Tree real skeleton point extraction method and system based on deep neural network and minimum spanning tree algorithm, terminal and storage medium

Abstract

The invention belongs to the technical field of optical engineering, and discloses a tree real skeleton point extraction method and system based on a deep neural network and a minimum spanning tree algorithm, a terminal and a storage medium, and the method comprises the steps: obtaining original point cloud data; predicting the original point cloud data based on a deep neural network, and obtaining the vector, radius and category information of each point cloud; performing centralization processing on the original point cloud data according to the vector, radius and category information of each point cloud, and performing initial skeleton extraction by using a minimum spanning tree algorithm to obtain initial skeleton points; simplifying and smoothing the initial skeleton point, and updating to obtain a real skeleton point based on a real skeleton point updating method of Euclidean clustering; and outputting the real skeleton point. The tree real skeleton point extraction precision is improved, and the reconstruction precision of the tree model is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of point cloud data processing, and particularly relates to a tree real skeleton point extraction method, system, terminal and storage medium based on a deep neural network and a minimum spanning tree algorithm. BACKGROUND

[0002] In the process of tree model reconstruction, skeleton points play an irreplaceable role. Real skeleton points can provide important references for model generation on the one hand, and can also provide various important information including radius. In addition, by adjusting the density of skeleton points, the accuracy and data volume of branch models can be indirectly controlled. Therefore, how to obtain real and accurate skeleton points is extremely important for tree model reconstruction.

[0003] Most skeleton point extraction methods are based on the minimum spanning tree algorithm, which takes the original point cloud as input data and generates the final result through a series of optimization steps. Since the actual collected point cloud data is huge, it takes a lot of time to calculate the skeleton points, and the final generated result is not accurate. Especially at the branch bifurcation position, in order to achieve a smooth effect, the position of the skeleton point has to be moved, which has a serious impact on the accuracy of the generated model. Since the skeleton points generated by this method do not have radius information, the radius of the generated model is calculated based on the radius of the position of the first skeleton point (for example, the root of the tree), and the radius of the remaining skeleton points is also assigned in a step-by-step decreasing manner, so the accuracy of the tree model generated based on this method is low and the visual effect is poor.

[0004] Therefore, the prior art still needs to be improved. SUMMARY

[0005] The technical problem to be solved by the present application is that, in view of the defects of the prior art, the present application provides a tree real skeleton point extraction method, system, terminal and storage medium based on a deep neural network and a minimum spanning tree algorithm, to solve the problem of low accuracy of the tree model generated by the existing skeleton point extraction method.

[0006] The technical scheme adopted by the present application to solve the technical problem is as follows: In a first aspect, the present application provides a tree real skeleton point extraction method based on a deep neural network and a minimum spanning tree algorithm, comprising: obtaining original point cloud data; predicting the original point cloud data based on a deep neural network to obtain the vector, radius and category information of each point cloud; The original point cloud data is centralized according to the vector, radius and category information of each point cloud, and an initial skeleton is extracted by using a minimum spanning tree algorithm to obtain initial skeleton points; The initial skeleton points are simplified and smoothed, and real skeleton points are updated based on a Euclidean clustering real skeleton point updating method to obtain the real skeleton points; The real skeleton points are output.

[0007] In an implementation manner, the deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The original point cloud data is predicted based on the deep neural network to obtain the vector, radius and category information of each point cloud, which comprises: The original point cloud data is voxelized according to a preset spatial resolution to obtain sparse voxel grid data; The sparse voxel grid data is down-sampled based on the deep neural network, multi-scale features are extracted, the features are restored by step-by-step up-sampling, the restored features are connected with the same scale features after down-sampling by jumping, and a multi-layer perceptron is used to output the vector, radius and category information of each point cloud.

[0008] In an implementation manner, the original point cloud data is centralized according to the vector, radius and category information of each point cloud, and an initial skeleton is extracted by using a minimum spanning tree algorithm to obtain initial skeleton points, which comprises: All point cloud data is classified based on the category information to obtain leaf point cloud data and trunk point cloud data; The trunk point cloud data is centralized by using a geometric center alignment centralization method based on the vector and the radius to obtain centralized trunk point cloud data; The centralized trunk point cloud data is subjected to initial skeleton extraction based on the minimum spanning tree algorithm to obtain the initial skeleton points.

[0009] In an implementation manner, the centralized trunk point cloud data is subjected to initial skeleton extraction based on the minimum spanning tree algorithm to obtain the initial skeleton points, which comprises: An empty set is initialized; All point clouds in the centralized trunk point cloud data are traversed, point clouds with a Euclidean distance greater than or equal to a preset distance threshold are selected, and the selected point clouds are added to the initialized set to obtain a trunk point cloud set after removal of redundant points; The trunk point cloud set after removal of redundant points is subjected to triangulation processing to generate a triangle set; traverse each triangle in the triangle set, add an undirected edge for each pair of vertices, and assign a weight to each edge, to obtain a weighted undirected graph; Based on the weighted undirected graph, the initial skeleton extraction is performed by using the minimum spanning tree algorithm to obtain the initial skeleton points.

[0010] In an implementation manner, the simplification and smoothing processing of the initial skeleton points comprises: The skeleton points of the tree branches in the initial skeleton points that do not meet the condition are filtered by using a subtree ratio method; The error data in the filtered skeleton points are removed based on an error data removal strategy of a tree branch length rule to obtain the simplified skeleton points; The root node of the simplified skeleton points is initialized, and the root node is taken as a starting point of the skeleton; Starting from the starting point of the skeleton, the skeleton graph is traversed by using a depth-first manner until the leaf node of each path, and when a bifurcation point or a path end is found, the current path is defined as a skeleton segment to obtain a single skeleton extraction result; Each skeleton segment of the extracted single skeleton is traversed point by point, the average value of the neighborhood points in a window is calculated for each skeleton point, and the coordinates of the skeleton point are updated to a smoothed result according to the average value to obtain the smoothed initial skeleton points.

[0011] In an implementation manner, the real skeleton point updating method based on the Euclidean clustering updates the real skeleton points, comprising: The smoothed initial skeleton points are distinguished according to a preset skeleton point radius threshold to obtain skeleton points of main trunks and skeleton points of intermediate layer branches; The skeleton points of the main trunks are taken as initial skeleton points to be updated; The real skeleton points are obtained by updating the initial skeleton points to be updated by using the real skeleton point updating method based on the Euclidean clustering.

[0012] In an implementation manner, the real skeleton point updating method based on the Euclidean clustering updates the real skeleton points, comprising: The original point cloud data are Euclidean clustered according to the initial skeleton points to be updated to form a set of point cloud clusters; The centroid positions of each point cloud cluster in the set of point cloud clusters are calculated, and the centroid positions are taken as updated skeleton point positions; Each skeleton point position in the initial skeleton points to be updated is updated to the centroid position of the corresponding point cloud cluster to obtain the real skeleton points.

[0013] In a second aspect, the present application provides a tree real skeleton point extraction system based on a deep neural network and a minimum spanning tree algorithm, comprising: An acquisition module is configured to acquire original point cloud data. A prediction module is configured to predict the original point cloud data based on a deep neural network to obtain a vector, a radius and category information of each point cloud. An initial skeleton extraction module is configured to perform centering processing on the original point cloud data according to the vector, the radius and the category information of each point cloud, and perform initial skeleton extraction by using a minimum spanning tree algorithm to obtain initial skeleton points. A real skeleton updating module is configured to perform simplification and smoothing processing on the initial skeleton points, and update the real skeleton points based on a Euclidean clustering real skeleton point updating method. An output module is configured to output the real skeleton points.

[0014] In a third aspect, the present application provides a terminal, comprising a processor and a memory, wherein the memory stores a tree real skeleton point extraction program based on a deep neural network and a minimum spanning tree algorithm, and the tree real skeleton point extraction program based on the deep neural network and the minimum spanning tree algorithm is configured to implement operations of the tree real skeleton point extraction method based on the deep neural network and the minimum spanning tree algorithm as described in the first aspect when executed by the processor.

[0015] In a fourth aspect, the present application further provides a computer readable storage medium, wherein the computer readable storage medium stores a tree real skeleton point extraction program based on a deep neural network and a minimum spanning tree algorithm, and the tree real skeleton point extraction program based on the deep neural network and the minimum spanning tree algorithm is configured to implement operations of the tree real skeleton point extraction method based on the deep neural network and the minimum spanning tree algorithm as described in the first aspect when executed by a processor.

[0016] The technical scheme of the present application has the following effects: The present application predicts original point cloud data based on a deep neural network, and extracts initial skeleton points by using a minimum spanning tree algorithm, thereby effectively reducing redundant information and generating representative and structured initial skeleton point data. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only constitute some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from the structures shown in the drawings without creative labor.

[0018] Figure 1 is a flowchart of the tree real skeleton point extraction method based on deep neural network and minimum spanning tree algorithm in the present application.

[0019] Figure 2 is a general flowchart of the tree real skeleton point extraction method based on deep neural network and minimum spanning tree algorithm in the present application.

[0020] Figure 3 is a schematic diagram of part of the training data set in the present application.

[0021] Figure 4 is a schematic diagram of the framework of the deep neural network in the present application.

[0022] Figure 5 is a schematic diagram of the branch and leaf separation effect in the present application.

[0023] Figure 6 is a schematic diagram of the centralized point cloud in the present application.

[0024] Figure 7 is a schematic diagram of the main branch skeleton point extraction in the present application.

[0025] Figure 8 is a schematic diagram of the smoothed skeleton point in the present application.

[0026] Figure 9 is a schematic diagram of the real skeleton point in the present application.

[0027] Figure 10 is a functional schematic diagram of the terminal in one implementation manner of the present application.

[0028] The purposes, functional features and advantages of the present application will be further described with reference to the drawings. DETAILED DESCRIPTION

[0029] In order to make the purposes, technical solutions and advantages of the present application more clear and explicit, the following will further describe the present application in detail with reference to the drawings and examples. It should be understood that the specific examples described here are only used to explain the present application, and are not used to limit the present application.

[0030] Exemplary method The existing skeleton point extraction method is realized based on a minimum spanning tree algorithm, which takes original point cloud as input data, and generates a final result through a series of optimization steps. Since the actual collected point cloud data is huge, a large amount of time is needed to calculate the skeleton point, and the final generated result is not accurate. Especially at the branch bifurcation position, in order to achieve a smooth effect, the position of the skeleton point has to be moved, which seriously affects the accuracy of the generated model. Since the skeleton point generated by this method does not have radius information, the radius of the generated model is calculated by the radius of the position of the first skeleton point (for example, the root of the tree), and the radius of the remaining skeleton points is also assigned in a step-by-step decreasing manner, so the accuracy of the tree model generated based on this method is low and the visual effect is poor.

[0031] In order to solve the above technical problems, the embodiment of the present application provides a tree real skeleton point extraction method based on a deep neural network and a minimum spanning tree algorithm, which comprises the following steps:

[0032] As shown in Figure 1 The embodiment of the present application provides a tree real skeleton point extraction method based on a deep neural network and a minimum spanning tree algorithm, which comprises the following steps: Step S100, acquiring original point cloud data.

[0033] In this embodiment, the method mainly classifies, predicts and shrinks the original point cloud data through a deep neural network to generate a centralized original point cloud result. Then, the centralized point cloud is processed by using a minimum spanning tree algorithm to obtain an initial skeleton point, which will be used for subsequent optimization and updating steps. In order to facilitate the reconstruction of the tree model, the real skeleton point is divided into main branch skeleton points and intermediate layer skeleton points in this embodiment, which are respectively used for the reconstruction of the main branch model and the crown model.

[0034] As shown in Figure 2 The tree real skeleton point extraction process based on a deep neural network and a minimum spanning tree algorithm in this embodiment mainly comprises three steps: Firstly, the original point cloud data is predicted by deep learning to obtain the vector, radius and category information of each point cloud; Then, the original point cloud data is centralized, optimized by a minimum spanning tree algorithm, and the like, to obtain initial skeleton points; Finally, the initial skeleton points are updated to obtain real skeleton points, and are classified into main branch skeleton points and intermediate layer skeleton points.

[0035] In the embodiment, in order to enable subsequent algorithms to accurately realize skeleton prediction and realize semantic segmentation of branches and leaves in the point cloud, before the original point cloud data is predicted based on a deep neural network, a point cloud simulation data set needs to be constructed to learn geometric characteristics in space.

[0036] As an example, a tree model can be generated in batches using simulation software (for example, SpeedTree Games, software for three-dimensional tree modeling), and is sampled into a point cloud to obtain a point cloud simulation data set. The specific steps are as follows: 1) Point cloud generation: The flight path of the unmanned aerial vehicle around the tree model is simulated to collect data, to generate an RGB image (red, green, and blue three-channel image) with a resolution of about 2.1MP (that is, 1920x1080 pixels). And the three-dimensional point cloud is restored by the method of aerial triangulation, to form a virtual tree point cloud data. The generated point cloud is denoted as: ; Each point cloud contains its three-dimensional space coordinates and color features .

[0037] 2) Noise enhancement: In view of the photogrammetry noise problem, the point cloud needs to be transformed, including point inflation and erosion, and the like. And a random light change is added, to assign a random sky box to each tree. The noise distribution characteristics of the processed point cloud can be described as: ; Wherein, (.) represents a point cloud transformation operation containing the above noise simulation.

[0038] Based on the above point cloud generation steps, the distribution characteristics of the generated point cloud simulation data include: 1) Point density variation, that is, the density of points in different regions is different; 2) Missing area, for example, partial point missing caused by self-occlusion; 3) Noise, simulating the error in photogrammetry.

[0039] The finally generated point cloud simulation dataset includes 100 apple trees, 100 cherry trees, 100 eucalyptus trees, 100 ginkgo trees, 100 pine trees and 100 walnut trees. In order to predict the radius and vector of each point, the corresponding skeleton points of each tree are also generated synchronously in this embodiment. Considering the generalization of the deep neural network, different data sizes are set for different types of trees, wherein the data size of the apple tree and the cherry tree is the smallest, the data size of the ginkgo tree and the pine tree is the largest, and the data size of the eucalyptus tree and the walnut tree is relatively moderate.

[0040] As shown in Figure 3 , Figure 3 , a part of the data in the point cloud simulation dataset is displayed. Taking the number of point clouds as the basis for judgment, the data size of the apple tree is about 880,000 points, the data size of the cherry tree is about 97,000 points, the data size of the eucalyptus tree is the smallest, which is about 226,000 points, the data size of the ginkgo tree is about 4,524,000 points, the data size of the pine tree is the largest, which can reach about 5,974,000 points, and the data size of the walnut tree is about 440,000 points. The number of point clouds of the same type of tree is similar.

[0041] In order to train the deep neural network, the point cloud simulation dataset is divided into a certain proportion in this embodiment, wherein a part is used as the pre-training data of the deep neural network; another part is used as the original point cloud data, so as to obtain the vector, radius and category information of each point cloud by predicting the original point cloud data through the trained deep neural network.

[0042] As shown in Figure 1 , the embodiment of the application provides a tree real skeleton point extraction method based on a deep neural network and a minimum spanning tree algorithm, which comprises the following steps: Step S200, predicting the original point cloud data based on the deep neural network to obtain the vector, radius and category information of each point cloud.

[0043] In order to predict the geometric features of each skeleton point cloud in the original point cloud data, a deep neural network is constructed and trained in this embodiment; wherein the geometric features predicted by the deep neural network include: 1) skeleton point cloud radius , used to represent the local space scale around the point ; 2) skeleton point cloud direction , used to point to the normalized direction vector of the central axis.

[0044] Specifically, the geometric features are represented as follows: ; wherein, .

[0045] Specifically, in an implementation manner of the embodiment, before step S200, the following steps are included: Step S201a, constructing the deep neural network; wherein the deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer, and a branch and leaf segmentation head. Step S201b, obtaining a pre-training data set, training the deep neural network based on the pre-training data set, and obtaining a trained deep neural network.

[0046] In the embodiment, the deep neural network adopts U-net (an encoder-decoder architecture based on a convolutional neural network) as a backbone network, and cooperates with a sparse architecture SpConv (a spatial sparse convolution library) to improve computational efficiency. Specifically, the deep neural network comprises the following main parts: 1) Sparse convolution: sparse convolution is used in the encoder for large-scale feature exchange; and submanifold convolution is used in the decoder to avoid feature expansion and improve computational efficiency. 2) Backbone network: the U-net backbone architecture is composed of a series of encoders and decoders; the encoder down-samples the feature map (with a step of 2), and utilizes residual connection to retain details. After layer-by-layer connection of the encoder and the decoder, the encoder and the decoder are transmitted to three fully connected network branches to extract the following features of each point cloud respectively: direction , radius , and category .

[0047] 3) Activation and normalization: a ReLU (Rectified Linear Unit) activation function and batch normalization are connected after each convolution layer. 4) Branch and leaf segmentation head: a fully connected network branch is added to realize semantic segmentation of branches and leaves, and each voxel point is classified as a branch or a leaf.

[0048] In the embodiment, the deep neural network is trained based on the pre-training data set described above to obtain a trained deep neural network; in a specific training process, cross-entropy loss or a processing class imbalance loss can be used for training, and an optimizer is configured to train parameters; or, a pre-trained model is used for transfer learning, and an automatic parameter adjustment framework is used for parameter adjustment. After training is completed, the deep neural network can analyze real point cloud data and assign labels, thereby providing basic information for subsequent point cloud processing.

[0049] Specifically, in an implementation manner of the embodiment, step S200 comprises the following steps: Step S201, voxelizing the original point cloud data according to a preset spatial resolution to obtain sparse voxel grid data. Step S202, based on the deep neural network, the sparse voxel grid data is down-sampled, multi-scale features are extracted, and the restored features are connected with the same scale features after down-sampling through step-by-step up-sampling, and the multi-layer perceptron is used to output the vector, radius and category information of each point cloud respectively.

[0050] In the embodiment, voxelized point cloud data is taken as input, so voxelization processing is also needed for the original point cloud data. Specifically, the point cloud data is voxelized at a certain spatial resolution and mapped into a sparse voxel grid: ; wherein, is an index set of a three-dimensional grid, represents the features of the point cloud in the voxel , including coordinate and color information. The voxelization adopts a resolution of 1 cm, so that the point cloud features in each voxel have a high density.

[0051] In the embodiment, the prediction process of the deep neural network is as shown in Figure 4 . The deep neural network takes NxF voxel input as the starting point, gradually down-samples through the encoder to extract multi-scale features. In the encoder branch, each stage contains a residual block (ResBlock) and a down-sampling operation (Encoder), and the number of feature channels gradually increases (2, 4, 8, 16, 32, 64 in turn). In the decoder branch, the corresponding up-sampling operation (Decoder) module restores the features through step-by-step up-sampling, while performing jump connection with the same scale features in the encoder. Finally, through the multi-layer perceptron (MLP), three branches of direction, radius and category are output respectively to meet the needs of branch and leaf separation and point cloud centralization.

[0052] Specifically, for each point cloud , the deep neural network outputs the following feature information: 1) Category label : Each point cloud is assigned a category label ∈ (tree leaf, tree branch) according to the network prediction result, so as to realize semantic segmentation of the point cloud. 2) Color information : The color of the point cloud is updated and optimized to generate a more realistic color distribution. 3) Vector information : The network further outputs a direction vector = ( ) to represent the direction of each point cloud in the point cloud pointing to the center of the trunk, and = 1. 4) local radius : The spatial scale relationship of points is represented by predicting the distance of each point cloud to the trunk center (i.e. radius). In combination with the above features, the predicted point cloud set can be formally represented as: ; wherein, ( ).

[0053] In this embodiment, the vector, radius and category information of each point cloud are predicted by a deep neural network, which facilitates subsequent centering operation of the trunk point cloud according to the predicted information and helps to accurately extract the initial skeleton points.

[0054] As shown in Figure 1 , the present embodiment provides a tree real skeleton point extraction method based on a deep neural network and a minimum spanning tree algorithm, comprising the following steps: Step S300, centering processing is performed on the original point cloud data according to the vector, radius and category information of each point cloud, and the minimum spanning tree algorithm is used for initial skeleton extraction to obtain initial skeleton points.

[0055] In this embodiment, for the predicted point cloud set , the leaf point cloud needs to be separated from the trunk point cloud, and the trunk point cloud needs to be centered to more indirectly reflect the real tree skeleton, which helps to extract the initial skeleton points.

[0056] Specifically, in one implementation manner of the present embodiment, step S300 comprises the following steps: Step S301, all point cloud data is classified based on the category information to obtain leaf point cloud data and trunk point cloud data; Step S302, the trunk point cloud data is centered by using a geometric center alignment centering method based on the vector and the radius to obtain centered trunk point cloud data; Step S303, the centered trunk point cloud data is subjected to initial skeleton extraction based on the minimum spanning tree algorithm to obtain the initial skeleton points.

[0057] In this embodiment, after the deep neural network prediction is completed, first, the point cloud is classified based on the category label , so as to separate the point cloud into two parts: 1) leaf point cloud : the point cloud data satisfying = leaf; 2) trunk point cloud :satisfy =Point cloud data of tree branches.

[0058] The results of point cloud classification are as follows Figure 5 As shown. For tree trunk point clouds In this embodiment, further processing using centralized operations is required.

[0059] For the classified tree trunk point cloud This embodiment proposes a centralization method based on geometric center alignment. It assumes that the geometric relationship of each point cloud is determined by its direction vector and local radius, and that the point cloud is aligned along the predicted direction vector. movement radius The distance allows mapping of the point cloud to the center of its skeleton, achieving centralization of the trunk point cloud: ; For point clouds = ( The location of its central point is calculated using the following formula: ; The centralization process moves each predicted trunk point cloud back to a position closer to the skeleton point, forming a centralized point cloud: .

[0060] like Figure 6 As shown, Figure 6 (a) is a point cloud diagram of the tree trunk. Figure 6 (b) is a schematic diagram of the point cloud after centering the trunk portion. Figure 6 (c) is a schematic diagram comparing the point cloud before and after centralization (blue is the point cloud before centralization, and red is the point cloud after centralization). The centralized point cloud after the movement is smaller in volume than the original point cloud, and each point is closer to the skeleton point, which makes up for the algorithm failure caused by the missing part of the point cloud.

[0061] The centering operation of the tree trunk point cloud used in this embodiment is crucial because it provides standardized input for obtaining the true tree skeleton. By eliminating geometric offsets caused by variations in branch complexity or density in the point cloud, the centered point cloud can more indirectly reflect the true tree skeleton, which is helpful for subsequent steps.

[0062] In this embodiment, the point cloud after the centering operation already has a rudimentary skeleton shape, but it is still necessary to extract a reasonable skeleton point path from these point clouds. Therefore, this embodiment uses the MST algorithm (Minimum Spanning Tree algorithm) to process the result of branch point cloud centering, aiming to further simplify the data and make the skeleton points more representative.

[0063] In an implementation form of the embodiment, step S303 comprises the following steps: Step S303a, initializing an empty set; Step S303b, traversing all point clouds in the centralized trunk point cloud data, selecting point clouds with a Euclidean distance greater than or equal to a preset distance threshold, and adding the selected point clouds to the initialized set to obtain a trunk point cloud set after removing redundant points; Step S303c, performing triangulation processing on the trunk point cloud set after removing redundant points to generate a triangle set; Step S303d, traversing each triangle in the triangle set, adding an undirected edge to each pair of vertices, and assigning a weight to each edge to obtain a weighted undirected graph; Step S303e, based on the weighted undirected graph, using the minimum spanning tree algorithm to extract an initial skeleton to obtain the initial skeleton point.

[0064] As an example, the initial skeleton point extraction process based on the minimum spanning tree algorithm in the embodiment is as follows: 1) Redundant point removal: Some point clouds in the centralized point cloud data have a too small distance from other point clouds and contribute less to the final skeleton result. Therefore, the centralized point cloud set P is used in the embodiment represents a point cloud, where The Euclidean distance between the two is defined as: ; In the embodiment, a distance threshold τ = 2mm is set. If two points and in the point cloud satisfy <τ, the point is determined as a redundant point and is removed from the set. The specific process of the algorithm is as follows: First, initialize an empty set . Then, traverse all points in the centralized point cloud set P: if ∈ , both satisfy d( , ) ≥ τ, then is added to . After the traversal is completed, the filtered point set is output. The point cloud set after removing redundant points will be used as the input for subsequent calculations.

[0065] 2) Constructing a weighted undirected graph: Let the point cloud data after the redundant points are deleted be denoted as = { | i = 1, 2,..., M} where M is the number of remaining points in the point cloud. Next, the embodiment performs Delaunay triangulation (a spatial partitioning method that divides a point set into triangles or tetrahedrons) processing on to generate a weighted graph of vertices and edges. The geometric constraints of Delaunay triangulation are as follows: For any triangle , the circumcircle of the triangle does not contain other points in the point cloud set ; among all possible triangular mesh structures, the minimum angle is maximized, thereby reducing the occurrence of long and thin triangles.

[0066] By performing Delaunay triangulation on the point cloud set , a point cloud graph composed of a set of triangles can be obtained. On this basis, a weight (i.e., the Euclidean distance between the two endpoints) is assigned to each edge , and the weight is defined as follows: ; The flow of constructing the weighted graph includes the following steps: First, perform Delaunay triangulation on to generate a set of triangles T.

[0067] Then, traverse each triangle in T, and for each pair of vertices , , add an undirected edge and calculate the weight according to the above formula. Finally, the weighted undirected graph G = (V, E, w) is obtained, where V is the node set of the graph (i.e., the filtered point cloud), E is the edge set, and w is the weight function.

[0068] 3) Extracting the initial skeleton result using the MST algorithm: After the weighted undirected graph G is constructed, the initial skeleton result is further extracted using the MST algorithm. The MST is a subgraph T = ( , ) that satisfies the following properties: T is a connected graph, and there is no cycle in T; the sum of the edge weights corresponding to T is minimum. Specifically, given a weighted undirected graph G = (V, E, w), its minimum spanning tree can be constructed by the Prim algorithm (a greedy algorithm for finding the minimum spanning tree in a weighted connected graph), which is as follows: First, initialize the point set = { }, the initial point is a determined root node. Second, find the point and in the point set , so that the edge weight is minimum and satisfies . Third, add to , and add the corresponding edge , to . Finally, the above steps are executed in a loop until = V. For the graph T, the following properties are guaranteed to be true: | | = |V |, | | = |V | 1.

[0069] After the extraction process of the minimum spanning tree, the complex point cloud network can be directly simplified into a tree structure, the main skeleton connectivity is retained, the redundant connection is removed, and it is the basis for subsequent extraction of real skeleton points.

[0070] As shown in Figure 1 , the embodiment of the application provides a tree real skeleton point extraction method based on a deep neural network and a minimum spanning tree algorithm, which comprises the following steps: Step S400, simplifying and smoothing the initial skeleton points, and updating the real skeleton points based on a Euclidean clustering real skeleton point updating method.

[0071] In this embodiment, after the initial skeleton points are generated, there may still be many redundant small branches or incorrect topological structures in the skeleton. In order to obtain a more concise and more accurate skeleton structure, it is also necessary to perform simplification processing.

[0072] Specifically, in one implementation manner of the embodiment, step S400 comprises the following steps: Step S401, filtering the skeleton points of the branches in the initial skeleton points that do not satisfy the condition by using a sub-tree ratio method; Step S402, removing the error data in the filtered skeleton points based on an error data removal strategy of a branch length rule, to obtain the simplified skeleton points.

[0073] In this embodiment, a skeleton simplification method based on subtree ratio and branch length is proposed. The method mainly includes the following steps: deleting unimportant small edges according to the subtree ratio, clearing erroneous data based on the rules of the end branch length, and simplifying the skeleton as a whole.

[0074] 1) Delete unimportant small edges based on subtree ratio: This embodiment uses the subtree ratio method to remove unimportant small edges. In the initial skeleton point results, some small branches have low importance to the overall skeleton structure and weak correlation with the trunk. To reduce the interference caused by these redundant structures, this embodiment uses the subtree ratio method to filter the skeleton points of small branches that do not meet the conditions. The subtree ratio is defined as the ratio of the total length of a subtree to the total length of the parent subtree, denoted as . The specific calculation method is as follows: Let the unsimplified skeleton structure be represented as a directed weighted graph T = (V, E, w), where V is the set of nodes and E is the set of edges. This represents the edge weight (or distance) from node i to node j. For node j... its subtree length Defined as: ; in, Indicates For all nodes in the subtree rooted at the root, express arrive Euclidean distance.

[0075] Subtree ratio It can be defined as: ; in, Represents child nodes The length of the subtree, From arrive Direct side rights. When subtree ratio When the edge is less than a given threshold τ, The corresponding smaller branches are considered unimportant and are deleted. The processing procedure is as follows: First, starting from the root node, traverse the skeleton nodes and calculate the subtree length of each node. . Then, for each edge Calculate its subtree ratio . Finally, a judgment is made: if ( , If τ (take τ = 0.219), it is determined that the edge does not belong to the skeleton important part, and is deleted together with its sub-tree.

[0076] Repeat the above steps until all nodes are traversed.

[0077] 2) Error data cleaning based on tree branch length rule: In the simplified skeleton structure, there may still be significantly longer end branches, which are usually derived from noise or unreasonable data sampling. To further improve the simplification effect, the embodiment proposes an error data cleaning strategy based on the tree branch length rule.

[0078] Let the direct child node of node be , and the judgment formula based on the length of the tree branch is: ; Wherein, represents the length of the tree branch , and α is the magnification factor (α = 3 in this embodiment), is the length of the previous section of the tree branch. If the tree branch is more than α times the length of the previous section of the tree branch , it is considered that the tree branch is error data and should be deleted from the skeleton. The actual operation includes the following steps: First, traverse from the end node to the root node of the skeleton, and for each edge , , if >α· (wherein represents the length of the previous section of the edge), the tree branch is deleted. Then, ensure that the tree branch is pruned to maintain the connectivity of the skeleton. Finally, update all related attributes of the skeleton (such as edge weight, node radius), and ensure the integrity of the topological structure of the skeleton.

[0079] Simplify the minimum spanning tree, and the simplified skeleton points are obviously closer to the true situation. After excluding the influence of small and chaotic skeleton points, it is beneficial to the subsequent steps.

[0080] In this embodiment, after the initial skeleton point extraction and simplification, the complete skeleton structure is usually composed of a series of connected segments. In order to further analyze the topological structure and geometric shape of the skeleton, it is also necessary to extract a single skeleton and perform smoothing processing to improve the smoothness.

[0081] Specifically, in one implementation manner of the embodiment, step S400 further includes the following steps: Step S403: Initialize the root node of the simplified skeleton point and use the root node as the starting point of the skeleton. Step S404: Starting from the starting point of the skeleton, traverse the skeleton graph in a depth-first manner until the leaf node of each path is reached. When a branch point or the end of a path is found, define the current path as a skeleton segment to obtain the single skeleton extraction result. Step S405: For each skeleton segment of the extracted single skeleton, traverse point by point. For each skeleton point, calculate the average value of the neighboring points within the window, and update the coordinates of the skeleton point to the smoothed result based on the average value to obtain the smoothed initial skeleton point.

[0082] In this embodiment, a single skeleton extraction and smoothing algorithm is proposed. This algorithm extracts skeleton segments one by one from a complex skeleton graph and performs post-processing to improve smoothness, ensuring the geometric continuity of the final skeleton segments and their consistency with the original data. The single skeleton extraction process is as follows: Figure 7 As shown, Figure 7 In the diagram, (1), (2), (3), and (4) represent different methods for extracting skeleton path segments.

[0083] The specific steps are as follows: 1) Single skeleton extraction: A multi-branch skeleton structure typically consists of several trunks, branching points, and branches. In order to extract each coherent skeleton segment, this embodiment starts from the branching point and explores the skeleton graph by using the Depth-First Search (DFS) algorithm (an algorithm used to traverse or search tree or graph structures) to determine the connected paths between the branching point and the leaf nodes.

[0084] Definition of a bifurcation point: In a skeleton graph T = (V, E), if the out-degree (i.e., the number of edges emanating from a node v∈V) satisfies If , then it is defined as a bifurcation point. Conversely, if If v is a leaf node, then v is considered a leaf node.

[0085] The method for extracting skeleton path segments is as follows: First, initialize the root node as This serves as the starting point for the skeleton; Then, from Initially, the skeleton graph is traversed using the DFS algorithm, and for each path, the process continues until a leaf node is reached; Whenever a fork point (or the end of a path) is found, the current path is defined as a skeleton segment and stored in the extraction results. Let the root node The first path is p = { , , , . . . ,}, the path extraction algorithm follows the following constraints: If the node (i>0) satisfies: or is a leaf node, then the path { , , ,..., } is intercepted and stored as a skeleton segment.

[0086] 2) Smooth each individual skeleton segment: During the skeleton segment extraction process, irregular small segments are generated due to the depth-first search, and the geometric distribution of the original points often contains large discreteness, which may lead to non-smooth skeleton points or geometric discontinuity. Therefore, the embodiment performs smoothing on each individual skeleton segment to enhance its geometric continuity.

[0087] In this embodiment, the window average smoothing method is used to smooth the point set in each skeleton segment. Let the skeleton segment P = { , ,..., }, and each skeleton point is a three-dimensional coordinate point ( , , ). The smoothed coordinate of the weighted average function of the sliding window is defined as: ; where: W is the size of the sliding window (W = 6 in this embodiment) ; represents the effective points in the window; to avoid missing data at the window boundary, the boundary is filled using the continuation method. The specific algorithm flow is as follows: First, each skeleton segment is traversed point by point. Then, for each skeleton point , the average value of the neighborhood points in the window is calculated.

[0088] Finally, the coordinates of the skeleton point are updated to the smoothed result .

[0089] The advantage of the window smoothing method is that the algorithm is simple and efficient, and is suitable for large-scale skeleton data processing. To verify the smoothing effect, the smoothed skeleton points are compared with the original point cloud set in this embodiment. All points of the smoothed skeleton segment satisfy: ; where, is the maximum allowable error distance is 0.001, represents the Euclidean distance.

[0090] As Figure 8 shown, the gray points are the points before smoothing, and the blue points are the points after smoothing. This algorithm can effectively extract each skeleton segment and post-process it to enhance the smoothness, providing higher quality data support for subsequent processing of the skeleton (such as modeling, analysis, etc.).

[0091] In this embodiment, after smoothing, the positions of some of the initial skeleton points are deep into the tree canopy, so the predicted results have larger deviations. Therefore, these skeleton points also need to be distinguished and updated.

[0092] Specifically, in an implementation manner of this embodiment, step S400 further includes the following steps: Step S406, according to a preset skeleton point radius threshold, distinguishing the smoothed initial skeleton points to obtain skeleton points of main trunks and skeleton points of intermediate layer branches; Step S407, taking the skeleton points of the main trunks as initial skeleton points to be updated; Step S408, updating the initial skeleton points to be updated based on a real skeleton point updating method based on Euclidean clustering to obtain the real skeleton points.

[0093] In this embodiment, a threshold rthres = 5 cm of the radius of the skeleton point is set to distinguish the skeleton points of the main branches and the skeleton points of the intermediate layer branches. The screening criteria are defined as follows: ; wherein, represents the position and radius of the i-th skeleton point.

[0094] After screening, in this embodiment, the skeleton points of the main branches are obtained and taken as the initial skeleton points to be updated for subsequent calculation. Due to the need for complex mathematical operations, the predicted results of these initial skeleton points to be updated may have certain errors. To ensure the accuracy of the skeleton points, in this embodiment, a real skeleton point updating method based on clustering is proposed. As Figure 9 shown, this method combines the initial skeleton points to be updated and the geometric distribution characteristics of the original point cloud, and obtains a skeleton point set with higher accuracy by cutting the original point cloud and recalculating the position of the skeleton point based on the centroid.

[0095] In an implementation manner of this embodiment, step S408 includes the following steps: Step S408a, according to the initial skeleton points to be updated, performing Euclidean clustering on the original point cloud data to form a set of point cloud clusters; Step S408b, calculate the centroid position of each point cloud cluster in the set of point cloud clusters, and take the centroid position as the updated skeleton point position; Step S408c, update the position of each skeleton point in the initial skeleton points to be updated to the centroid position of the corresponding point cloud cluster to obtain the real skeleton points.

[0096] In this embodiment, the specific steps of the clustering-based real skeleton point updating method are as follows: 1) Perform Euclidean clustering on the original point cloud data according to the initial skeleton points to be updated: Given the initial skeleton point set N = { | i = 1, 2,..., m} to be updated, where represents the i-th initial skeleton point to be updated, and the original point cloud is: P = { | j = 1, 2,..., k}, where represents the j-th point in the point cloud. In order to perform Euclidean clustering on the point cloud data using the initial skeleton points to be updated, the midpoint distance and direction vector of each pair of adjacent skeleton points , are defined as the basis for judgment.

[0097] The midpoint position is: , and the length of the midpoint vector is .

[0098] Then, for each point of the original point cloud, calculate its distance from two skeleton points and , and determine whether it is within the range of less than distance in the direction of the vector. If this distance condition is met, the point is assigned to the skeleton point cluster closest to it. This operation is essentially a segmentation of the point cloud P according to the geometric constraints defined by the skeleton points, thereby forming a set of point cloud clusters: ; where each point cloud cluster is composed of a set of points assigned to the skeleton point . .

[0099] 2) Calculate the centroid of the point cloud cluster: For each point cloud cluster ∈ , calculate its centroid position as the updated skeleton point position. The centroid is calculated according to the following formula: ; wherein, representing the number of point cloud clusters representing the number of point cloud clusters

[0100] 3) Update the initial skeleton point positions: updating each skeleton point position in the initial skeleton point set to be updated to the centroid position of its corresponding point cloud cluster , that is: ; After the centroid calculation of all point cloud clusters is completed, an updated skeleton point set is generated: ; The clustering-based real skeleton point updating method effectively improves the accuracy of the skeleton nodes and reduces errors caused by point cloud noise.

[0101] As shown in Figure 1 , the embodiment of the present application provides a tree real skeleton point extraction method based on a deep neural network and a minimum spanning tree algorithm, comprising the following steps: Step S500, outputting the real skeleton point.

[0102] In this embodiment, the output real skeleton point is a skeleton point that has been classified, which contains main branch skeleton points and intermediate layer skeleton points.

[0103] To sum up, the embodiment realizes the goals of redundant point removal, Delaunay triangulation, and minimum spanning tree extraction of the centralized point cloud through the MST-based optimization algorithm, effectively reduces redundant information, and generates representative and structured skeleton point initial data; and by combining point cloud geometric information and initial skeleton points, the original point cloud is subjected to Euclidean clustering, and the initial skeleton point position is updated according to the centroid position of the point cloud cluster, and finally the real skeleton point is obtained. The entire process improves the accuracy of skeleton point prediction and reduces errors caused by point cloud noise and complex geometric structures. The finally generated skeleton point can more accurately reflect the geometric structure of the tree, and provides a high-precision skeleton data basis for tree model reconstruction.

[0104] The embodiment achieves the following technical effects through the above technical solutions: The embodiment is based on a deep neural network to predict original point cloud data, and an initial skeleton is extracted through a minimum spanning tree algorithm, which effectively reduces redundant information and generates initial skeleton point data with strong representation and structure; and through simplification and smoothing processing of the initial skeleton point data, real skeleton points can be updated based on a Euclidean clustering real skeleton point updating method, errors caused by point cloud noise and complex geometric structures are reduced, and the finally generated skeleton points can more accurately reflect the geometric structure of the tree; the embodiment improves the extraction accuracy of real skeleton points of the tree and improves the reconstruction accuracy of the tree model.

[0105] Exemplary device Based on the above embodiment, the application further provides a tree real skeleton point extraction system based on a deep neural network and a minimum spanning tree algorithm, comprising: An acquisition module is configured to acquire original point cloud data. A prediction module is configured to predict the original point cloud data based on a deep neural network, and acquire vector, radius and category information of each point cloud. An initial skeleton extraction module is configured to perform centering processing on the original point cloud data according to the vector, radius and category information of each point cloud, and perform initial skeleton extraction by using a minimum spanning tree algorithm to obtain initial skeleton points. A real skeleton updating module is configured to perform simplification and smoothing processing on the initial skeleton points, and update real skeleton points based on a Euclidean clustering real skeleton point updating method. An output module is configured to output the real skeleton points.

[0106] The above technical solution achieves the following technical effects: The embodiment is based on a deep neural network to predict original point cloud data, and an initial skeleton is extracted through a minimum spanning tree algorithm, which effectively reduces redundant information and generates initial skeleton point data with strong representation and structure; and through simplification and smoothing processing of the initial skeleton point data, real skeleton points can be updated based on a Euclidean clustering real skeleton point updating method, errors caused by point cloud noise and complex geometric structures are reduced, and the finally generated skeleton points can more accurately reflect the geometric structure of the tree; the embodiment improves the extraction accuracy of real skeleton points of the tree and improves the reconstruction accuracy of the tree model.

[0107] Based on the above embodiment, the application further provides a terminal, and a principle block diagram thereof can be as shown in Figure 10 .

[0108] The terminal comprises a processor, a memory, an interface, a display screen and a communication module connected by a system bus; the processor of the terminal is configured to provide computing and control capabilities; the memory of the terminal comprises a computer readable storage medium and an internal memory; the computer readable storage medium stores an operating system and a computer program; the internal memory provides an environment for the operating system and the computer program in the computer readable storage medium to run; the interface is configured to connect external devices; the display screen is configured to display corresponding information; and the communication module is configured to communicate with a cloud server or other devices.

[0109] The computer program is configured to implement the operation of the tree real skeleton point extraction method based on the deep neural network and the minimum spanning tree algorithm when executed by the processor.

[0110] Those skilled in the art can understand that, Figure 10 The principle block diagram shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the terminal to which the scheme of the present application is applied; specifically, the terminal can comprise more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement.

[0111] In one embodiment, a terminal is provided, comprising a processor and a memory, wherein the memory stores a tree real skeleton point extraction program based on a deep neural network and a minimum spanning tree algorithm, and the tree real skeleton point extraction program based on the deep neural network and the minimum spanning tree algorithm is configured to implement the operation of the tree real skeleton point extraction method based on the deep neural network and the minimum spanning tree algorithm as above when executed by the processor.

[0112] In one embodiment, a computer readable storage medium is provided, wherein the computer readable storage medium stores a tree real skeleton point extraction program based on a deep neural network and a minimum spanning tree algorithm, and the tree real skeleton point extraction program based on the deep neural network and the minimum spanning tree algorithm is configured to implement the operation of the tree real skeleton point extraction method based on the deep neural network and the minimum spanning tree algorithm as above when executed by the processor.

[0113] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the computer program can be stored in a non-volatile storage medium, and when executed, can include the processes of the above-mentioned embodiments. Any reference to memory, storage, database or other medium used in each embodiment of the present application can include non-volatile and volatile memory.

[0114] In conclusion, the application provides a tree real skeleton point extraction method, system, terminal and storage medium based on a deep neural network and a minimum spanning tree algorithm, comprising: obtaining original point cloud data; predicting the original point cloud data based on a deep neural network to obtain the vector, radius and category information of each point cloud; performing centralization processing on the original point cloud data according to the vector, radius and category information of each point cloud, and performing initial skeleton extraction by using a minimum spanning tree algorithm to obtain initial skeleton points; performing simplification and smoothing processing on the initial skeleton points, and updating the real skeleton points based on a Euclidean clustering real skeleton point updating method; and outputting the real skeleton points. The application improves the tree real skeleton point extraction precision and improves the reconstruction precision of the tree model.

[0115] It should be understood that the application is not limited to the above examples, and can be improved or changed according to the above description for those of ordinary skill in the art, and all these improvements and changes shall belong to the protection scope of the appended claims of the application.

Claims

1. A method for extracting real skeleton points of a tree based on a deep neural network and a minimum spanning tree algorithm, characterized in that, The method comprises the following steps: acquiring original point cloud data; predicting the original point cloud data based on a deep neural network to obtain vector, radius and category information of each point cloud; centralizing the original point cloud data according to the vector, radius and category information of each point cloud, and extracting an initial skeleton by using a minimum spanning tree algorithm to obtain initial skeleton points; simplifying and smoothing the initial skeleton points, and updating the real skeleton points based on a real skeleton point updating method based on Euclidean clustering; outputting the real skeleton points.

2. The method of claim 1, wherein the method is based on a deep neural network and a minimum spanning tree algorithm. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head.

3. The method of claim 1, wherein the method further comprises: The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head.

4. The method of claim 3, wherein the method further comprises: The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head.

5. The method of claim 1, wherein the method further comprises: The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. The deep neural network comprises a sparse convolution layer, a backbone network layer, an activation and normalization layer and a branch and leaf segmentation head. Starting from the starting point of the skeleton, the skeleton graph is traversed in a depth-first manner until the leaf nodes of each path, and when a branching point or a path end is found, the current path is defined as a skeleton segment to obtain a single skeleton extraction result; Each skeleton segment of the extracted single skeleton is traversed point by point, and for each skeleton point, the average value of the neighborhood points in the window is calculated, and the coordinates of the skeleton point are updated to the smoothed result according to the average value to obtain the smoothed initial skeleton point.

6. The method of claim 5, wherein the method further comprises: The real skeleton point updating method based on Euclidean clustering updates the real skeleton point, including: According to the preset skeleton point radius threshold, the smoothed initial skeleton point is distinguished to obtain the skeleton points of the main trunk and the skeleton points of the intermediate layer branches; The skeleton points of the main trunk are taken as the initial skeleton points to be updated; The real skeleton point updating method based on Euclidean clustering updates the initial skeleton points to be updated to obtain the real skeleton point.

7. The method of claim 6, wherein the method further comprises: The real skeleton point updating method based on Euclidean clustering updates the initial skeleton points to be updated to obtain the real skeleton point, including: According to the initial skeleton points to be updated, the original point cloud data is Euclidean clustered to form a set of point cloud clusters; The centroid positions of each point cloud cluster in the set of point cloud clusters are calculated, and the centroid positions are taken as the updated skeleton point positions; Each skeleton point position in the initial skeleton points to be updated is updated to the centroid position of the corresponding point cloud cluster to obtain the real skeleton point.

8. A tree real skeleton point extraction system based on a deep neural network and a minimum spanning tree algorithm, characterized in that, It includes: An acquisition module is configured to acquire original point cloud data; A prediction module is configured to predict the original point cloud data based on a deep neural network to obtain vector, radius and category information of each point cloud; An initial skeleton extraction module is configured to perform centering processing on the original point cloud data according to the vector, radius and category information of each point cloud, and perform initial skeleton extraction by using a minimum spanning tree algorithm to obtain initial skeleton points; A real skeleton updating module is configured to simplify and smooth the initial skeleton points, and update the real skeleton points based on a real skeleton point updating method based on Euclidean clustering; An output module is configured to output the real skeleton points.

9. A terminal, characterized by comprising: It includes: A processor and a memory, wherein the memory stores a tree real skeleton point extraction program based on a deep neural network and a minimum spanning tree algorithm, and the tree real skeleton point extraction program based on the deep neural network and the minimum spanning tree algorithm is used to implement the operations of the tree real skeleton point extraction method based on the deep neural network and the minimum spanning tree algorithm in any one of claims 1-7 when executed by the processor.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores a tree real skeleton point extraction program based on a deep neural network and a minimum spanning tree algorithm, and the tree real skeleton point extraction program based on the deep neural network and the minimum spanning tree algorithm is used to implement the operations of the tree real skeleton point extraction method based on the deep neural network and the minimum spanning tree algorithm when executed by the processor.

Images