Tree structure parameter extraction and three-dimensional reconstruction method based on ground laser point cloud data

Abstract

The invention discloses a tree structure parameter extraction and three-dimensional reconstruction method based on ground laser point cloud data, belongs to the technical field of tree three-dimensional modeling computer programs based on laser point cloud data in forestry scientific research, takes tree point cloud as input, realizes prediction of skeleton points by iterating a sparse convolution-based residual U-shaped network (SpConv-based ResUNet) for multiple times, and provides a tree structure parameter extraction and three-dimensional reconstruction method based on ground laser point cloud data. An initial skeleton is constructed through graph analysis, connection point search of automatic driving path smoothing and angle constraint is used for skeleton smoothing, an arterial snake is used for constructing a branch model based on the radius estimated by a constraint B spline curve and the skeleton, and a simple phyllotaxis algorithm is used for realizing insertion of virtual leaves so as to realize tree simulation. And automatically extracting parameters by using the input point cloud, the extracted skeleton and the reconstructed model.

Description

Technical Field

[0001] The invention belongs to the technical field of computer programs for three-dimensional modeling of trees based on laser point cloud data in forestry scientific research, and relates to a method for extracting tree structural parameters and three-dimensional reconstruction based on ground laser point cloud data. Background Art

[0002] Tree structure is influenced by genetic, biotic, and abiotic factors, resulting in specific physiological and ecological characteristics that underlie analyses of growth dynamics, developmental patterns, reproduction, and environmental responses. However, its complexity makes modeling and quantification challenging. With the increasing application of laser scanning technology, which can generate high-throughput three-dimensional surface information, in forestry, the expression of tree structure has evolved into quantitative structural modeling methods that utilize simple geometric primitives for visualization and parametric calculations. This approach has become the mainstream method for 3D tree modeling based on laser point clouds and has been applied to research in forest ecology and tree breeding.

[0003] Due to the inevitable noise and occlusion issues inherent in point cloud data, various current quantitative structural models suffer from various problems when processing single-tree point clouds from real-world laser scans: low reconstruction quality, inefficient algorithms, weak parameter extraction capabilities, and complex hyperparameter settings. Furthermore, these methods are not interactive, limiting the analysis of branch properties to statistical analysis at the global level. Summary of the Invention

[0004] In response to the existing technical problems, the present invention provides a method for extracting tree structure parameters and reconstructing three-dimensionally based on ground laser point cloud data.

[0005] A method for extracting tree structural parameters and reconstructing three-dimensional (3D) reconstruction from terrestrial laser point cloud data comprises the following steps: using a tree point cloud as input, predicting skeleton points through multiple iterations of a sparse convolution-based residual U-net (SpConv-based ResUNet), constructing an initial skeleton through graph analysis, applying path smoothing and angle-constrained connection point search for autonomous driving to skeleton refinement, constructing a branch model using an arterial snake based on the radius and skeleton estimated by a constrained B-spline curve, and inserting virtual leaves using a simple phyllotaxy algorithm to simulate the tree. Parameters are automatically extracted using the input point cloud, the extracted skeleton, and the reconstructed model.

[0006] The residual U-shaped network of sparse convolution is embedded in the recurrent structure to shrink the point cloud and obtain the skeleton points of the tree.

[0007] Graph analysis is to find the shortest path tree on the connected graph obtained by De Launay tetrahedronization of the skeleton points, and perform path selection based on the tree graph.

[0008] Path smoothing is a three-dimensional implementation of path smoothing in Baidu's autonomous driving platform Apollo, with angle constraints within a loose branching angle range of 15° to 165°.

[0009] The radius estimation uses a variety of constraint curves (power, exponential, quadratic, Bailey, and B-spline) with zero-crossing points and monotonically increasing value ranges to output stem shape analysis results, and only B-spline is used for modeling.

[0010] The model of the arterial snake establishes a mapping relationship between cross-section vertices to facilitate volume and surface area calculations.

[0011] The simple leaf arrangement algorithm can realize basic simulation of opposite, alternate, whorled and clustered leaves according to parameterized leaf size and arrangement.

[0012] The advantages of the present invention are: a method for extracting tree structure parameters and three-dimensional reconstruction of ground laser point cloud data, which can achieve more accurate, rapid and stable parameter extraction and quantitative structure modeling, can meet the needs of forestry research in various scenarios, and efficiently provide materials for fields that require tree models such as digital twins.

[0013] The sparse convolutional neural network is introduced into point cloud shrinkage and integrates methods that are beneficial to the modeling process. It can efficiently process single tree ground laser data in batches; it supports the extraction and visual analysis of 29 individual-level, 20 organ-level, and 9 plot-level parameters, and supports the analysis of stem shape and horizontal spatial distribution patterns; it achieves high-precision, efficient, and robust parameter extraction and modeling. In the experimental data, the extracted stem length, stem volume, and branch height R 2 The scores reached 0.999, 0.99, and 0.9, and the branch length, branch angle, and branch angle R 2 The scores were all around 0.95, effectively alleviating the overestimation of biomass. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work. As shown in the figure:

[0015] Figure 1 It is a technical flow chart of the present invention.

[0016] Figure 2 This is the distribution map of marker trees in the sample plot of the present invention.

[0017] Figure 3 This is the laser data processing flow of the present invention.

[0018] Figure 4 This is the single wood reconstruction effect of the present invention.

[0019] Figure 5 This is the sample site reconstruction effect of the present invention. DETAILED DESCRIPTION

[0020] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0021] Example 1: Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 and Figure 5 As shown in the figure, a tree structure parameter extraction and 3D reconstruction method based on ground laser point cloud data mainly involves single tree parameter extraction, tree quantitative structure modeling and forest visualization simulation technology in computer science.

[0022] A method for extracting tree structural parameters and reconstructing three-dimensional (3D) tree structures from terrestrial laser point cloud data comprises the following steps: using a tree point cloud as input, predicting skeleton points through multiple iterations of a sparse convolution-based residual U-net (SpConv-based ResUNet), constructing an initial skeleton through graph analysis, applying path smoothing and angle-constrained connection point search for autonomous driving to skeleton smoothing, constructing a branch model using an arterial snake based on the radius and skeleton estimated by a constrained B-spline curve, and inserting virtual leaves using a simple phyllotaxy algorithm to simulate the tree. Utilizing the input point cloud, the extracted skeleton, and the reconstructed model, 58 parameters (29 individual tree measurement factors, 20 branch attributes, and 9 horizontal spatial structure parameters) are automatically extracted.

[0023] A method for extracting tree structural parameters and reconstructing three-dimensional (3D) forests from terrestrial laser point cloud data addresses the difficulties faced by current quantitative structural modeling methods in achieving comprehensive coordination of reconstruction quality, algorithm efficiency, parameter extraction comprehensiveness and accuracy, and interactivity and functionality. This method provides a robust and easy-to-use solution for tree structural analysis and rapid reconstruction of forest scenes using point cloud data, offering technical support for precision forestry and related fields requiring realistic tree models.

[0024] It mainly includes two aspects: point cloud contraction achieved by repeated iteration of residual U-shaped network based on sparse convolution and a novel reconstruction process for efficiently generating three-dimensional models and parameters of the above-ground tree structure. The overall technical process is as follows: Figure 1 shown.

[0025] Depend on Figure 1 , including a training process and an inference process. The training process uses a synthetic tree dataset to train a residual U-shaped network based on sparse convolution. The inference process uses the trained network to predict the centripetal displacement of the point cloud, and shrinks the point cloud through multiple cycles to obtain skeleton points. The obtained skeleton points are subjected to the standard quantitative structural modeling process of initial skeleton construction-skeleton refinement-radius estimation-Mesh reconstruction to obtain the skeleton and Mesh model of the tree, and also include the results of trunk shape analysis. This process integrates different methods to improve quality. Different levels of parameter extraction are achieved using the input point cloud, skeleton and mesh, and visual analysis is achieved through the secondary development plug-in of CloudCompare.

[0026] like Figure 1 As shown, the input of the entire program can be a single tree point cloud file or additional information in the form of a .yaml data set, which includes leaf images, size, angle, distribution pattern, bark texture, tree species information, and stem curve type (quadratic / exponential / Behre / B-spline). This input information will be fed into the inference process.

[0027] The inference process first divides the input point cloud into blocks (each block includes a buffer zone), batch voxelizes it, and then passes it to the trained neural network model. The model uses the voxelized voxels with attributes as input and predicts the movement direction and distance through a residual U-shaped network based on sparse convolution. Based on the model results, the predicted centripetal displacement corresponding to the point cloud can be calculated. The buffer zone is removed by masking to obtain the shrunken point cloud. This process is repeated until the last iteration ends. The shrunken point cloud is voxel-downsampled to obtain skeleton points, and the initial radius of each skeleton point is calculated. These are then passed to the subsequent quantitative structure modeling process.

[0028] The neural network model used in inference is trained on input data randomly cube-cropped and batch-voxelized on the synthetic tree dataset. The model uses the voxelized point cloud portion of the input data to predict movement direction and distance. The ground-truth centripetal displacement in the input data can be decomposed into the ground-truth movement direction and distance, which are then used to calculate the loss and update the weights.

[0029] The quantitative structure modeling process includes: initial skeleton construction, skeleton refinement, radius estimation and mesh reconstruction.

[0030] The initial skeleton construction includes: graph construction (using Delaunay tetrahedronization), shortest path tree generation (using edge weight graph solution of squared distance) and path selection (using sphere allocation).

[0031] Skeleton refinement includes: skeleton smoothing (using the 3D version of autonomous driving trajectory smoothing) and breakpoint connection (using angle constraint connection point search). This step outputs the tree skeleton.

[0032] Radius estimation includes: initial radius filtering (using filters used for rTwig), B-splines for modeling, and various curve fitting for stem shape analysis.

[0033] Mesh reconstruction includes: branches are represented by arterial snakes, and leaves are represented by a simple leaf order algorithm. The output of this step includes a layered model and a texture model.

[0034] Multi-scale parameters can be extracted based on point cloud, skeleton and mesh. Including:

[0035] Single tree measurement factors (individual-level parameters): diameter at breast height, tree height, timber volume, crown width and height below branches, totaling 29.

[0036] Branch attributes (organ-level parameters): branch length, branching angle, branch attachment angle, bow height and chord length, totaling 20.

[0037] Horizontal spatial structure parameters (plot-level parameters): a total of 9 parameters, including crowding, intermixing, size ratio, angular scale and competition index.

[0038] The visual interaction of these parameters is realized through the ccEntQueri er plug-in developed by CloudCompare to realize the request and response to the database where the parameters are saved.

[0039] The specific steps are as follows:

[0040] enter

[0041] It supports batch processing of single-tree point cloud files and setting different parameters for each tree in the form of a .yaml file. The configuration file needs to have a key name for each tree, and the value is a dictionary configuration, as shown in Table 1.

[0042] Table 1 Custom configuration of each tree

[0043]

[0044]

[0045] 1. Point cloud shrinkage based on sparse convolution

[0046] (1) Structure of sparse convolutional neural network

[0047] The neural network takes the sparse convolution tensor (SparseConvTensor) generated by batch voxelized point cloud as input, and can use the voxel matrix V with N rows (i.e. N voxels) and 7 columns containing attributes N×7=[b,X N×3 , (IIP) N×3 ], where b is the batch vector, X is the voxel coordinate matrix, and IIP is the matrix resulting from voxel downsampling (equivalent to multiplying the matrix π on the left) of a point cloud (matrix) P containing at least N points. Vectors are column vectors by default, with the subscript t used to denote their tth element (i.e., tth row). Matrices are represented by the subscript (t) to denote the vector consisting of their tth row.

[0048] Each row of the matrix

[0049] V (t) =[b t , X (t) , (ΠP) (t) ]=[b t ,i t ,j t , k t , x t ,y t , z t ], t=1, 2, ..., N,

[0050] Here b t is the batch number of the t-th voxel, i t ,j t , k t is the voxel coordinate or voxel index, x t ,y t , z t is the coordinate of the only sampling point within the voxel.

[0051] The voxel downsampling operation is based on the index of the point in the point cloud, and only the first point within a voxel will be retained.

[0052] The neural network is based on a residual UNet implementation using sparse convolution. A convolutional (Conv) layer containing a sparse submanifold convolution (SubMConv3d) maps the three-channel feature map (i.e., the three-dimensional coordinates of the sampling points) into the feature map required for input to the backbone encoder-decoder architecture. The encoder block uses another sparse convolution (SparseConv3d) for downsampling. It differs from SubMConv3d in that only the kernel covering the input points is required to calculate the output points. The upsampling decoder block uses SparseInverseConv3d to ensure that the output results have the same index as the corresponding downsampled input. Residual connections are added to both the downsampling and upsampling steps. Finally, the output channels are integrated. Batch normalization (BatchNorm1d) is used before each layer and block, followed by a ReLU activation function.

[0053] The output of the neural network is the predicted unit movement direction matrix and logarithmic movement distance vector, respectively represented by To express.

[0054] in, Represents dimensional real space.

[0055] With the help of each row of them, the centripetal displacement matrix can be calculated, which is recorded as Each row of is the predicted centripetal displacement of the sampling point in the corresponding row of v, defined as

[0056] For row t=1, 2, ..., N

[0057] in This is the movement direction and logarithmic movement distance corresponding to the tth row of matrix V. It describes the motion of the point toward the ideal central axis. Before obtaining the unit movement direction matrix, matrix row normalization is required. The reason for using the logarithmic movement distance vector instead of the movement distance vector directly is that the output is signed and differs from the true value by multiple orders of magnitude. Therefore, the logarithmic transformation is used to eliminate these variations.

[0058] (2) Loss Function

[0059] During training, the overall loss function used is a combination of the cosine similarity loss for unit direction and the L1 loss for logarithmic distance, defined as:

[0060]

[0061] Among them, the predicted moving direction and logarithmic moving distance The corresponding truth value is denoted as D (t) 、(ln d) t , which come from the synthetic tree dataset. This dataset contains the virtual generated single wood point cloud for training and the corresponding true centripetal displacement of each point. The required true logarithmic displacement can be obtained by inverse deduction from the previous formula

[0062] lnd=ln(||L (1) ||2,||L (2) ||2, ..., ||L (N) ||2) T ,

[0063] And the direction of true value movement

[0064]

[0065] Where L is the true centripetal displacement of the batch sampling point ΠP in each round of training or evaluation; ⊙ represents the Hadamard product or element-by-element product of the matrix; Represents the Kronecker product or tensor product of a vector and a vector.

[0066] (3) Iterative Point Cloud Contraction of Network Embedding

[0067] Point cloud shrinkage based on sparse convolution can be expressed as:

[0068]

[0069] Among them, P (n-1) 、P (n) Respectively represent the point clouds formed by the block batch points in the previous round (n-1) and the current round (n round); M (n) is the mask operation of the current round, used to eliminate all the results in the supplemented buffer in the previous round; (n) Represents the sampling operation of the current round, including re-segmentation of the point cloud, block-by-block buffer increase, and voxel downsampling; They are respectively the moving direction and logarithmic moving distance predicted by the neural network for the current round.

[0070] We use a voxel size and cube size twice that used during training, 0.02m and 8m respectively. This is because the trees in the synthetic tree dataset are relatively short and simple in structure. If they are scaled to the original size, the shrinking effect will be extremely insignificant. However, a larger size means a loss of detail, which is detrimental to the skeleton of delicate tree branches. For very small trees, the current size does not meet the runnable conditions, so the program will try to change these sizes by continuously dividing by 2 to ensure execution.

[0071] The maximum number of iterations is set to 10, and the point cloud is terminated early if the number of points after a round of shrinkage is less than 1000 to prevent loss of branch structure details. After shrinkage, the points are further downsampled, and the points at this point are identified as skeleton points s. Each skeleton point is obtained by moving a single tree surface point, and the distance between the two is considered to be the initial radius of the circular cross-section at that skeleton point.

[0072] 2. Quantitative Structural Modeling

[0073] (1) Initial skeleton construction

[0074] Delaunay tetrahedronization enables the graph constructed from skeleton points to contain only one connected component. On this graph, the Dijkstra algorithm is used with the lowest skeleton point as the origin to find the shortest path to other points. To increase the preference for more adjacent reachable skeleton points, the edge weight uses the squared distance. The shortest path tree at this point can be considered a skeleton, but it contains a large number of short branches. These short branches are usually inside the branches and should be eliminated, so path selection is performed. Before path selection, the skeleton points are divided into different branches in an orderly manner according to the length of the branches supported by each skeleton point. With the help of the initial radius, a series of spheres with the skeleton points on the branches as the sphere center are constructed according to the priority of the branch level. Skeleton points within these spheres that do not belong to the current branch and have not been discarded previously should be discarded, and the topology of the skeleton points of each branch should be reconstructed.

[0075] (2) Skeleton refinement

[0076] Skeleton smoothing and trajectory smoothing are essentially similar; both aim to reduce or eliminate collisions within a path constructed from discrete sampling points. The fem position deviation smoother currently used in Baidu's Apollo autonomous driving platform simultaneously considers smoothness, length, and offset from the original point, resulting in fast and effective solutions. Therefore, it has been modified to smooth the topology of each branch individually in three dimensions. In this case, disconnected skeletons are caused by branches not being connected to their parent branches, requiring the search for a connection point where the starting point of each branch is on the parent branch. Compared to the computational complexity and significant topological modifications associated with finding the optimal connection point, directly finding a point on the parent branch is less expensive. The desired connection point can be found by constructing a right triangle from the nearest neighboring points on the parent branch to the starting point. The parent branch point closest to this point is used as the connection point. However, sometimes the connection point found is too far away, so a constraint on the branch angle is imposed (must be within 15° to 165°). For any failure to meet this constraint, the parent branch point that minimizes the change in the branch's growth direction within this angle range is used as the connection point.

[0077] (3) Radius estimation

[0078] A radius range filter was used to remove noisy data points. Several commonly used trunk curves were then used to fit the functional relationship between the trunk length from each node to the tip and the initial radius. These filters eliminated noise with large oscillations by constraining the trunk length from each node to the tip to various proportional relationships with the initial radius. Experiments were conducted using power, exponential, quadratic, Bailey, and B-spline functions to fit the trunk contour. These curves were constrained to monotonically increase within their range and pass through the origin to prevent overfitting. The results of each curve fit were recorded in a file, and the B-spline curve was used exclusively for subsequent model development, as it was found to be particularly immune to overfitting.

[0079] (4) 3D modeling of trees

[0080] A 3D tree model was constructed using the refined skeleton and estimated radius. The branch structure was represented using an arterial snake model. This representation enforces point mapping relationships between cross sections, facilitating efficient volume and area calculations and texture UV mapping. Leaves generated using a parameterized simple phyllotaxy rule (see Table 2) were combined with the bark texture to enhance the visualization of the 3D model.

[0081] Table 2 Simple phyllotaxy

[0082]

[0083] 3. Multi-level parameter extraction

[0084] The multi-level parameters supported for extraction are shown in Table 3.

[0085] Table 3 Extractable parameters

[0086]

[0087]

[0088]

[0089]

[0090] Example 2: Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 and Figure 5 As shown in the figure, a tree structure parameter extraction and 3D reconstruction method based on terrestrial laser point cloud data is proposed. The specific implementation case is as follows:

[0091] An example of 3D reconstruction and parameter extraction of a typical temperate natural secondary forest plot observed long-term at the Qingyuan Forest Ecosystem Observation and Research Station

[0092] 1. Data Collection and Preprocessing

[0093] A typical temperate natural secondary forest plot with 22 tree species, including conifers, broadleaf trees, vines, shrubs, and small trees, was selected at the Qingyuan Forest Ecosystem Observation and Research Station for long-term observation. The plot is 30 meters by 30 meters in size and was scanned by ground laser scanning in late autumn (end of October) in 2020 and 2024 after the leaves fell. The scanning was carried out using a multi-station ground-based scanner and a backpack mobile laser scanner. Figure 2 shown.

[0094] The two phases of ground laser point cloud data were obtained through Figure 3 The processing flow shown obtains single tree point cloud data.

[0095] like Figure 3 As shown, the 2020 sample site TLS data was obtained by scanning with the RlEGL VZ400i, and the project results were stitched together at multiple stations using the accompanying RiSCANPro (version 2.10) software to obtain the entire sample site point cloud. The scanning device used for the 2024 sample site MLS data was the LiGrip H120, and the project was SLAM solved using the accompanying LiDAR360MLS (version 8.0) software. The solution results were divided into multiple files, and the files were merged in LiDAR360 (version 8.0) to form a complete sample site point cloud. After obtaining the sample site point cloud, a standardized processing process was performed in LiDAR360:

[0096] (1) Preprocessing: This includes cropping the required area, coordinate system matching, downsampling, denoising, ground point classification, normalization, CSF ground point filtering, and ground point / non-ground point segmentation. Cropping the required area refers to manually deleting points outside the sample plot, and expanding the actual sample plot range to a certain extent to prevent incomplete trees at the edge of the sample plot; coordinate system matching refers to manually registering point clouds from different sources through data registration; and ground point / non-ground point segmentation is achieved using category extraction. The output is a normalized ground point cloud and a normalized non-ground point cloud.

[0097] (2) Normalized ground point processing: including denormalization. The output is a ground point cloud.

[0098] (3) Normalized non-ground point processing: This includes denoising, tree segmentation, denormalization, extracting point clouds into multiple files by tree ID, and manual proofreading and interpretation. Tree segmentation uses point cloud segmentation; manual proofreading and interpretation are used to assign or match sample tree numbers to the coarse tree segmentation results split into individual files and to correct misclassified or missed points. The output is a point cloud of all individual trees in the plot.

[0099] 2. Reconstruction quality

[0100] Figure 4 、 Figure 5 Visual display is provided from the perspectives of single tree reconstruction effect and sample plot reconstruction effect.

[0101] Figure 4 It shows the global and local hierarchical model reconstruction effects of a Japanese larch from two perspectives; the global and local reconstruction effects of multiple tree species; and the two-phase mapping model of one tree from each of the 22 tree species.

[0102] Figure 5 The hierarchical models of all trees in the sample plot are displayed simultaneously, and the digital twin scene is constructed.

[0103] 3. Parameter extraction

[0104] Table 4 shows several important single tree tree measurement factors. Figure 4 The overall performance of the 22 trees in the two phases of data.

[0105] Table 4 Several single tree measurement factors

[0106]

[0107]

[0108]

[0109] Table 5 gives the results of several important branch attributes.

[0110] Table 5 Several branch attributes

[0111]

[0112] 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 method for extracting tree structure parameters and 3D reconstruction from terrestrial laser point cloud data, characterized in that: The method includes the following steps: taking tree point cloud as input, predicting skeleton points through multiple iterations of sparse convolution-based residual U-type network (SpConv-basedResUNet), constructing the initial skeleton through graph analysis, using the path smoothing of autonomous driving and the connection point search of angle constraints for skeleton refinement, constructing the branch model using arterial snake based on the radius and skeleton estimated by constrained B-spline curve, and using a simple leaf order algorithm to insert virtual leaves to realize tree simulation, and realizing automatic parameter extraction using the input point cloud, extracted skeleton and reconstructed model.

2. The method for extracting tree structure parameters and 3D reconstruction from terrestrial laser point cloud data according to claim 1, characterized in that: The residual U-shaped network of sparse convolution is embedded in the recurrent structure to shrink the point cloud and obtain the skeleton points of the tree.

3. The method for extracting tree structure parameters and 3D reconstruction from terrestrial laser point cloud data according to claim 1, characterized in that: Graph analysis is to find the shortest path tree on the connected graph obtained by Delaunay tetrahedronization of the skeleton points, and perform path selection based on the tree graph.

4. The method for extracting tree structure parameters and 3D reconstruction from terrestrial laser point cloud data according to claim 1, characterized in that: Path smoothing is a three-dimensional implementation of path smoothing in Baidu's autonomous driving platform Apollo, with angle constraints within a loose branching angle range of 15° to 165°.

5. The method for extracting tree structure parameters and 3D reconstruction from terrestrial laser point cloud data according to claim 1, characterized in that: The radius estimation uses a variety of constraint curves (power, exponential, quadratic, Bailey, and B-spline) with zero-crossing points and monotonically increasing value ranges to output stem shape analysis results, and only B-spline is used for modeling.

6. The method for extracting tree structure parameters and 3D reconstruction from terrestrial laser point cloud data according to claim 1, characterized in that: The model of the arterial snake establishes a mapping relationship between cross-section vertices to facilitate volume and surface area calculations.

7. The method for extracting tree structure parameters and 3D reconstruction from terrestrial laser point cloud data according to claim 1, characterized in that: The simple leaf arrangement algorithm can realize basic simulation of opposite, alternate, whorled and clustered leaves according to the parameterized leaf size and arrangement.

Images