A method, system, terminal equipment, and media for tree point cloud branch and leaf separation based on multidimensional hypervoxel random forest.

Abstract

This invention discloses a method, system, terminal device, and medium for separating tree point cloud branches and leaves based on multidimensional supervoxel random forest, relating to the field of tree point cloud branch and leaf separation technology. The method includes: clustering the tree point cloud to obtain a tree supervoxel set; extracting features from the tree supervoxel set to obtain multidimensional features; inputting the multidimensional features of the tree supervoxel set into a pre-trained random forest classifier model to obtain an initial branch supervoxel set and an initial leaf supervoxel set; and applying class distance constraints to the initial branch supervoxel set and the initial leaf supervoxel set to obtain the branch supervoxel set and the leaf supervoxel set, respectively. This invention reduces the impact of missing points and noise by using supervoxel units, and combines multidimensional features and the robustness of random forests to accurately separate branches and leaves, providing support for tree parameter extraction, 3D modeling, and forestry management, and is applicable to complex tree structure scenarios.

Description

Technical Field

[0001] This invention relates to the field of tree point cloud branch and leaf separation technology, and in particular to a method, system, terminal equipment and medium for tree point cloud branch and leaf separation based on multidimensional hypervoxel random forest. Background Technology

[0002] In most forestry applications, distinguishing between wood and leaves is a crucial first step when using terrestrial lidar systems for tree point cloud analysis. Understanding the precise structure of trees is essential for accurate parameter assessment and the construction of high-precision 3D models. While visualization and other fields often focus on virtual tree modeling, they neglect the precision and accuracy of branch and trunk geometry, making it difficult to quantitatively extract structural parameters and factors that accurately reflect the tree's true growth status, such as branch grade and number, branch volume, trunk volume, aboveground biomass, diameter at breast height (DBH), and tree height. Structural modeling after separating branches and leaves from tree point clouds not only enhances the realism of 3D scenes but, more importantly, provides effective methods for the scientific management of trees. Therefore, researching precise methods for separating branches and leaves from tree point clouds is necessary and crucial for forestry and multiple other disciplines. However, due to the existence of different tree structures or species, separating wood and leaves remains a challenging task. This is especially true for complex trees, where clearly distinguishing between wood points and leaf points is often difficult.

[0003] Current methods for separating branches and leaves from tree point clouds mainly rely on two basic processes: first, extracting point cloud features and using supervised or unsupervised classification algorithms to determine the semantic category of each point; then, classifying it as either a branch or a leaf. For supervised classification methods, feature extraction primarily focuses on choosing different neighborhood sizes, which can lead to instability in feature calculation due to issues such as local missing data and noise. Methods using point cloud feature computation networks for branch and leaf separation require substantial data support, and even with ample data, the model's generalization ability remains a critical issue requiring further research and improvement. Unsupervised classification methods, on the other hand, require extensive parameter settings, and their robustness to tree species and size needs further testing and improvement. Furthermore, balancing robustness and compactness remains a significant challenge.

[0004] Therefore, there is an urgent need for a branch and leaf separation method that is less affected by local missing data and noise in point cloud data, and has robustness and compactness, to fill the gap in the existing technology. Summary of the Invention

[0005] The technical problem this invention aims to solve is that, in the field of tree point cloud branch and leaf separation, existing separation methods suffer from several drawbacks. Supervised classification is affected by missing points in the point cloud and noise, leading to unstable feature calculations and insufficient model generalization ability. Unsupervised classification suffers from strong parameter dependence and poor robustness to tree species and size, making it difficult to accurately distinguish between tree points and leaf points in complex tree structures. Therefore, an effective solution is urgently needed to address these technical problems.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention provides a method for separating tree point cloud branches and leaves based on a multidimensional hypervoxel random forest, the method comprising:

[0008] Clustering of tree point clouds yields a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several tree point cloud neighboring points and a center point.

[0009] Feature extraction is performed on the tree hypervoxel set to obtain multi-dimensional features of the tree hypervoxel set;

[0010] The multi-dimensional features of the tree supervoxel set are input into a pre-trained random forest classifier model to obtain the initial tree branch supervoxel set and the initial leaf supervoxel set.

[0011] The initial tree branch hypervoxel set and the initial leaf hypervoxel set are subjected to class distance constraint processing to obtain the tree branch hypervoxel set and the leaf hypervoxel set, wherein the tree branch hypervoxel set includes only tree branch hypervoxels to represent trees, and the leaf hypervoxel set includes only leaf hypervoxels to represent leaves.

[0012] In one implementation, the clustering of tree point clouds to obtain a set of tree supervoxels includes:

[0013] Randomly select several seed points in the tree point cloud;

[0014] Based on the seed point and the feature distance, the neighboring points of each seed point are clustered to obtain the corresponding number of tree hypervoxels;

[0015] The center point corresponding to each tree hypervoxel was calculated;

[0016] The neighborhood information corresponding to each tree hypervoxel is calculated;

[0017] The tree hypervoxels and their corresponding center points are integrated with the neighborhood information to obtain the tree hypervoxel set.

[0018] In one implementation, the neighborhood information is used to represent the adjacency relationship between tree supervoxels, and the calculation of the neighborhood information corresponding to each tree supervoxel includes:

[0019] For each boundary between adjacent tree supervoxels, a surface-based convexity detection is performed to obtain the corresponding concavity-convexity relationship features;

[0020] When the concavity-convexity relationship feature is a convex connection, it indicates that the adjacent tree hypervoxels belong to the same category;

[0021] When the concavity-convexity relationship feature is a concave connection, it indicates that the adjacent tree hypervoxels belong to different categories;

[0022] By integrating all the aforementioned concavity-convexity relation features, neighborhood information corresponding to each tree supervoxel is obtained.

[0023] In one implementation, the multi-dimensional features include local features, and the feature extraction of the tree supervoxel set to obtain the multi-dimensional features of the tree supervoxel set includes:

[0024] Calculate the ratio of the number of point clouds within each tree hypervoxel to the volume of the tree hypervoxel to obtain local density features;

[0025] For each tree hypervoxel, a preset number of feature values ​​are calculated through feature decomposition, and the geometric shape features corresponding to the tree hypervoxel are obtained based on the combination of the feature values.

[0026] By integrating all the local density features with the geometric features, local features are obtained.

[0027] In one implementation, the multi-dimensional features include global structural features, and the feature extraction of the tree supervoxel set to obtain the multi-dimensional features of the tree supervoxel set includes:

[0028] Calculate the distance between the center point of each tree hypervoxel and the ground in a preset direction to obtain the relative height feature;

[0029] Calculate the angle between the normal of the center point of each tree hypervoxel and the horizontal plane to obtain the point cloud directional features;

[0030] By integrating all the relative height features and the point cloud orientation features, a global structural feature is obtained.

[0031] In one implementation, the multi-dimensional features include neighborhood features, and the feature extraction of the tree supervoxel set to obtain the multi-dimensional features of the tree supervoxel set includes:

[0032] Based on the neighborhood information, the number of neighboring supervoxels of the same type for each tree supervoxel is calculated to obtain the adjacency connection quantity feature.

[0033] Based on the neighborhood information, the geometric shape features of the same type of adjacent supervoxels of each tree supervoxel are calculated to obtain the neighborhood shape features of the supervoxel.

[0034] By integrating all the adjacency connection quantity features and the supervoxel neighborhood shape features, the neighborhood features are obtained.

[0035] In one implementation, the pre-trained random forest classifier model includes several decision trees, and the training process of the pre-trained random forest classifier model includes:

[0036] Using the multi-dimensional features of several tree point clouds and their corresponding tree supervoxes as a training set, and randomly sampling to generate multiple training subsets, a decision tree is constructed for each training subset, and a preset number of features are randomly selected when each decision tree splits at a node.

[0037] Based on the constructed decision trees, the categories of tree hypervoxels are voted on to generate preliminary classification results;

[0038] Based on the preliminary classification results, the number of decision trees is determined through cross-validation.

[0039] The performance of the random forest classifier model was verified using a test set, and a pre-trained random forest classifier model with expected performance was obtained.

[0040] Secondly, embodiments of the present invention also provide a tree point cloud branch and leaf separation system based on multidimensional hypervoxel random forest, the system comprising:

[0041] The tree supervoxel clustering module is used to cluster tree point clouds to obtain a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several tree point cloud neighboring points and a center point.

[0042] A multi-dimensional feature acquisition module is used to extract features from the tree hypervoxel set to obtain multi-dimensional features of the tree hypervoxel set.

[0043] The initial classification module is used to input the multi-dimensional features of the tree supervoxel set into a pre-trained random forest classifier model to obtain the initial branch supervoxel set and the initial leaf supervoxel set.

[0044] The classification optimization module is used to perform category distance constraint processing on the initial tree branch hypervoxel set and the initial leaf hypervoxel set to obtain the tree branch hypervoxel set and the leaf hypervoxel set, wherein the tree branch hypervoxel set includes only tree branch hypervoxels to represent trees, and the leaf hypervoxel set includes only leaf hypervoxels to represent leaves.

[0045] Thirdly, embodiments of the present invention also provide a terminal device, the terminal device including a memory, a processor, and a tree point cloud branch and leaf separation program based on multidimensional supervoxel random forest stored in the memory and executable on the processor. When the processor executes the tree point cloud branch and leaf separation program based on multidimensional supervoxel random forest, it implements the steps of the tree point cloud branch and leaf separation method based on multidimensional supervoxel random forest described in any of the above schemes.

[0046] Fourthly, embodiments of the present invention also provide a computer-readable storage medium storing a tree point cloud branch and leaf separation program based on a multidimensional supervoxel random forest. When the tree point cloud branch and leaf separation program based on a multidimensional supervoxel random forest is executed by a processor, it implements the steps of the tree point cloud branch and leaf separation method based on a multidimensional supervoxel random forest as described in any of the above schemes.

[0047] Beneficial Effects: This invention discloses a method, system, terminal device, and medium for separating tree point clouds branches and leaves based on a multidimensional supervoxel random forest. The method first clusters the tree point cloud to obtain a tree supervoxel set, wherein each tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several neighboring tree point cloud points and a center point. Next, feature extraction is performed on the tree supervoxel set to obtain multidimensional features. Then, the multidimensional features of the tree supervoxel set are input into a pre-trained random forest classifier model to obtain an initial branch supervoxel set and an initial leaf supervoxel set. Finally, class distance constraint processing is applied to the initial branch supervoxel set and the initial leaf supervoxel set to obtain a branch supervoxel set and a leaf supervoxel set, wherein the branch supervoxel set only includes branch supervoxels to represent branches, and the leaf supervoxel set only includes leaf supervoxels to represent leaves. This invention transforms point clouds into structured supervoxel units through supervoxel clustering, effectively reducing local missing data and noise interference, and improving feature stability. Simultaneously, it expresses differences in branch and leaf structure through multi-dimensional feature extraction, including local features, global structural features, and neighborhood features, and enhances the generalization ability of the classification model by combining it with a random forest classifier. Finally, it further optimizes the classification boundary through class distance constraint processing, improving the accuracy of branch and leaf separation. This invention provides support for tree parameter extraction, 3D modeling, and forestry management, and is applicable to scenarios with complex tree structures. Attached Figure Description

[0048] Figure 1 The flowchart illustrates a specific implementation method for the tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest provided in this embodiment of the invention.

[0049] Figure 2 This is a comparison image before and after separation in the tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest provided in the embodiments of the present invention.

[0050] Figure 3 This is a flowchart of the steps of the tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest provided in the embodiments of the present invention.

[0051] Figure 4 This is a comparison diagram before and after class distance constraint in the tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest provided in the embodiments of the present invention.

[0052] Figure 5 This is a schematic diagram of the tree point cloud branch and leaf separation device based on multidimensional hypervoxel random forest provided in the embodiments of the present invention.

[0053] Figure 6 This is a block diagram illustrating the internal structure of the terminal device provided in an embodiment of the present invention. Detailed Implementation

[0054] To make the objectives, technical solutions, and effects of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0055] The flowchart shown in the attached diagram is for illustrative purposes only and does not necessarily include all content, operations, or steps, nor does it require execution in the described order. For example, some operations or steps can be broken down, combined, or partially merged, so the actual execution order may change depending on the actual situation.

[0056] It should be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0057] It should be understood that, in order to clearly describe the technical solutions of the embodiments of the present invention, the terms "first" and "second" are used in the embodiments of the present invention to distinguish identical or similar items with essentially the same function and effect. For example, "first control information" and "second control information" are only used to distinguish different control information and do not limit their order.

[0058] Those skilled in the art will understand that the words "first" and "second" do not limit the quantity or the order of execution, and that the words "first" and "second" do not necessarily imply that they are different.

[0059] It should also be understood that the term “and / or” as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0060] In the field of digital management and 3D modeling of forestry resources, the separation of tree point clouds into branches and leaves is a crucial foundation for accurately obtaining structural parameters such as branch grade, volume, and biomass. While terrestrial lidar can efficiently collect tree point cloud data, it suffers from problems such as large data volume, local missing data, and noise interference in complex scenarios. Existing separation methods mainly rely on supervised and unsupervised classification: supervised classification achieves separation through neighborhood feature extraction and classifier training, but is affected by missing points and noise, leading to unstable feature calculations, and deep models require a large amount of labeled data, resulting in insufficient generalization ability; unsupervised classification faces challenges such as complex parameter settings and poor robustness to tree species and size, making it difficult to adapt to the structural differences of different trees. Furthermore, traditional techniques often focus on the visualization effects of virtual modeling, neglecting the accuracy requirements of branch geometry, resulting in the inability to quantitatively extract key parameters reflecting the true growth status of trees, thus hindering the development of forestry scientific management and 3D model construction.

[0061] Therefore, existing technologies face many problems in practical applications, including the instability of feature calculation and insufficient generalization ability of supervised classification methods due to local missing points in the point cloud and noise interference. Unsupervised classification also suffers from strong parameter dependence and poor adaptability to tree species and size, making it difficult to accurately distinguish tree points from leaf points in complex trees. Furthermore, existing methods lack refined processing of branch and trunk geometry, failing to meet the need for accurate extraction of tree structural parameters. How to combine structured point cloud processing technology with robust classification models to optimize the shortcomings of existing methods in noise resistance, generalization, and parameter adaptability, and achieve high-precision tree point cloud branch and leaf separation, has become a core technical problem that urgently needs to be solved.

[0062] To address the aforementioned technical challenges, this invention proposes a method for separating tree branches and leaves from point clouds based on multi-dimensional supervoxel feature constraints. Specifically, it's a method for automatically separating the woody parts and leaves of tree point clouds from 3D LiDAR point clouds of ground-based trees. The core idea is to transform the classification problem of the original point cloud into the classification of supervoxels, eliminating the influence of local errors in the point cloud on the classification results. Specifically, firstly, the point cloud is converted into a supervoxel structure. Local features, global structural features, and neighborhood features of the supervoxels are extracted, and a random forest classification method is used for coarse classification of tree branches and leaves. Subsequently, class distance constraints are applied to the coarsely classified branch and leaf point clouds to further optimize the branch and leaf separation results. Figure 2 This paper demonstrates the comparative effect of branch and leaf separation before and after the proposed tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest.

[0063] This embodiment provides a tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest, such as... Figure 1 As shown, the specific steps include the following:

[0064] Step S100: Cluster the tree point cloud to obtain a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several tree point cloud neighboring points and a center point.

[0065] In this embodiment, as Figure 3 The separation method flowchart is shown. In the coarse classification stage of tree point clouds, the original tree point cloud is first converted into a group of supervoxels using a point cloud supervoxel clustering method. This group of supervoxels is called tree supervoxels, and each tree supervoxel represents that the point cloud in that supervoxel is a branch or leaf. The center point representing the supervoxel is then extracted. Point clouds in the same class of supervoxels usually have similar feature information, and there is a certain adjacency relationship between the supervoxels.

[0066] In one implementation, the clustering of tree point clouds to obtain a tree supervoxel set specifically includes the following steps:

[0067] Step S110: Randomly select several seed points in the tree point cloud;

[0068] Step S120: Based on the seed point and feature distance, cluster the neighboring points of each seed point to obtain the corresponding number of tree hypervoxels;

[0069] Step S130: Calculate the center point corresponding to each tree hypervoxel;

[0070] Step S140: Calculate the neighborhood information corresponding to each tree hypervoxel;

[0071] Step S150: Integrate all the tree hypervoxels and their corresponding center points with neighborhood information to obtain the tree hypervoxel set.

[0072] In this embodiment, the original point cloud of the tree is first converted into a group of supervoxels by the point cloud VCCS (Voxel Cloud Connectivity Segmentation) supervoxel clustering method, and the centroid of each supervoxel is calculated. The center point representing the supervoxel is extracted from it. Point clouds in the same class of supervoxels usually have similar feature information, and there is a certain adjacency relationship between the supervoxels.

[0073] The construction of hypervoxels involves two main steps. First, several seed points are randomly selected from the point cloud data. Second, neighboring points are clustered based on feature distance to obtain hypervoxels.

[0074] Specifically, after randomly selecting several seed points from the point cloud data and using these seed points as the centers of clusters, starting from each seed point, neighboring points are added to the corresponding cluster based on feature distance. This process can be based on a threshold to determine which points are similar enough to be classified into the same cluster.

[0075] In the aforementioned clustering algorithm for neighboring points, the feature distance is used to evaluate the similarity of points in the point cloud data. That is, in the supervoxel construction step, clustering is based on feature distance. Preferably, the feature distance can include multiple dimensions such as color, spatial location, and normal information to estimate the uniformity of points. This uniformity estimate is used to identify groups of points similar in the aforementioned features, thereby forming supervoxels. It can be understood that feature distance is a measure of the differences between points in specific features such as color, spatial location, and normal. In supervoxel clustering, feature distance is used to evaluate the uniformity of points, i.e., the degree of similarity of points in these features.

[0076] Preferably, the supervoxel clustering algorithm estimates the uniformity of points using color, spatial, and normal dimensions, and uses this uniformity as the feature distance for clustering to determine similarity, as shown in the following formula:

[0077]

[0078]

[0079]

[0080] in, This represents a summary estimate of uniformity across all dimensions. This represents the Euclidean distance between the seed point and its surrounding points. This represents the normal to the plane fitted by the least squares fitting method based on nearest neighbors. Indicates the sampling step size or sampling interval in the spatial direction. This indicates the sampling step size or sampling interval in the normal direction. , , These represent the coordinate differences between the seed point and its surrounding points along the three coordinate axes. Indicates the resolution or size of a voxel. and Let's represent two vectors that are used to calculate the normal distance. Use when needed.

[0081] Based on a summary estimate of uniformity across all dimensions, neighboring points are added to their respective clusters, and this step is repeated until no more points can be added to any cluster, or a preset stopping condition is met. Preferably, a preset number of clusters and cluster size can be used. After clustering is complete, the clustering results can be further optimized, for example, by merging smaller clusters or adjusting cluster boundaries to improve cluster quality and consistency. After clustering is complete, the quality of the clustering results can be evaluated by calculating the uniformity within clusters and the separation between clusters.

[0082] After the clustering algorithm is completed, the centroid of each obtained hypervoxel is calculated and recorded as the center point of the hypervoxel. In this way, all the hypervoxels obtained by clustering and their center points can be regarded as a first-level hypervoxel model, which includes a set of hypervoxels, where each hypervoxel represents a branch or leaf.

[0083] Subsequently, the neighborhood information corresponding to each tree hypervoxel is calculated, and all tree hypervoxels and their corresponding center points are integrated with the neighborhood information to obtain the tree hypervoxel set.

[0084] In one implementation, the neighborhood information is used to represent the adjacency relationship between tree hypervoxels, and the calculation of the neighborhood information corresponding to each tree hypervoxel specifically includes the following steps:

[0085] Step S141: Perform surface-based convexity detection on the boundary between each adjacent tree supervoxel to obtain the corresponding concavity-convexity relationship features;

[0086] Step S142: When the concavity-convexity relationship feature is a convex connection, it indicates that the adjacent tree hypervoxels belong to the same category;

[0087] Step S143: When the concavity-convexity relationship feature is a concave connection, it indicates that the adjacent tree hypervoxels are of different categories;

[0088] Step S144: Integrate all the concavity-convexity relation features to obtain the neighborhood information corresponding to each tree supervoxel.

[0089] In this embodiment, to accurately define the neighborhood relationships between supervoxels, the LCCP (Locally Convex Connected Patches) algorithm is further applied to extract the second-level supervoxel model based on the first-level supervoxel model described above. This second-level supervoxel model contains several groups of supervoxels, where all supervoxels within each group can be considered to represent either leaves or branches. It can be understood that the first-level model is used for preliminary clustering, while the second-level model is used for refining adjacency relationships. The LCCP algorithm determines neighborhood information by utilizing the inherent connectivity relationships within supervoxels; this neighborhood information represents the adjacency relationships between tree supervoxels.

[0090] Through surface-based convexity detection, the boundaries between adjacent supervoxels are assigned concave or convex feature identifiers. When two supervoxels have a concave relationship (i.e., they are concavely connected), they are considered to belong to different objects; when two supervoxels have a convex relationship (i.e., they are convexly connected), they are considered to belong to the same object. Therefore, after processing based on the LCCP algorithm, the adjacency relationships of supervoxels are assigned concave and convex properties, which are used to quickly and accurately obtain robust neighborhood information during feature calculation. It is understood that the neighborhood information includes the concave / convex connectivity relationship between adjacent supervoxels, i.e., feature information on whether adjacent supervoxels belong to the same type of object.

[0091] The calculation method for determining the concavity / convexity relationship is shown in the following formula:

[0092]

[0093] in, , and This represents the center of mass of the two observed supervoxels. and Let them represent their normal vectors. Used to determine whether a concave-convex relationship is a convex connection or a concave connection, when When the value is greater than 0, this relationship is considered a convex connection, which means that the normal vector of the current hypervoxel is connected to the normal vector of the hypervoxel. The angle between the defined linear vectors is very small; when When <0, the relationship is considered a concave connection.

[0094] Step S200: Extract features from the tree supervoxel set to obtain multi-dimensional features of the tree supervoxel set.

[0095] In this embodiment, as Figure 3 As shown, after performing clustering and related calculations on the original point cloud to obtain supervoxel clusters, supervoxel center points, and supervoxel neighborhood information, it is necessary to extract the features of the supervoxels. Specifically, the extracted features are divided into three categories: local features, global structural features, and neighborhood features.

[0096] In one implementation, the multi-dimensional features include local features. The step of extracting features from the tree supervoxel set to obtain the multi-dimensional features of the tree supervoxel set specifically includes the following steps:

[0097] Step S210: Calculate the ratio of the number of point clouds in each tree hypervoxel to the volume of the tree hypervoxel to obtain the local density features;

[0098] Step S220: For each tree hypervoxel, calculate a preset number of feature values ​​through feature decomposition, and obtain the geometric shape features corresponding to the tree hypervoxel based on the combination of the feature values.

[0099] Step S230: Integrate all the local density features and the geometric shape features to obtain local features.

[0100] In this embodiment, local features are first extracted from the multidimensional features. Specifically, local features involve the eigenvalues ​​and zenith angles of the tree point cloud feature vectors at a local spatial scale. These features primarily reflect the different spatial distributions of the trunk, branches, and leaves, and this difference in spatial distribution is used to separate leaf points from branch points. When extracting local features, hypervoxels are used as the basic unit of computation to obtain the structural information of the point cloud data within the hypervoxels. This structural information constitutes the local features of the hypervoxels.

[0101] Specifically, the local density of the point cloud is first calculated as the local density feature of the supervoxel. The local density feature is calculated as the ratio of the number of point clouds within a supervoxel to the volume of the supervoxel. In the point cloud branch and leaf separation task, there is a significant difference between the supervoxel point cloud density of trees and the supervoxel point cloud density of leaves.

[0102] Secondly, the shape features of the supervoxel point cloud are obtained as the geometric shape features corresponding to the supervoxels. Three eigenvalues ​​λ1, λ2, and λ3, and their corresponding zenith angles θ1, θ2, and θ3 are calculated through eigenvalue decomposition. The zenith angle refers to the angle between the observation point starting from the plumb line and the plumb line. The eigenvalues ​​are sorted in descending order of λ1≥λ2≥λ3≥0 and θ1≥θ2≥θ3≥0. Based on the mathematical meaning of the eigenvalues, different combinations of eigenvalues ​​exhibit different shape features.

[0103] The specific methods for calculating the local density and geometric features of point clouds are shown in Table 1:

[0104] Table 1

[0105]

[0106] Volume density represents the density of the point cloud within the neighborhood of a supervoxel. This represents the total number of point clouds within the neighborhood of a supervoxel. (neighbourhood volume) represents the volume of the supervoxel's neighborhood; λ1, λ2, and λ3 represent the eigenvalues ​​obtained from eigenvalue decomposition; θ1, θ2, and θ3 represent the zenith angles corresponding to the eigenvalues; anisotropy represents the degree of non-uniformity of the distribution of the supervoxel point cloud in different spatial directions; the proportion of the second principal component (PCA2) represents the contribution of the point cloud distribution in the secondary principal direction, i.e., the direction corresponding to λ2; surface variation represents the contribution of the supervoxel point cloud distribution in the surface direction, i.e., the direction corresponding to λ3; and sphericity represents the degree to which the distribution of the supervoxel point cloud approaches sphericity.

[0107] Finally, by integrating the aforementioned local density and geometric features of the hypervoxel, its local features are obtained.

[0108] In one implementation, the multi-dimensional features include global structural features. The step of extracting features from the tree supervoxel set to obtain the multi-dimensional features of the tree supervoxel set specifically includes the following steps:

[0109] Step S240: Calculate the distance between the center point of each tree hypervoxel and the ground in a preset direction to obtain the relative height feature;

[0110] Step S250: Calculate the angle between the normal of the center point of each tree hypervoxel and the horizontal plane to obtain the point cloud direction features;

[0111] Step S260: Integrate all the relative height features and the point cloud direction features to obtain global structural features.

[0112] In this embodiment, global structural features are extracted from the multidimensional features. To further improve the classification performance of point cloud data, global features are used in addition to utilizing local features of supervoxels. Specifically, features of the supervoxel center points in the original point cloud data are extracted to enhance the robustness of the model. Two types of point cloud features are mainly used: relative height features and point cloud orientation features. The fusion of these features enriches the descriptive dimensions of the data and can more accurately capture the inherent structure and morphology of the point cloud data, thereby achieving higher accuracy in classification tasks.

[0113] Specifically, the relative height feature is first obtained by calculating the distance from the center point of the hypervoxel to the ground in the extended z-direction.

[0114] Next, the angle between the normal to the supervoxel center point and the horizontal plane is calculated to obtain the point cloud orientation features. The formula for calculating the angle between the normal to the supervoxel center point and the horizontal plane is as follows:

[0115]

[0116] in, c The value of cosine. Let be the normal vector of the hypervoxel. Let be the normal vector of the horizontal plane, defined as (0,0,1). , , It is the supervoxel normal vector The three components are in a three-dimensional Cartesian coordinate system. To facilitate feature normalization, this paper uses cosine values ​​to represent the directional features of the supervoxel.

[0117] Finally, all the relative height features and the point cloud orientation features are integrated to obtain the global structure features.

[0118] In one implementation, the multi-dimensional features include neighborhood features, and the feature extraction of the tree supervoxel set to obtain the multi-dimensional features of the tree supervoxel set specifically includes the following steps:

[0119] Step S270: Based on the neighborhood information, calculate the number of neighboring supervoxels of the same type for each tree supervoxel to obtain the adjacency connection number feature.

[0120] Step S280: Based on the neighborhood information, calculate the geometric shape features of the same type of adjacent supervoxels of each tree supervoxel to obtain the supervoxel neighborhood shape features.

[0121] Step S290: Integrate all the adjacent connection quantity features and the supervoxel neighborhood shape features to obtain neighborhood features.

[0122] In this embodiment, neighborhood features are extracted from the multidimensional features. Compared with the original point cloud data, the point cloud processed by supervoxel clustering is used to represent the connection relationships between supervoxels. These connections can be used to express the adjacency of different elements in the point cloud. For example, supervoxels located in the branches usually have a simpler adjacency structure, while supervoxels in the leaves exhibit more complex adjacency relationships. By analyzing the above connection relationships, branches can be avoided from being incorrectly classified as leaves, thereby improving the classification accuracy of the point cloud data. Based on this, in addition to using random forest to extract local and global features of supervoxels, further classification is performed based on the neighborhood features of supervoxels to optimize the classification effect. After the neighborhood information is calculated using the LCCP supervoxel clustering algorithm, the boundaries of different targets can be accurately defined. Subsequently, supervoxels that meet the conditions are used for neighborhood-based feature value calculation. The combination of the number of adjacencies of the supervoxel center point and the neighborhood geometric features is used as the neighborhood features to enhance the classification features.

[0123] Specifically, the number of adjacencies of a supervoxel's center point is first calculated and used as a feature of adjacency connectivity. This feature is defined as the number of adjacent supervoxel center points of a given supervoxel in the supervoxel partitioning. It is important to emphasize that for any given supervoxel, supervoxels with convex connections (i.e., supervoxels of the same type) can be considered adjacent supervoxels. In other words, when calculating the number of adjacencies of adjacent supervoxel center points, supervoxels with concave connections are excluded. By searching the supervoxel neighborhood information using the LCCP segmentation algorithm, the adjacency relationship between each supervoxel's center point and its corresponding adjacent supervoxel center points can be obtained, i.e., the concave-convex connectivity relationship in the neighborhood information. This adjacency relationship can reflect the complexity and structural characteristics of tree point cloud supervoxels; for example, supervoxel center points located on branches have fewer adjacencies, while those located on leaves have more.

[0124] Secondly, the geometric shape features of the neighborhood point cloud are calculated as the shape features of the supervoxel neighborhood. Feature decomposition is performed on the neighborhood point cloud to obtain feature values ​​λ1, λ2, and λ3. By employing three types of features, the geometric shape features of the neighborhood point cloud are calculated. The calculation methods for each feature are shown in Table 2.

[0125] Table 2

[0126]

[0127] Wherein, λ1, λ2, and λ3 represent the eigenvalues ​​obtained from the eigenvalue decomposition of the neighborhood point cloud, linearity represents the degree to which the spatial distribution of the supervoxel neighborhood point cloud approaches a straight line, anisotropy represents the degree of non-uniformity of the distribution of the supervoxel point cloud in different spatial directions, and sphericity represents the degree to which the distribution of the supervoxel point cloud approaches a spherical shape.

[0128] Finally, all the adjacency connection quantity features and the supervoxel neighborhood shape features are integrated to obtain the neighborhood features.

[0129] Step S300: Input the multi-dimensional features of the tree supervoxel set into the pre-trained random forest classifier model to obtain the initial tree branch supervoxel set and the initial tree leaf supervoxel set.

[0130] In this embodiment, as Figure 3 By constructing a decision tree, performing voting classification, conducting cross-validation, and validating the model, a pre-trained random forest classifier model is built and trained. Then, the calculated features are fed into the random forest classifier model for training, yielding coarse classification results.

[0131] In one implementation, the pre-trained random forest classifier model includes several decision trees, and the training process of the pre-trained random forest classifier model specifically includes the following steps:

[0132] Step S310: Using the multi-dimensional features of several tree point clouds and their corresponding tree supervoxes as a training set, and randomly sampling to generate multiple training subsets, constructing a decision tree for each training subset, and randomly selecting a preset number of features when splitting nodes in each decision tree.

[0133] Step S320: Based on the constructed decision trees, vote on the categories of tree hypervoxels to generate preliminary classification results;

[0134] Step S330: Based on the preliminary classification results, determine the number of decision trees through cross-validation;

[0135] Step S340: Use the test set to verify the performance of the random forest classifier model and obtain a pre-trained random forest classifier model that meets the expected performance.

[0136] In this embodiment, to classify point cloud data by integrating the above three types of features, a supervoxel-based random forest classification model needs to be constructed. First, since random forests have good robustness in handling outliers, they are chosen as the classifier. A portion of the dataset containing point clouds of several trees is used as the training set, and the remainder as the test set. Specifically, by using randomly sampled data and feature subsets during the construction of each tree, overfitting of the model can be effectively avoided. Preferably, 70% of the dataset is randomly selected as the training set, and a data and feature subset is randomly selected from the training set as the training subset.

[0137] In the task of classifying leaves and wood, the random forest classifier used demonstrates superior performance compared to other single machine learning algorithms, such as Naive Bayes and neural networks. In this embodiment, the construction of the random forest model is constrained by two key parameters: the number of input features and the number of decision trees. The number of input features determines the size of the feature subset randomly selected by the model at each decision tree node split, affecting the model's bias and variance, and consequently its generalization ability. For feature selection, default values ​​are used. The number of features used to construct each decision tree was considered. Furthermore, to determine the optimal number of decision trees in the model, cross-validation was used to systematically analyze the performance of models with 100 to 600 trees, and based on the analysis results, 500 decision trees were determined to be the most accurate.

[0138] During training, supervoxes are used as the basic classification units. The three-dimensional features of the extracted supervoxes are used as training information input to supervise the training of the random forest classification model. The performance of the random forest classifier model is verified using the test set, and a pre-trained random forest classifier model with expected performance is obtained.

[0139] Step S400: Perform category distance constraint processing on the initial tree branch hypervoxel set and the initial leaf hypervoxel set to obtain the tree branch hypervoxel set and the leaf hypervoxel set, wherein the tree branch hypervoxel set only includes tree branch hypervoxels to represent trees, and the leaf hypervoxel set only includes leaf hypervoxels to represent leaves.

[0140] In this embodiment, after the random forest model completes the initial classification of the supervoxel center point cloud, a classification optimization strategy based on class distance constraints is adopted to further improve the accuracy of point cloud data classification. Specifically, for the branch point cloud that may remain in the leaf point cloud, it is compared and analyzed with the already classified branch point cloud, and a preferred parameter of 0.5m is introduced as the distance threshold. Through this distance constraint, errors that may occur during the classification process can be effectively identified and corrected, thereby ensuring the accuracy of the classification results. Figure 4 As shown, after applying class distance constraints to the two classes of supervoxels obtained from the coarse classification by the pre-trained random forest classifier model, the classification effect is better.

[0141] Finally, the result set is obtained as the tree branch hypervoid set and the leaf branch hypervoid set, representing the tree branch and the leaf branch, respectively.

[0142] In summary, the technical solution described in the above embodiments effectively reduces the impact of local missing data and noise on the classification results by using supervoxels as the basic units for classifying 3D point cloud data of trees. Furthermore, based on the different characteristics of branch and leaf point clouds, the adjacency relationship features between supervoxels are fully considered, and random forests are used for classification based on multi-dimensional feature extraction. A class distance constraint method is also introduced to optimize the classification results. Ultimately, this provides support for tree parameter extraction, 3D modeling, and forestry management, and is suitable for scenarios with complex tree structures.

[0143] like Figure 5 As shown in the figure, this embodiment of the invention provides a tree point cloud branch and leaf separation system based on multidimensional hypervoxel random forest. The system includes: a tree hypervoxel clustering module 10, a multidimensional feature acquisition module 20, an initial classification module 30, and a classification optimization module 40.

[0144] Specifically, the tree supervoxel clustering module 10 is used to cluster the tree point cloud to obtain a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several neighboring tree point cloud points and a center point; the multi-dimensional feature acquisition module 20 is used to extract features from the tree supervoxel set to obtain multi-dimensional features of the tree supervoxel set; the initial classification module 30 is used to input the multi-dimensional features of the tree supervoxel set into a pre-trained random forest classifier model to obtain an initial branch supervoxel set and an initial leaf supervoxel set; the classification optimization module 40 is used to perform class distance constraint processing on the initial branch supervoxel set and the initial leaf supervoxel set to obtain a branch supervoxel set and a leaf supervoxel set, wherein the branch supervoxel set only includes branch supervoxels to represent branches, and the leaf supervoxel set only includes leaf supervoxels to represent leaves.

[0145] In one implementation, the tree hypervoxel clustering module includes:

[0146] The seed point selection unit is used to randomly select several seed points in the tree point cloud;

[0147] Clustering unit, used to cluster the neighboring points of each seed point based on the seed point and the feature distance, to obtain the corresponding number of tree hypervoxels;

[0148] The center point acquisition unit is used to calculate the center point corresponding to each tree hypervoxel.

[0149] The neighborhood information acquisition unit is used to calculate the neighborhood information corresponding to each tree hypervoxel.

[0150] The tree supervoxel set integration unit is used to integrate all the tree supervoxels and their corresponding center points and neighborhood information to obtain the tree supervoxel set.

[0151] In one implementation, the neighborhood information is used to represent the adjacency relationship between tree supervoxels, and the neighborhood information acquisition unit includes:

[0152] The concavity-convexity relationship feature detection subunit is used to perform surface-based convexity detection on the boundary between each adjacent tree supervoxel to obtain the corresponding concavity-convexity relationship features.

[0153] A convex connection determination subunit is used to indicate that the adjacent tree hypervoxels belong to the same category when the concavity-convexity relationship feature is a convex connection.

[0154] The concave connection determination subunit is used to indicate that the adjacent tree hypervoxels are of different categories when the concave-convexity relationship feature is a concave connection;

[0155] The concavity-convexity relation feature integration subunit is used to integrate all the concavity-convexity relation features to obtain the neighborhood information corresponding to each tree supervoxel.

[0156] In one implementation, the multi-dimensional features include local features, and the multi-dimensional feature acquisition module includes:

[0157] An ontology model setting unit is used to set the ontology model as a constraint condition in the large language model recognition process;

[0158] The local density feature acquisition unit is used to calculate the ratio of the number of point clouds in each tree hypervoxel to the volume of the tree hypervoxel, and obtain the local density feature.

[0159] The geometric shape feature acquisition unit is used to calculate a preset number of feature values ​​for each tree hypervoxel through feature decomposition, and obtain the geometric shape features corresponding to the tree hypervoxel based on the combination of the feature values.

[0160] The local feature integration unit is used to integrate all the local density features and the geometric shape features to obtain local features.

[0161] In one implementation, the multi-dimensional features include global structural features, and the multi-dimensional feature acquisition module includes:

[0162] The relative height feature acquisition unit is used to calculate the distance between the center point of each tree hypervoxel and the ground in a preset direction to obtain the relative height feature;

[0163] The point cloud orientation feature acquisition unit is used to calculate the angle between the normal of the center point of each tree hypervoxel and the horizontal plane to obtain the point cloud orientation feature.

[0164] The global structural feature integration unit is used to integrate all the relative height features and the point cloud orientation features to obtain global structural features.

[0165] In one implementation, the multi-dimensional features include neighborhood features, and the multi-dimensional feature acquisition module includes:

[0166] The adjacency connection quantity feature acquisition unit is used to calculate the number of similar adjacent supervoxels of each tree supervoxel based on the neighborhood information, and obtain the adjacency connection quantity feature.

[0167] The supervoxel neighborhood shape feature acquisition unit is used to calculate the geometric shape features of the same type of adjacent supervoxels of each tree supervoxel based on the neighborhood information, so as to obtain the supervoxel neighborhood shape features.

[0168] The neighborhood feature integration unit is used to integrate all the adjacent connection quantity features and the supervoxel neighborhood shape features to obtain neighborhood features.

[0169] In one implementation, the pre-trained random forest classifier model includes several decision trees, and the initial classification module includes:

[0170] The decision tree construction unit is used to use the multi-dimensional features of several tree point clouds and their corresponding tree supervoxes as a training set, and randomly sample to generate multiple training subsets. A decision tree is constructed for each training subset, and each decision tree randomly selects a preset number of features when splitting a node.

[0171] The preliminary classification result generation unit is used to vote on the categories of tree hypervoxels based on the constructed decision trees to generate preliminary classification results.

[0172] The decision tree number determination unit is used to determine the number of decision trees based on the preliminary classification results through cross-validation.

[0173] The model performance verification unit is used to verify the performance of the random forest classifier model using the test set, and obtain a pre-trained random forest classifier model that meets the expected performance.

[0174] Based on the above embodiments, the present invention also provides a terminal device, the principle block diagram of which can be as follows: Figure 6 As shown, the terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used to communicate with external terminals via a network connection. When executed by the processor, the computer program implements a tree point cloud branch and leaf separation method based on a multidimensional hypervoxel random forest. The display screen can be an LCD screen or an e-ink screen. The temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.

[0175] Those skilled in the art will understand that Figure 6 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0176] In one embodiment, a terminal device is provided, including a memory and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by one or more processors. The one or more programs include instructions for performing the following operations:

[0177] Clustering of tree point clouds yields a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several tree point cloud neighboring points and a center point.

[0178] Feature extraction is performed on the tree hypervoxel set to obtain multi-dimensional features of the tree hypervoxel set;

[0179] The multi-dimensional features of the tree supervoxel set are input into a pre-trained random forest classifier model to obtain the initial tree branch supervoxel set and the initial leaf supervoxel set.

[0180] The initial tree branch hypervoxel set and the initial leaf hypervoxel set are subjected to class distance constraint processing to obtain the tree branch hypervoxel set and the leaf hypervoxel set, wherein the tree branch hypervoxel set includes only tree branch hypervoxels to represent trees, and the leaf hypervoxel set includes only leaf hypervoxels to represent leaves.

[0181] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0182] In summary, this invention discloses a method, system, terminal device, and medium for tree point cloud branch and leaf separation based on multidimensional supervoxel random forest, relating to the field of tree point cloud branch and leaf separation technology. The method first clusters the tree point cloud to obtain a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several neighboring tree point cloud points and a center point. Next, feature extraction is performed on the tree supervoxel set to obtain multidimensional features of the tree supervoxel set. Then, the multidimensional features of the tree supervoxel set are input into a pre-trained random forest classifier model to obtain an initial branch supervoxel set and an initial leaf supervoxel set. Finally, class distance constraint processing is applied to the initial branch supervoxel set and the initial leaf supervoxel set to obtain a branch supervoxel set and a leaf supervoxel set, wherein the branch supervoxel set only includes branch supervoxels to represent branches, and the leaf supervoxel set only includes leaf supervoxels to represent leaves. This invention transforms point clouds into structured supervoxel units through supervoxel clustering, effectively reducing local missing data and noise interference, and improving feature stability. Simultaneously, it expresses differences in branch and leaf structure through multi-dimensional feature extraction, including local features, global structural features, and neighborhood features, and enhances the generalization ability of the classification model by combining it with a random forest classifier. Finally, it further optimizes the classification boundary through class distance constraint processing, improving the accuracy of branch and leaf separation. This invention provides support for tree parameter extraction, 3D modeling, and forestry management, and is applicable to scenarios with complex tree structures.

[0183] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0184] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for separating tree point cloud branches and leaves based on multidimensional hypervoxel random forest, characterized in that, The method includes: Clustering of tree point clouds yields a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several tree point cloud neighboring points and a center point. Feature extraction is performed on the tree supervoxel set to obtain multi-dimensional features of the tree supervoxel set, including local features, global structural features, and neighborhood features. The multi-dimensional features of the tree supervoxel set are input into a pre-trained random forest classifier model to obtain the initial tree branch supervoxel set and the initial leaf supervoxel set. The initial tree branch hypervoxel set and the initial leaf hypervoxel set are subjected to class distance constraint processing to obtain the tree branch hypervoxel set and the leaf hypervoxel set, wherein the tree branch hypervoxel set only includes tree branch hypervoxels to represent trees, and the leaf hypervoxel set only includes leaf hypervoxels to represent leaves. The feature extraction of the tree supervoxel set yields multi-dimensional features of the tree supervoxel set, including: Calculate the ratio of the number of point clouds within each tree hypervoxel to the volume of the tree hypervoxel to obtain local density features; For each tree hypervoxel, a preset number of feature values ​​are calculated through feature decomposition, and the geometric shape features corresponding to the tree hypervoxel are obtained based on the combination of the feature values. By integrating all the local density features with the geometric features, local features are obtained.

2. The tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest according to claim 1, characterized in that, The clustering of tree point clouds yields a set of tree supervoxels, including: Randomly select several seed points in the tree point cloud; Based on the seed point and the feature distance, the neighboring points of each seed point are clustered to obtain the corresponding number of tree hypervoxels; The center point corresponding to each tree hypervoxel was calculated; The neighborhood information corresponding to each tree hypervoxel is calculated; The tree hypervoxels and their corresponding center points are integrated with the neighborhood information to obtain the tree hypervoxel set.

3. The tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest according to claim 2, characterized in that, The neighborhood information is used to represent the adjacency relationships between tree hypervoxels. The calculation of the neighborhood information corresponding to each tree hypervoxel includes: For each boundary between adjacent tree supervoxels, a surface-based convexity detection is performed to obtain the corresponding concavity-convexity relationship features; When the concavity-convexity relationship feature is a convex connection, it indicates that the adjacent tree hypervoxels belong to the same category; When the concavity-convexity relationship feature is a concave connection, it indicates that the adjacent tree hypervoxels belong to different categories; By integrating all the aforementioned concavity-convexity relation features, neighborhood information corresponding to each tree supervoxel is obtained.

4. The tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest according to claim 1, characterized in that, The feature extraction of the tree supervoxel set yields multi-dimensional features of the tree supervoxel set, including: Calculate the distance between the center point of each tree hypervoxel and the ground in a preset direction to obtain the relative height feature; Calculate the angle between the normal of the center point of each tree hypervoxel and the horizontal plane to obtain the point cloud directional features; By integrating all the relative height features and the point cloud orientation features, a global structural feature is obtained.

5. The tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest according to claim 3, characterized in that, The feature extraction of the tree supervoxel set yields multi-dimensional features of the tree supervoxel set, including: Based on the neighborhood information, the number of neighboring supervoxels of the same type for each tree supervoxel is calculated to obtain the adjacency connection quantity feature. Based on the neighborhood information, the geometric shape features of the same type of adjacent supervoxels of each tree supervoxel are calculated to obtain the neighborhood shape features of the supervoxel. By integrating all the adjacency connection quantity features and the supervoxel neighborhood shape features, the neighborhood features are obtained.

6. The tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest according to claim 1, characterized in that, The pre-trained random forest classifier model includes several decision trees, and the training process of the pre-trained random forest classifier model includes: Using the multi-dimensional features of several tree point clouds and their corresponding tree supervoxes as a training set, and randomly sampling to generate multiple training subsets, a decision tree is constructed for each training subset, and a preset number of features are randomly selected when each decision tree splits at a node. Based on the constructed decision trees, the categories of tree hypervoxels are voted on to generate preliminary classification results; Based on the preliminary classification results, the number of decision trees is determined through cross-validation. The performance of the random forest classifier model was verified using a test set, and a pre-trained random forest classifier model with expected performance was obtained.

7. A tree point cloud branch and leaf separation system based on multidimensional hypervoxel random forest, characterized in that, The system, applied to the steps of implementing the tree point cloud branch and leaf separation method based on multidimensional hypervoxel random forest as described in any one of claims 1-6, comprises: The tree supervoxel clustering module is used to cluster tree point clouds to obtain a tree supervoxel set, wherein the tree supervoxel set includes several tree supervoxels, and each tree supervoxel includes several tree point cloud neighboring points and a center point. A multi-dimensional feature acquisition module is used to extract features from the tree hypervoxel set to obtain multi-dimensional features of the tree hypervoxel set. The initial classification module is used to input the multi-dimensional features of the tree supervoxel set into a pre-trained random forest classifier model to obtain the initial branch supervoxel set and the initial leaf supervoxel set. The classification optimization module is used to perform category distance constraint processing on the initial tree branch hypervoxel set and the initial leaf hypervoxel set to obtain the tree branch hypervoxel set and the leaf hypervoxel set, wherein the tree branch hypervoxel set includes only tree branch hypervoxels to represent trees, and the leaf hypervoxel set includes only leaf hypervoxels to represent leaves.

8. A terminal device, characterized in that, The terminal device includes a memory, a processor, and a tree point cloud branch and leaf separation program based on multidimensional supervoxel random forest stored in the memory and executable on the processor. When the processor executes the tree point cloud branch and leaf separation program based on multidimensional supervoxel random forest, it implements the steps of the tree point cloud branch and leaf separation method based on multidimensional supervoxel random forest as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a tree point cloud branch and leaf separation program based on multidimensional supervoxel random forest. When the tree point cloud branch and leaf separation program based on multidimensional supervoxel random forest is executed by the processor, it implements the steps of the tree point cloud branch and leaf separation method based on multidimensional supervoxel random forest as described in any one of claims 1-6.

Images