Three-dimensional point cloud rapid sampling method and device based on graph structure

Abstract

The invention provides a three-dimensional point cloud rapid sampling method and device based on a graph structure, and relates to the technical field of computer simulation and method optimization. Comprising the following steps: generating a weighted adjacency matrix of three-dimensional point cloud data by using a K-nearest neighbor algorithm; segmenting the three-dimensional point cloud data and the weighted adjacency matrix to obtain a plurality of sub-blocks and a plurality of sub-adjacency matrixes, and determining a sub-graph and a graph Laplacian matrix corresponding to each sub-block according to the plurality of sub-blocks and the plurality of sub-adjacency matrixes; according to the characteristic value of the graph Laplacian matrix, a preset frequency spectrum truncation threshold and the pulse signal of the node in each sub-graph, determining local energy of the filter and the node in each sub-graph by using a Chebyshev approximation algorithm; and according to the filter, the local energy and a preset regularization parameter, sampling a plurality of target nodes in the nodes of each sub-graph, and determining a sampling signal set of the three-dimensional point cloud data according to the plurality of target nodes. When large-scale point cloud data is processed, the calculation complexity is reduced, and the sampling efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer simulation and method optimization, and in particular to a three-dimensional point cloud fast sampling method and device based on graph structure. BACKGROUND

[0002] Three-dimensional point cloud sampling technology reduces redundant points in point cloud data, realizes data compression while preserving key geometric information, and is widely used in automatic driving, virtual reality, industrial detection and other fields. Three-dimensional point cloud data is usually collected by devices such as laser radar or depth camera. Since the amount of three-dimensional point cloud data collected by the device is large and redundant, directly processing the collected three-dimensional point cloud data will bring huge computing and storage overhead. Therefore, a fast sampling method is one of the key technologies for point cloud processing.

[0003] At present, three-dimensional point cloud sampling methods include curvature-based sampling methods and deep learning-based sampling methods. The curvature-based sampling method preferentially retains points with large curvature, which can better maintain geometric features. However, when processing large-scale point cloud data, the computational complexity is high. The deep learning-based sampling method learns the internal structure of the point cloud to realize adaptive sampling, but the model training corresponding to the deep learning-based sampling method requires a large amount of labeled data, making the sampling efficiency of large-scale three-dimensional point cloud data low. SUMMARY

[0004] The purpose of the embodiments of the present application is to provide a three-dimensional point cloud fast sampling method and device based on graph structure, which solves the problems of high computational complexity and low sampling efficiency in processing large-scale point cloud data in the prior art.

[0005] To solve the above technical problems, the embodiments of the present application provide the following technical solutions: The first aspect of the present application provides a three-dimensional point cloud fast sampling method based on graph structure, comprising: Generating a weighted adjacency matrix of three-dimensional point cloud data using a K-nearest neighbor algorithm; According to the preset number, the three-dimensional point cloud data and the weighted adjacency matrix are respectively divided to obtain a plurality of sub-blocks and a plurality of corresponding sub-adjacency matrices, and the corresponding sub-graphs of each sub-block and the graph Laplacian matrices corresponding to each sub-graph are determined according to the plurality of sub-blocks and the plurality of sub-adjacency matrices; According to the eigenvalues corresponding to the graph Laplacian matrix, the preset spectral truncation threshold and the impulse signals of the nodes in each sub-graph, a filter and the local energy of the nodes in each sub-graph are determined using a Chebyshev approximation algorithm; According to the filter, the local energy and the preset regularization parameter, a plurality of target nodes are sampled in the nodes of each sub-graph, and a sampling signal set corresponding to the three-dimensional point cloud data is determined according to the plurality of target nodes.

[0006] The second aspect of the present application provides a device for fast sampling of three-dimensional point cloud based on graph structure, comprising: A generating module is configured to generate a weighted adjacency matrix of the three-dimensional point cloud data by using a K-neighbor algorithm. A dividing module is configured to divide the three-dimensional point cloud data and the weighted adjacency matrix respectively according to a preset number, to obtain a plurality of sub-blocks and a plurality of sub-adjacency matrices respectively, and to determine a sub-graph corresponding to each sub-block and a graph Laplacian matrix corresponding to each sub-graph according to the plurality of sub-blocks and the plurality of sub-adjacency matrices. A determining module is configured to determine a filter and a local energy of a node in each sub-graph by using a Chebyshev approximation algorithm according to an eigenvalue corresponding to the graph Laplacian matrix, a preset spectral truncation threshold, and an impulse signal of the node in each sub-graph. A sampling module is configured to sample a plurality of target nodes in the node of each sub-graph according to the filter, the local energy, and a preset regularization parameter, and to determine a sampling signal set corresponding to the three-dimensional point cloud data according to the plurality of target nodes.

[0007] Compared with the prior art, the device for fast sampling of three-dimensional point cloud based on graph structure provided by the present application generates a weighted adjacency matrix of the three-dimensional point cloud data by using a K-neighbor algorithm, divides the three-dimensional point cloud data and the weighted adjacency matrix respectively according to a preset number, to obtain a plurality of sub-blocks and a plurality of sub-adjacency matrices respectively, and determines a sub-graph corresponding to each sub-block and a graph Laplacian matrix corresponding to each sub-graph according to the plurality of sub-blocks and the plurality of sub-adjacency matrices. The filter and the local energy of the node in each sub-graph are determined by using a Chebyshev approximation algorithm according to an eigenvalue corresponding to the graph Laplacian matrix, a preset spectral truncation threshold, and an impulse signal of the node in each sub-graph. A plurality of target nodes are sampled in the node of each sub-graph according to the filter, the local energy, and a preset regularization parameter, and a sampling signal set corresponding to the three-dimensional point cloud data is determined according to the plurality of target nodes. In this way, the three-dimensional point cloud data can be divided into a plurality of sub-blocks, the weighted adjacency matrix can be divided into a plurality of sub-adjacency matrices, and a plurality of sub-graphs corresponding to the plurality of sub-blocks can be determined according to the plurality of sub-blocks and the plurality of sub-adjacency matrices after division. Each sub-graph is sampled, and the sampling signal set can be obtained through the plurality of target nodes sampled by each sub-graph, so that the computational complexity is low when processing large-scale point cloud data. A plurality of target nodes are sampled in the node of each sub-graph according to the filter, the local energy, and a preset regularization parameter, so that the local energy can be maximized and the correlation interference can be minimized, and thus the sampling efficiency is high. BRIEF DESCRIPTION OF DRAWINGS

[0008] The above and other objects, features and advantages of the present application will become more apparent from the following detailed description when taken in conjunction with the accompanying drawings in which a few embodiments of the present application are shown by way of illustration. In the drawings: Figure 1 A flow chart of the method for quickly sampling a three-dimensional point cloud based on a graph structure is schematically shown; Figure 2 A structural diagram of the device for quickly sampling a three-dimensional point cloud based on a graph structure is schematically shown. DETAILED DESCRIPTION

[0009] Exemplary embodiments of the present application will be described in detail with reference to the drawings, wherein the same or similar components are denoted by the same or similar reference numerals, and wherein:

[0010] It should be noted that the technical terms or scientific terms used in the present application should be understood as their common meanings understood by those skilled in the art, unless otherwise specified.

[0011] The method in the embodiments of the present application will be described in detail below.

[0012] Figure 1 A flow chart of the method for quickly sampling a three-dimensional point cloud based on a graph structure in the embodiments of the present application is schematically shown, referring to Figure 1 As shown in the figure, the method for quickly sampling a three-dimensional point cloud based on a graph structure can include: S101, generating a weighted adjacency matrix of the three-dimensional point cloud data by using a K-nearest neighbor algorithm.

[0013] Specifically, the weighted adjacency matrix of the three-dimensional point cloud data is generated by using a K-nearest neighbor algorithm (KNN), including: Step A1: constructing an adjacency graph of the three-dimensional point cloud data by using the K-nearest neighbor algorithm.

[0014] Each node in the adjacency graph corresponds to each point in the three-dimensional point cloud data, and each point corresponds to K nearest neighbor points.

[0015] The adjacency graph is an adjacency graph based on Euclidean distance.

[0016] Step A2: calculating the edge weight between each point and the K nearest neighbor points according to each point and the K nearest neighbor points.

[0017] The expression of the edge weight is: ; wherein, is the point and the point corresponding to the edge weight between the adjacent points, is the coordinate vector of the point is the coordinate vector of the point corresponding to the coordinate vector of the adjacent point, is the coordinate vector of the point corresponding to the coordinate vector of the adjacent point, is the coordinate vector of the point is the scaling parameter for controlling the distance influence, is the coordinate vector of the point corresponding to the coordinate vector of the adjacent point corresponding to the coordinate vector of the adjacent point, is the Euclidean distance between the coordinate vector of the point is the square of the two-norm.

[0018] Step A3: generating a weighted adjacency matrix according to the edge weight.

[0019] The non-zero elements in the weighted adjacency matrix correspond to the edge weight.

[0020] S102, according to the preset number, the three-dimensional point cloud data and the weighted adjacency matrix are respectively segmented, a plurality of sub-blocks and a plurality of corresponding sub-adjacency matrices are obtained, and according to the plurality of sub-blocks and the plurality of sub-adjacency matrices, the sub-graph corresponding to each sub-block and the graph Laplacian matrix corresponding to each sub-graph are determined.

[0021] Specifically, according to the preset number, the three-dimensional point cloud data and the weighted adjacency matrix are respectively segmented, a plurality of sub-blocks and a plurality of corresponding sub-adjacency matrices are obtained, and according to the plurality of sub-blocks and the plurality of sub-adjacency matrices, the sub-graph corresponding to each sub-block and the graph Laplacian matrix corresponding to each sub-graph are determined, including: Step B1: according to the preset number, the three-dimensional point cloud data is segmented into a plurality of sub-blocks.

[0022] Wherein, the three-dimensional point cloud data corresponding to each sub-block is different, and each sub-block has a sub-block index.

[0023] The preset number can be The three-dimensional point cloud data is segmented into sub-blocks. The sub-blocks are mutually non-overlapping sub-blocks, The sub-blocks are represented as , wherein, is the first sub-block, is the sub-block,For the first sub-block, the three-dimensional point cloud data satisfies , i.e. the union of all sub-blocks is equal to the three-dimensional point cloud data. The size of the first sub-block is represented as , wherein is the cardinality of the set , i.e. the number of points in the point cloud sub-block, is the size of the first sub-block. The sub-blocks of the partition are displayed by green small boxes.

[0024] Step B2: according to the sub-block index, the weighted adjacency matrix is partitioned to obtain a sub-adjacency matrix corresponding to each sub-block.

[0025] The sub-block index refers to , and the weighted adjacency matrix is partitioned into sub-adjacency matrices , , wherein is the first sub-adjacency matrix, is the sub-adjacency matrix corresponding to the first sub-block .

[0026] The number of all sub-adjacency matrices is the same as the number of all sub-blocks.

[0027] Step B3: taking the point set of each sub-block as a vertex set, and constructing a sub-graph corresponding to each sub-block according to the vertex set and the sub-adjacency matrix.

[0028] Exemplarily, the sub-graph corresponding to the first sub-block is represented as , wherein the vertex set of the sub-graph corresponds to the point set of the first sub-block . The sub-graph is used to express the local structure relationship of each point in the corresponding sub-block.

[0029] Step B4: according to the elements in the sub-adjacency matrix, a corresponding degree matrix is calculated.

[0030] The degree matrix corresponding to the sub-adjacency matrix is a diagonal matrix.

[0031] Step B5: according to the degree matrix and the sub-adjacency matrix, a graph Laplacian matrix corresponding to each sub-graph is determined.

[0032] Exemplarily, the first sub-adjacency matrix The corresponding degree matrix is represented as , subgraph The corresponding graph Laplacian matrix .

[0033] S103, according to the eigenvalue corresponding to the graph Laplacian matrix, the preset spectral truncation threshold and the pulse signal of the node in each subgraph, the Chebyshev approximation algorithm is used to determine the filter and the local energy of the node in each subgraph.

[0034] Wherein, the node of each subgraph includes selected node and unselected node. The pulse signal of the node in each subgraph includes the pulse signal of the selected node and the pulse signal of the unselected node. The local energy includes the first local energy of the selected node and the second local energy of the unselected node.

[0035] Specifically, according to the eigenvalue corresponding to the graph Laplacian matrix, the preset spectral truncation threshold and the pulse signal of the node in each subgraph, the Chebyshev approximation algorithm is used to determine the filter and the local energy of the node in each subgraph, including: Step C1: according to the eigenvalue corresponding to the graph Laplacian matrix and the preset spectral truncation threshold, the kernel function is determined.

[0036] The expression of the kernel function is: ; Wherein, is the kernel function of the eigenvalue corresponding to the graph Laplacian matrix, is the eigenvalue corresponding to the graph Laplacian matrix, is the preset spectral truncation threshold.

[0037] The kernel function is usually a monotonically decreasing function, which realizes the reservation of the low frequency component of the subgraph signal by low pass filtering, and realizes the gradual suppression of the high frequency component of the subgraph signal.

[0038] Step C2: the kernel function is processed by using the Chebyshev approximation algorithm to obtain the filter.

[0039] The operation of the kernel function is converted into the cascade of function variables by the Chebyshev approximation algorithm, and the random high spectral feature matrix is generated. The filter is obtained by transforming the random high spectral feature matrix by the Chebyshev approximation algorithm.

[0040] Step C3: according to the filter and the pulse signal of the node in each subgraph, the local energy of the node in each subgraph is determined.

[0041] The expression of the local energy of the node in each subgraph is: ; Wherein, is the local energy of the node in each subgraph The local energy of the nodes in each subgraph The square of the second norm of the localized filter response, For each node in the subgraph The localized filter response of is the filter corresponding to the kernel function, For each node in the subgraph The local energy.

[0042] S104 , sampling multiple target nodes from the nodes of each subgraph according to the filter, local energy, and preset regularization parameters, and determining a sampling signal set corresponding to the three-dimensional point cloud data according to the multiple target nodes.

[0043] The sampling object of the present invention is a large-scale three-dimensional point cloud data. The sampling goal is to effectively sample multiple target nodes from the nodes of each subgraph, reconstruct the sampling signal set determined by the multiple target nodes, and reconstruct the color attributes of the three-dimensional point cloud data.

[0044] Specifically, based on the filter, local energy, and preset regularization parameters, multiple target nodes are sampled from the nodes of each subgraph, and a sampling signal set corresponding to the three-dimensional point cloud data is determined based on the multiple target nodes, including: Step D1: Determine a first localized filter response of the selected node according to the filter and the pulse signal of the selected node.

[0045] The expression for the first localized filter response of the selected node is: ; in, For selected nodes The first localized filter response, is the filter, For selected nodes pulse signal.

[0046] The first node selected among the selected nodes is the node with the largest energy norm.

[0047] Step D2: Determine a second localized filter response of the unselected node based on the filter and the pulse signal of the unselected node.

[0048] ; in, Unselected nodes The second localized filter response, is the filter, Unselected nodes pulse signal.

[0049] Step D3: determining a correlation square weighted accumulation and an energy coupling accumulation between the selected nodes and the unselected nodes according to the first localized filter response, the second localized filter response, the first local energy, and a preset regularization parameter.

[0050] Specifically, an expression of the correlation square weighted accumulation is: ; wherein, is the correlation square weighted accumulation between the selected nodes and the unselected nodes , is the first localized filter response of the selected nodes , is the transpose of the first localized filter response of the selected nodes , is the second localized filter response of the unselected nodes , is the first local energy of the selected nodes , i.e., the square of the two-norm of the first localized filter response of the selected nodes , is the preset regularization parameter, is a sampling signal set corresponding to the three-dimensional point cloud data. The preset regularization parameter is used to enhance numerical stability.

[0051] An expression of the energy coupling accumulation is: ; wherein, is the energy coupling accumulation between the selected nodes and the unselected nodes , is the first localized filter response of the selected nodes , is the transpose of the first localized filter response of the selected nodes , is the second localized filter response of the unselected nodes , is the first local energy of the selected nodes , is the preset regularization parameter, is a sampling signal set corresponding to the three-dimensional point cloud data.

[0052] Step D4: sampling a plurality of target nodes in the nodes of each subgraph according to the correlation square weighted accumulation, the energy coupling accumulation, the second local energy, and the preset regularization parameter, to obtain all target nodes of all subgraphs.

[0053] The multiple target nodes are nodes in the selected nodes.

[0054] Sample multiple target nodes corresponding to each subgraph in the nodes of each subgraph, and determine the set of multiple target nodes corresponding to each subgraph as the local sampling point set .

[0055] Specifically, the expression of multiple target nodes is: ; in, For multiple target nodes, Unselected nodes The second localized filter response, Unselected nodes The second local energy of the unselected nodes The square of the second norm of the second localized filter response, To preset the regularization parameter, For selected nodes and unselected nodes The weighted cumulative squared correlation between For selected nodes and unselected nodes The accumulated energy coupling between The unselected sampling signal set consists of unselected nodes.

[0056] The algorithm corresponding to the expression of multiple target nodes in step D4 can be called a greedy algorithm. Through this greedy algorithm, the remaining independent energy can be maximized and the correlation interference can be minimized. The maximization of the remaining independent energy corresponds to the numerator of the expression of multiple target nodes, that is, , minimizing the correlation interference corresponds to the denominator of the expression of multiple target nodes, namely .

[0057] Step D5: Merge all target nodes of all subgraphs to obtain a sampling signal set corresponding to the three-dimensional point cloud data.

[0058] The operations from steps D1 to D5 above can be called Kernel-based Sampling using Localized Graph Filter impulse responses (KSLF). The core of the KSLF algorithm is to optimize nodes using filter responses to select multiple target nodes.

[0059] At the time of sampling, a plurality of target nodes are sampled in the nodes of each subgraph, which can meet the global sampling rate limit and retain important information as much as possible while significantly reducing the computational overhead.

[0060] In order to evaluate the effect of sampling and the representativeness of the sampling point cloud in the sampled signal set, the sampling point cloud in the sampled signal set needs to be reconstructed.

[0061] The effect of point cloud reconstruction is visualized by reconstruction of the RGB color signal, and after step S104, the graph structure based three-dimensional point cloud fast sampling method further comprises: reconstructing the sampled signal set by using the expression of graph Laplacian regularization to obtain a target reconstructed point cloud signal.

[0062] The target reconstructed point cloud signal is used for high-quality modeling, improving the accuracy of shape-based object recognition and tracking algorithms, etc.

[0063] Specifically, the sampled signal set is reconstructed by using the expression of graph Laplacian regularization to obtain the target reconstructed point cloud signal, comprising: Step E1: calculating the sampled signal set by using the expression of graph Laplacian regularization to obtain a reconstructed signal set.

[0064] The expression of the reconstructed signal set is: , wherein, is the reconstructed signal set, is the sampled signal set corresponding to the three-dimensional point cloud data, is the Laplacian matrix between unsampled nodes in the sampled signal set, is the Laplacian matrix between unsampled nodes and sampled nodes in the sampled signal set.

[0065] Step E2: merging the reconstructed signal set and the sampled signal set to obtain the target reconstructed point cloud signal.

[0066] Based on the above Figure 1As can be seen from the implementation mode, the embodiment of the present application utilizes the K-neighbor algorithm to generate a weighted adjacency matrix of the three-dimensional point cloud data; the three-dimensional point cloud data and the weighted adjacency matrix are respectively segmented according to a preset number, a plurality of sub-blocks and a plurality of corresponding sub-adjacency matrices are obtained, and a sub-graph corresponding to each sub-block and a graph Laplacian matrix corresponding to each sub-graph are determined according to the plurality of sub-blocks and the plurality of sub-adjacency matrices; a filter and a local energy of a node in each sub-graph are determined by using the Chebyshev approximation algorithm according to an eigenvalue corresponding to the graph Laplacian matrix, a preset spectral truncation threshold and a pulse signal of the node in each sub-graph; and a plurality of target nodes are sampled in the node of each sub-graph according to the filter, the local energy and a preset regularization parameter, and a sampling signal set corresponding to the three-dimensional point cloud data is determined according to the plurality of target nodes. In this way, the three-dimensional point cloud data can be segmented into a plurality of sub-blocks, the weighted adjacency matrix can be segmented into a plurality of sub-adjacency matrices, a plurality of sub-graphs corresponding to the plurality of sub-blocks can be determined according to the plurality of segmented sub-blocks and the plurality of segmented sub-adjacency matrices, a plurality of target nodes can be sampled in each sub-graph, and the sampling signal set can be obtained through the plurality of target nodes sampled in each sub-graph, so that when large-scale point cloud data is processed, the calculation complexity is low, the robustness is strong and the error is small; the plurality of target nodes can be sampled in the node of each sub-graph according to the filter, the local energy and the preset regularization parameter, so that the local energy can be maximized and the correlation interference can be minimized, and thus the sampling efficiency is high.

[0067] Based on the same inventive concept, as an implementation of the above-mentioned method for rapidly sampling three-dimensional point cloud based on graph structure, the embodiment of the present application further provides a device for rapidly sampling three-dimensional point cloud based on graph structure. Figure 2 The structure diagram of the device for rapidly sampling three-dimensional point cloud based on graph structure in the embodiment of the present application is shown in Figure 2 The device for rapidly sampling three-dimensional point cloud based on graph structure can include: The generating module 201 is configured to utilize the K-neighbor algorithm to generate a weighted adjacency matrix of the three-dimensional point cloud data; The segmentation module 202 is configured to segment the three-dimensional point cloud data and the weighted adjacency matrix respectively according to a preset number, obtain a plurality of sub-blocks and a plurality of corresponding sub-adjacency matrices, and determine a sub-graph corresponding to each sub-block and a graph Laplacian matrix corresponding to each sub-graph according to the plurality of sub-blocks and the plurality of sub-adjacency matrices; The determining module 203 is configured to utilize the Chebyshev approximation algorithm to determine a filter and a local energy of a node in each sub-graph according to an eigenvalue corresponding to the graph Laplacian matrix, a preset spectral truncation threshold and a pulse signal of the node in each sub-graph; The sampling module 204 is configured to sample a plurality of target nodes in the node of each sub-graph according to the filter, the local energy and a preset regularization parameter, and determine a sampling signal set corresponding to the three-dimensional point cloud data according to the plurality of target nodes.

[0068] The three-dimensional point cloud fast sampling device based on a graph structure can further include a reconstruction module configured to, after sampling a plurality of target nodes in each subgraph node according to a filter of a kernel function, local energy, and a preset regularization parameter, and determining a sampling signal set corresponding to the three-dimensional point cloud data according to the plurality of target nodes, reconstruct the sampling signal set by using an expression of graph Laplacian regularization to obtain a target reconstructed point cloud signal.

[0069] The generation module 201 is specifically configured to construct an adjacency graph of the three-dimensional point cloud data by using a K-nearest neighbor algorithm, the adjacency graph including each point in the three-dimensional point cloud data and K nearest neighbor points corresponding to each point; calculate an edge weight between each point and the K nearest neighbor points according to each point and the K nearest neighbor points; and generate a weighted adjacency matrix according to the edge weight.

[0070] The segmentation module 202 is specifically configured to segment the three-dimensional point cloud data into a plurality of subblocks according to a preset number, the three-dimensional point cloud data corresponding to each subblock being different, each subblock having a subblock index; segment the weighted adjacency matrix according to the subblock index to obtain a subadjacency matrix corresponding to each subblock; take a point set of each subblock as a vertex set, and construct a subgraph corresponding to each subblock according to the vertex set and the subadjacency matrix; calculate a degree matrix corresponding to the subadjacency matrix according to elements in the subadjacency matrix; and determine a graph Laplacian matrix corresponding to each subgraph according to the degree matrix and the subadjacency matrix.

[0071] The determination module 203 is specifically configured to determine a kernel function according to eigenvalues corresponding to the graph Laplacian matrix and a preset spectral truncation threshold; process the kernel function by using a Chebyshev approximation algorithm to obtain a filter; and determine local energy of a node in each subgraph according to the filter and an impulse signal of the node in each subgraph.

[0072] The sampling module 204 is specifically configured to determine a first localized filter response of the selected node according to the filter and the pulse signal of the selected node; determine a second localized filter response of the unselected node according to the filter and the pulse signal of the unselected node; determine a correlation square weighted accumulation between the selected node and the unselected node according to the first localized filter response, the second localized filter response, the first local energy and a preset regularization parameter; sample a plurality of target nodes in the nodes of each subgraph according to the correlation square weighted accumulation, the energy coupling accumulation, the second local energy and the preset regularization parameter, to obtain all target nodes of all subgraphs, the plurality of target nodes being nodes in the selected nodes; and merge all target nodes of all subgraphs to obtain a sampling signal set corresponding to the three-dimensional point cloud data, wherein the nodes of each subgraph include the selected nodes and the unselected nodes, the pulse signals of the nodes in each subgraph include the pulse signals of the selected nodes and the pulse signals of the unselected nodes, and the local energy includes the first local energy of the selected nodes and the second local energy of the unselected nodes.

[0073] In the sampling module 204, the expression of the correlation square weighted accumulation is: ; wherein, is the correlation square weighted accumulation between the selected node and the unselected node , is the first localized filter response of the selected node , is the transpose of the first localized filter response of the selected node , is the second localized filter response of the unselected node , is the first local energy of the selected node , is the preset regularization parameter, is the sampling signal set corresponding to the three-dimensional point cloud data.

[0074] In the sampling module 204, the expression of the energy coupling accumulation is: ; wherein, is the energy coupling accumulation between the selected node and the unselected node , is the first localized filter response of the selected node , is the transpose of the first localized filter response of the selected node , is the unselected node a second localized filter response of the selected node, a first local energy of the selected node a first local energy of the selected node, a preset regularization parameter, a set of sampling signals corresponding to the three-dimensional point cloud data.

[0075] In the sampling module 204, the expression of the plurality of target nodes is: ; wherein, the plurality of target nodes, a second localized filter response of the selected node, a second localized filter response of the selected node, a second local energy of the selected node, a second local energy of the selected node, a preset regularization parameter, a correlation square weighted cumulative amount between the selected node and the unselected node a correlation square weighted cumulative amount between the selected node and the unselected node an energy coupling cumulative amount between the selected node and the unselected node a set of unselected sampling signals composed of the unselected nodes.

[0076] It should be noted that the above description of the embodiment of the three-dimensional point cloud fast sampling device based on the graph structure is similar to the above description of the embodiment of the three-dimensional point cloud fast sampling method based on the graph structure, and has similar beneficial effects as the embodiment of the three-dimensional point cloud fast sampling method based on the graph structure. For technical details of the embodiment of the three-dimensional point cloud fast sampling device based on the graph structure of the present application that are not disclosed, please refer to the description of the embodiment of the three-dimensional point cloud fast sampling method based on the graph structure of the present application for understanding.

[0077] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for fast sampling of a three-dimensional point cloud based on a graph structure, characterized in that, The method comprises the following steps: generating a weighted adjacency matrix of the three-dimensional point cloud data by using a K-nearest neighbor algorithm; segmenting the three-dimensional point cloud data and the weighted adjacency matrix according to a preset number, respectively, to obtain a plurality of sub-blocks and a plurality of corresponding sub-adjacency matrices, and determining a sub-graph corresponding to each sub-block and a graph Laplacian matrix corresponding to each sub-graph according to the plurality of sub-blocks and the plurality of sub-adjacency matrices; determining a filter and a local energy of a node in each sub-graph by using a Chebyshev approximation algorithm according to an eigenvalue corresponding to the graph Laplacian matrix, a preset spectral truncation threshold, and an impulse signal of the node in each sub-graph; sampling a plurality of target nodes in the node of each sub-graph according to the filter, the local energy, and a preset regularization parameter, and determining a sampling signal set corresponding to the three-dimensional point cloud data according to the plurality of target nodes.

2. The method of claim 1, wherein, After sampling a plurality of target nodes in the node of each sub-graph according to the filter, the local energy, and a preset regularization parameter, and determining a sampling signal set corresponding to the three-dimensional point cloud data according to the plurality of target nodes, the method further comprises: reconstructing the sampling signal set by using an expression of graph Laplacian regularization to obtain a target reconstructed point cloud signal.

3. The method of claim 1, wherein, The method of generating a weighted adjacency matrix of the three-dimensional point cloud data by using a K-nearest neighbor algorithm comprises the following steps: constructing an adjacency graph of the three-dimensional point cloud data by using the K-nearest neighbor algorithm, each node in the adjacency graph corresponding to each point in the three-dimensional point cloud data, each point corresponding to K nearest neighbor points; calculating an edge weight between each point and the K nearest neighbor points according to each point and the K nearest neighbor points; generating the weighted adjacency matrix according to the edge weight.

4. The method of claim 1, wherein, The method of segmenting the three-dimensional point cloud data and the weighted adjacency matrix according to a preset number, respectively, to obtain a plurality of sub-blocks and a plurality of corresponding sub-adjacency matrices, and determining a sub-graph corresponding to each sub-block and a graph Laplacian matrix corresponding to each sub-graph according to the plurality of sub-blocks and the plurality of sub-adjacency matrices comprises the following steps: segmenting the three-dimensional point cloud data into the plurality of sub-blocks according to the preset number, the three-dimensional point cloud data corresponding to each sub-block being different, each sub-block having a sub-block index; segmenting the weighted adjacency matrix according to the sub-block index to obtain a sub-adjacency matrix corresponding to each sub-block; taking a point set of each sub-block as a vertex set, and constructing a sub-graph corresponding to each sub-block according to the vertex set and the sub-adjacency matrix; calculating a degree matrix according to elements in the sub-adjacency matrix; determining a graph Laplacian matrix corresponding to each sub-graph according to the degree matrix and the sub-adjacency matrix.

5. The method of claim 1, wherein, The method of determining a filter and a local energy of a node in each sub-graph by using a Chebyshev approximation algorithm according to an eigenvalue corresponding to the graph Laplacian matrix, a preset spectral truncation threshold, and an impulse signal of the node in each sub-graph comprises the following steps: determining a kernel function according to the eigenvalue corresponding to the graph Laplacian matrix and the preset spectral truncation threshold; The Chebyshev approximation algorithm is used to process the kernel function to obtain the filter; According to the filter and the impulse signal of the node in each subgraph, the local energy of the node in each subgraph is determined.

6. The method of claim 1, wherein, The node of each subgraph includes a selected node and an unselected node, the impulse signal of the node in each subgraph includes the impulse signal of the selected node and the impulse signal of the unselected node, and the local energy includes the first local energy of the selected node and the second local energy of the unselected node; According to the filter, the local energy and the preset regularization parameter, a plurality of target nodes are sampled in the node of each subgraph, and a sampling signal set corresponding to the three-dimensional point cloud data is determined according to the plurality of target nodes, including: According to the filter and the impulse signal of the selected node, a first localized filter response of the selected node is determined; According to the filter and the impulse signal of the unselected node, a second localized filter response of the unselected node is determined; According to the first localized filter response, the second localized filter response, the first local energy and the preset regularization parameter, a correlation square weighted cumulative quantity and an energy coupling cumulative quantity between the selected node and the unselected node are determined; According to the correlation square weighted cumulative quantity, the energy coupling cumulative quantity, the second local energy and the preset regularization parameter, the plurality of target nodes are sampled in the node of each subgraph, to obtain all target nodes of all subgraphs, and the plurality of target nodes are nodes in the selected node; The all target nodes of all subgraphs are merged to obtain the sampling signal set corresponding to the three-dimensional point cloud data.

7. The method of claim 6, wherein, The expression of the correlation square weighted cumulative quantity is: ； wherein, is a correlation square weighted cumulant between the selected node and the unselected node , is a first localized filter response of the selected node , is a transpose of the first localized filter response of the selected node , is a second localized filter response of the unselected node , is a first local energy of the selected node , is the preset regularization parameter, is a set of sampling signals corresponding to the three-dimensional point cloud data.

8. The method of claim 6, wherein, The expression of the energy coupling cumulative quantity is: ； wherein, is an accumulated amount of energy coupling between the selected node and the unselected node , is a first localized filter response of the selected node , is a transpose of the first localized filter response of the selected node , is a second localized filter response of the unselected node , is a first local energy of the selected node , is the preset regularization parameter, is a set of sampling signals corresponding to the three-dimensional point cloud data.

9. The method of claim 6, wherein, The expression of the plurality of target nodes is: ； wherein, is a second localized filter response of the plurality of target nodes, is a second localized filter response of the unselected nodes is a second localized energy of the unselected nodes, is a second localized energy of the unselected nodes is a second localized energy of the unselected nodes, is the preset regularization parameter, is a correlation square weighted accumulation between the selected nodes and the unselected nodes is an energy coupling accumulation between the selected nodes and the unselected nodes is an energy coupling accumulation between the selected nodes and the unselected nodes, is an unselected sampling signal set composed of the unselected nodes.

10. A device for fast sampling of a three-dimensional point cloud based on graph structure, characterized in that, including: The generating module is configured to generate a weighted adjacency matrix of the three-dimensional point cloud data by using a K-nearest neighbor algorithm; The segmentation module is configured to segment the three-dimensional point cloud data and the weighted adjacency matrix according to a preset number, respectively, to obtain a plurality of subblocks and a plurality of subadjacency matrices corresponding thereto, and determine a subgraph corresponding to each subblock and a graph Laplacian matrix corresponding to each subgraph according to the plurality of subblocks and the plurality of subadjacency matrices; The determining module is configured to determine a filter and a local energy of a node in each subgraph by using a Chebyshev approximation algorithm according to an eigenvalue corresponding to the graph Laplacian matrix, a preset spectral truncation threshold and an impulse signal of the node in each subgraph; The sampling module is configured to sample a plurality of target nodes in the node of each subgraph according to the filter, the local energy and a preset regularization parameter, and determine a sampling signal set corresponding to the three-dimensional point cloud data according to the plurality of target nodes.

Images