data from the inclined model data, then extracts single tree structures in the scene based on the multi-source data, uses a shape similarity measurement algorithm based on three-dimensional Fourier shape descriptor to calculate and match the tree precision models in the three-dimensional model asset library to obtain high-simulation tree models, and finally plants and instantiates and renders the tree models in the scene according to the calculated single tree crown width, bottom elevation and top elevation parameters. The following is a detailed introduction to the implementation steps, and the overall process is shown in Figure 1 .

[0022] Figure 1 A tree single modeling method flowchart for city twin scene construction provided by the present application is shown in Figure 1 . The method includes the following steps:

[0023] Step S1, three-dimensional node sampling of the inclined model data of the specified area is performed to generate color point cloud data, and vertical sampling is performed to generate a true orthographic image TDOM.

[0024] It can be understood that the inclined model data (OSGB) is traversed to read the Mesh data vertices and perform grid sampling and texture color extraction to obtain color point cloud data. Vertical projection sampling is performed to generate a 0.1m resolution true orthographic image data (TDOM).

[0025] Step S2, the color point cloud data is subjected to semantic segmentation based on a deep learning model to obtain ground point cloud, grass point cloud and tree point cloud.

[0026] It can be understood that the color point cloud generated in step S1 is subjected to semantic segmentation based on the improved RandLA-Net deep learning model to obtain classified point cloud data of trees, grass, ground and the like. RandLA-Net is a high-efficiency and lightweight deep learning model for point cloud semantic segmentation. In order to further improve the point cloud classification accuracy and robustness, the grid sampling method is adopted to replace the default random sampling method of RandLA-Net to resample the color point cloud, but RandLA-Net is still used as the point cloud semantic segmentation framework. The improved RandLA-Net ensures the accuracy and efficiency of large-scale point cloud semantic segmentation, and improves the adaptability and generalization of point clouds of different densities.

[0027] In step S3, the ground point cloud and the grass point cloud are merged, and DEM terrain raster data is generated based on Kriging interpolation.

[0028] It can be understood that the ground point cloud and the grass point cloud in the point cloud classification result of step S2 are derived, and the DBSCAN clustering algorithm is used for point cloud denoising. After the ground points and the grass points are merged, the discrete XYZ coordinates are subjected to Kriging interpolation to generate continuous DEM terrain raster data. The DEM terrain raster data is used for subsequent step calculation and correction of the bottom elevation of the single tree.

[0029] In step S4, the ground point cloud and the tree point cloud are merged, and DTSM tree surface raster data expressing the relief of the ground trees is generated based on Kriging interpolation.

[0030] It can be understood that the ground point cloud and the tree point cloud in the point cloud classification result of step S2 are derived, and Kriging interpolation is performed based on the XYZ coordinates of the ground and tree points to generate DTSM raster data expressing the relief of the ground trees. The DTSM raster data is used for subsequent calculation and correction of the top elevation of the single tree.

[0031] In step 5, a single tree crown circumscribed rectangle vector data is extracted from the true orthophoto TDOM based on a preset network model.

[0032] It can be understood that a network model is pre-trained, and a single tree crown circumscribed rectangle vector data is extracted from the true orthophoto TDOM based on the trained network model.

[0033] The process of pre-training a network model is: collecting image data of different cities and different regions to make single tree annotation training set, and the annotation training set is the single tree crown circumscribed rectangle in the image range. Considering the recognition accuracy requirement of single tree in the image and the demand for large scene extraction efficiency, YoloV8 is selected as the deep learning network model for image single tree target recognition monitoring. YoloV8 network model inherits the efficient characteristics of Yolo series model, can realize fast detection speed while maintaining high accuracy, benefits from rich data enhancement and regularization strategy, and improved label assignment and loss function, etc. It has strong adaptability to image data under different environments, light, seasons and other conditions, and can stably identify single tree targets. Using dynamic scaling feature pyramid network and other technologies, it can effectively process single tree targets of different scales, whether tall trees or small shrubs, can be well detected and identified. After completing the single tree annotation training set, the YoloV8 network model is used for model training, and the model accuracy is evaluated and verified on the validation set and test set. According to the test results, the learning rate and regularization parameters are adjusted to optimize the model performance, and the final image single tree extraction network model program and weight file are output.

[0034] The true projection image TDOM data obtained in step S1 is subjected to the YoloV8 network model trained in step S5 to automatically extract single trees in the image, and the single tree crown circumscribed rectangle vector data is output. Each rectangle box represents a single tree object in the urban scene.

[0035] Step S6, based on the single tree crown circumscribed rectangle vector data, the tree point cloud in step S2 is instance segmented, and each piece of point cloud cluster after segmentation is a single tree point cloud.

[0036] It can be understood that the tree point cloud obtained in step S2 is subjected to point cloud instance segmentation according to the single tree crown circumscribed rectangle range obtained in step S5, and each piece of point cloud cluster after segmentation is a single tree point cloud.

[0037] Step S7, each single tree point cloud is subjected to Mesh network processing to obtain a simplified single model structure of the single tree.

[0038] It can be understood that Mesh network processing is performed on each single tree point cloud cluster obtained in step S6. Since the amount of Mesh data directly constructed based on the point cloud cluster is large, in order to improve the efficiency of subsequent calculation of the geometric shape similarity of the tree model, the embodiment of the application performs Mesh simplification on each single tree point cloud cluster based on a secondary error grid simplification algorithm based on edge contraction. The core idea of the secondary error (Quadric Error Metrics, QEM) grid simplification algorithm based on edge contraction is to reduce the number of vertices by iteratively contracting the edges in the grid, while minimizing the geometric error introduced in the simplification process. The algorithm defines the error by the sum of the squares of the distances from the vertices to the adjacent surfaces, constructs an error matrix, and then adopts a greedy strategy to contract the edge with the smallest error increment each time, so as to ensure that the geometric shape after simplification is as close as possible to the original grid. At the same time, the priority queue is used to dynamically maintain the contraction order of the edges, so as to quickly select the optimal operation and realize efficient three-dimensional grid simplification. The secondary error grid simplification algorithm based on edge contraction is used to perform Mesh simplification on each single tree point cloud cluster, and a simplified single tree model structure is obtained.

[0039] In step S8, the simplified single tree model structure is matched with each tree model in the tree three-dimensional model asset library, and the tree model with the highest similarity is matched as the single tree model expressing the specified area.

[0040] It can be understood that the tree three-dimensional model asset library is established in advance by a manual method, and the model library includes each tree model of common urban area tree species types such as trees and shrubs.

[0041] Since the geometric data amount of the data fine model is large, the secondary error grid simplification algorithm based on edge contraction is also used to perform Mesh simplification on each tree model, and a simplified tree model structure is obtained, so as to simplify the model geometric data amount as much as possible while keeping the basic form of the tree model unchanged.

[0042] The shape similarity measurement algorithm based on the three-dimensional shape Fourier descriptor is used to measure the shape similarity of the simplified single tree model structure and each tree model in the tree three-dimensional model asset library, and the tree model with the highest similarity is matched as the single tree model expressing the area. The algorithm encodes the three-dimensional body structure into frequency domain coefficients through three-dimensional Fourier transform, extracts the frequency spectrum features as the representation of the shape features, and then compares the frequency spectrum features of the two three-dimensional body shapes to measure the shape similarity. The implementation steps of the shape similarity measurement algorithm based on the three-dimensional Fourier shape descriptor are as follows:

[0043] (1) Three-dimensional shape voxelization and normalization processing:

[0044] First, the voxelization process is performed, and the three-dimensional model of the tree (including the extracted simplified single tree model structure and each tree detailed model in the three-dimensional asset library of the tree) is converted into a binary voxel grid. A voxel value of 1 represents the inside of the body, and 0 represents the outside of the body. If the surface of the three-dimensional model is S , the voxel grid V ( x , y , z ) is defined as:

[0045] ;

[0046] wherein ( x , y , z ) represents the original x, y, and z axis coordinates in the three-dimensional model. Then, the binary voxel grid of the three-dimensional model is subjected to shape normalization processing, the three-dimensional centroid is moved to the coordinate origin, the principal axis is aligned through principal component analysis, the rotation effect is eliminated, and the unit cube with a side length of 1 is scaled to ensure the translational, rotational, and scaling invariance.

[0047] (2) Three-dimensional discrete Fourier transform:

[0048] The normalized voxel grid is subjected to three-dimensional Fourier transform to obtain the frequency domain coefficients:

[0049] ;

[0050] wherein kx,ky,kz is the frequency domain coordinate, N is the size of the voxel grid.

[0051] (3) Fourier shape descriptor extraction:

[0052] The amplitude of the frequency spectrum is retained while the phase is ignored, the low-frequency part is intercepted to retain the main shape features, and the three-dimensional amplitude spectrum is flattened into a one-dimensional vector V as the three-dimensional shape descriptor, i.e., the frequency spectrum feature.

[0053] (4) Similarity measurement calculation:

[0054] The Euclidean distance d between two three-dimensional shape descriptor vectors V1 and V2 is calculated:

[0055] ;

[0056] wherein is the frequency spectrum feature of the simplified single tree model structure, is the frequency spectrum feature of the tree detailed model, i is the i-th element in the frequency spectrum feature, and the Euclidean distance dThe closer the value is to 0, the higher the similarity of the two three-dimensional body shapes is. Through the shape similarity measurement algorithm based on the three-dimensional Fourier shape descriptor, a geometric shape closest to a tree model in a three-dimensional tree model asset library can be matched, thereby providing a basic condition for subsequent restoration, placement and rendering of a high-simulation tree model in a city twin scene.

[0057] In step S9, the single tree crown circumscribed rectangle vector data extracted in step S5, the DEM terrain raster data generated in step S3 and the DTSM tree surface raster data generated in step S4 are subjected to spatial overlay analysis, and the bottom elevation, top elevation and crown width of each single tree are calculated to obtain a single tree bounding box.

[0058] It can be understood that the circumscribed rectangle of the single tree crown obtained in step S5 is subjected to spatial overlay analysis with the DEM terrain raster and the DTSM tree surface raster obtained in steps S3 and S4 respectively, the bottom elevation, top elevation and crown width of each single tree are calculated, and the single tree bounding box is obtained, as shown in FIG. 6. Figure 2 The single tree bounding box represents the spatial position of the single tree model in the three-dimensional scene, and provides a prerequisite for model planting, placement and three-dimensional spatial transformation in the next step.

[0059] In step S10, the single tree model matched in step S8 is scaled according to the single tree bounding box and is translated to the center point position of the single tree, so that the single tree modeling and instance rendering of the city twin scene are completed.

[0060] It can be understood that the fine tree model matched in step S9 is scaled according to the single tree bounding box geometry obtained in step S10, and is translated to the center point position of the single tree in the city twin scene, so that the single tree modeling and efficient instance rendering of the city twin scene are finally completed. In this way, the position, size and tree shape characteristics of the single tree in the city scene are accurately restored, and the model rendering pressure is significantly reduced by using the three-dimensional model instance rendering, thereby creating favorable conditions for the construction of a large-scale city twin scene.

[0061] The present application provides a single tree modeling method for city twin scene construction, which has the following advantages:

[0062] 1. The single tree reconstruction is automatically performed based on the inclined model data, without human intervention, so that the time consumption of single tree modeling is greatly reduced, and the single tree reconstruction efficiency of the entire city scene is improved.

[0063] 2. The inclined model data which is more easily obtained and widely used is used for tree reconstruction, and the requirement for modeling data input is lower.

[0064] 3. First, a single tree model is built based on the tilted 3D model and its derived point cloud and image data. Then, a shape similarity measurement algorithm based on the 3D Fourier shape descriptor is used to perform shape matching retrieval of simulation-level tree models. The individual tree modeling is completed by planting them in 3D space. Only limited manual modeling of high-precision tree models is required to complete the simulation-level rendering modeling of trees in super-large urban scenes.

[0065] 4. Rendering is performed by instantiating single-building models, which significantly reduces video memory usage and improves rendering efficiency, meeting the requirements for tree simulation-level rendering in large-scale urban scenes.

[0066] 5. Tree individualization is based on the currently more mainstream inclined 3D model and its derived 3D color point cloud, TDOM true orthogonal image and DSM digital surface model and other multi-source data. The data is widely available and the comprehensive use of multi-source data also improves the robustness and generalization of the individual tree modeling algorithm.

[0067] It should be noted that, in the above embodiments, the description of each embodiment has its own focus. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant description of other embodiments.

[0068] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0069] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded computer, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0070] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0071] The computer program instructions can also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions that are executed on the computer or other programmable apparatus provide steps for implementing the Figure 1 function specified in the flow or flows and / or blocks Figure 1 of the block or blocks.

[0072] While the preferred embodiments of the application have been described, additional variations and modifications can be made to the embodiments by those skilled in the art once they learn of the basic inventive concepts. Therefore, the appended claims are intended to cover all such additional variations and modifications as fall within the scope of the application.

[0073] Obviously, numerous modifications and variations of the present application are possible in light of the above teachings. It is therefore to be understood that within the scope of the appended claims and their equivalents, the application can be practiced otherwise than as specifically described.

Claims

1. A tree individualization modeling method for urban twin scenario construction, characterized in that, The method comprises the following steps: Step S1, three-dimensional node sampling of the tilt model data of the specified area to generate color point cloud data, and vertical sampling to generate a true orthophoto map TDOM; Step S2, semantic segmentation of the color point cloud data based on a deep learning model to obtain ground point cloud, grass point cloud and tree point cloud; Step S3, merging the ground point cloud and the grass point cloud, and generating DEM terrain grid data based on Kriging interpolation; Step S4, merging the ground point cloud and the tree point cloud, and generating DTSM tree surface grid data expressing the relief of the ground trees based on Kriging interpolation; Step 5, extracting single tree crown circumscribed rectangle vector data from the true orthophoto map TDOM based on a preset network model; Step S6, instance segmentation of the tree point cloud in step S2 based on the single tree crown circumscribed rectangle vector data, each piece of point cloud cluster after segmentation being a single tree point cloud; Step S7, Mesh network processing of each single tree point cloud to obtain a single tree simplified single model structure; Step S8, similarity matching of the single tree simplified single model structure with each tree detailed model in a tree three-dimensional model asset library, and matching the tree detailed model with the highest similarity as a single tree model expressing the specified area; Step S9, spatial overlay analysis of the single tree crown circumscribed rectangle vector data extracted in step S5, the DEM terrain grid data generated in step S3 and the DTSM tree surface grid data generated in step S4 to calculate the bottom elevation, top elevation and crown width of each single tree, and obtain a single tree bounding box; Step S10, scaling and translation of the single tree model matched in step S8 according to the single tree bounding box to the center point position of the single tree, to complete the single tree modeling and instantiation rendering of the city twin scene.

2. The tree individualization modeling method of claim 1, wherein, The step S3 of merging the ground point cloud and the grass point cloud and generating DEM terrain grid data based on Kriging interpolation comprises: Respectively denoising the ground point cloud and the grass point cloud based on a DBSCAN clustering algorithm; Merging the denoised ground point cloud and the grass point cloud, and generating continuous DEM terrain grid data by Kriging interpolation on discrete XYZ coordinates.

3. The tree individualization modeling method of claim 1, wherein, The preset network model in step S5 is a Yolov8 network model, and the training of the preset network model comprises: Collecting image data of multiple different cities and multiple different areas, labeling each single tree crown circumscribed rectangle in the image data to form a training set; Training the Yolov8 network model based on the training set.

4. The tree individualization modeling method of claim 1, wherein, The step S7 of Mesh network processing of each single tree point cloud to obtain a single tree simplified single model structure comprises: Mesh network processing of each single tree point cloud based on a secondary error grid simplification algorithm based on edge contraction.

5. The tree individualization modeling method of claim 1, wherein, The step S8, the monomer tree simplification monomer model structure is matched with each tree model in the tree three-dimensional model asset library Similarity, the highest similarity tree model is matched as the monomer tree model of the specified area, comprising: For each tree model in the tree three-dimensional model asset library, based on the secondary error grid simplification algorithm of edge contraction Mesh simplification, obtain the simplified tree model structure; Based on the shape similarity measurement algorithm of three-dimensional shape Fourier descriptor, the shape similarity of the monomer tree simplification monomer model structure and the simplified tree model structure corresponding to each tree model in the tree three-dimensional model asset library is measured, and the highest similarity tree model is matched as the monomer tree model of the specified area.

6. The tree individualization modeling method of claim 5, wherein, The shape similarity measurement algorithm based on three-dimensional shape Fourier descriptor, the shape similarity of the monomer tree simplification monomer model structure and the simplified tree model structure corresponding to each tree model in the tree three-dimensional model asset library is measured, comprising: Respectively, the monomer tree simplification monomer model structure and the simplified tree model structure corresponding to each tree model in the tree three-dimensional model asset library is three-dimensional Fourier transform coding, obtain the frequency domain coefficient, and extract the frequency spectrum feature; The frequency spectrum feature of the monomer tree simplification monomer model structure and the frequency spectrum feature of each tree model are measured.

7. The tree individualization modeling method of claim 6, wherein, The monomer tree simplification monomer model structure and the simplified tree model structure corresponding to each tree model in the tree three-dimensional model asset library is three-dimensional Fourier transform coding, obtain the frequency domain coefficient, and extract the frequency spectrum feature, comprising: convert the simplified tree model structure of each of the monomer tree and the simplified tree model structure corresponding to each of the tree fine model in the tree three-dimensional model asset library into a binary voxel grid, wherein a voxel value of the voxel grid is 1, indicating that the voxel grid is in the body, and a voxel value of the voxel grid is 0, indicating that the voxel grid is outside the body; wherein if the three-dimensional model surface is S , the voxel grid V ( x , y , z ) is defined as: ； wherein, x , y , z ) are the x, y, z coordinates of the point in the three-dimensional model, respectively; The voxel grid is normalized: the three-dimensional body center is moved to the coordinate origin, the principal axis is aligned by principal component analysis, and the unit cube with side length of 1 is scaled; The normalized voxel grid is three-dimensional Fourier transform, and the frequency domain coefficient is obtained: ； wherein, kx,ky,kz is the frequency domain coordinate, N is the voxel grid size; The amplitude of the frequency spectrum is retained and the phase is ignored, the low frequency part of the frequency domain coefficient is intercepted, and the three-dimensional amplitude spectrum is flattened into a one-dimensional vector V as a three-dimensional shape descriptor, as a frequency spectrum feature.

8. The tree individualization modeling method of claim 6, wherein, The frequency spectrum feature of the monomer tree simplification monomer model structure and the frequency spectrum feature of each tree model are measured. ； wherein, is the spectral feature of the simplified individual tree model structure for the individual tree, is the spectral feature of the refined tree model, i is the i-th element in the spectral feature, and d is the Euclidean distance.

Images