Three-dimensional point cloud reconstruction method and device based on graph cross attention network

Abstract

The invention provides a three-dimensional point cloud reconstruction method and device based on a graph cross attention network. The method comprises the steps that three-dimensional point cloud data are acquired and preprocessed; constructing a polyhedral graph structure according to the preprocessed three-dimensional point cloud data; extracting point cloud features of the preprocessed three-dimensional point cloud data by using a convolution encoder, projecting the point cloud features to three orthogonal view angle planes of XY, XZ and YZ to generate potential features, and sampling from the interior of a polyhedron by using a skeleton sampling algorithm to obtain query points; interpolating the query points on the potential features by adopting a bilinear interpolation method to obtain polyhedral features; a graph cross attention network is adopted to fuse spatial structure modeling of graph convolution and dynamic feature fusion capability of cross attention, a connection relation of a polyhedral graph structure is dynamically adjusted according to polyhedral features to obtain graph node features, and three-dimensional point cloud reconstruction is executed according to the graph node features; therefore, the robustness and the precision in complex part topology reconstruction are improved.

Description

Technical Field

[0001] The present invention relates to the field of point cloud data processing technology, and in particular to a three-dimensional point cloud reconstruction method based on a graph cross-attention network, a three-dimensional point cloud reconstruction device based on a graph cross-attention network, a computer-readable storage medium and a computer device. Background Art

[0002] Among the related technologies, visual computing in the field of intelligent manufacturing is closely related to three-dimensional reconstruction technology, and the two form a synergistic relationship in data acquisition and processing, model optimization, production monitoring, quality management, resource scheduling, and intelligent decision support. For example, three-dimensional reconstruction technology can be used to build a high-precision digital twin model of a manufacturing workshop, providing data support for production process optimization, intelligent detection, equipment operation and maintenance, etc. However, traditional three-dimensional reconstruction methods are highly dependent on professional software and technicians, and usually use manual or semi-automatic methods to extract and construct three-dimensional models from point cloud data. This process involves complex calculations such as feature extraction, point cloud matching, depth estimation, surface reconstruction, etc., which leads to large consumption of computing resources and long processing time, especially in large-scale industrial scenarios. It is difficult to meet real-time requirements.

[0003] With the advancement of deep learning, intelligent 3D reconstruction methods that integrate machine learning are becoming an important research direction in intelligent manufacturing. Graph neural networks (GNNs), due to their advantages in processing graph-structured data, have been widely used in point cloud parsing and 3D reconstruction tasks. However, in manufacturing scenarios, traditional graph convolutional networks (GCNs) still face the following challenges when processing 3D point clouds: 1) Static adjacency relationships are difficult to adapt to dynamic manufacturing environments: Traditional GCNs rely on a fixed adjacency matrix to define node connections, but point cloud data in manufacturing scenarios often suffer from uneven density and drastic local variations. For example, in the complex structures of industrial equipment and assembly line components, static adjacency relationships cannot accurately capture spatial correlations, affecting modeling accuracy. 2) Insufficient efficient processing capabilities in large-scale production environments: Modern intelligent manufacturing involves large-scale industrial scenarios, such as entire production lines and complex assembly systems. Existing graph convolutional models struggle to efficiently process such massive point cloud data. 3) Unimodal feature interactions limit intelligent recognition capabilities: Traditional GCNs primarily process geometric information (such as coordinates and normal vectors), but different components in manufacturing scenarios have unique semantic features, such as weld areas and defect detection points. Relying solely on geometric features makes it difficult to accurately distinguish between different parts, limiting the intelligent manufacturing system's ability to discern details. 4) 3D point cloud construction methods based on geometric features ignore process dependencies: Existing point cloud processing methods primarily rely on spatial proximity to construct an adjacency matrix, ignoring process information from the manufacturing process, such as material properties and process parameters. This limits the performance of intelligent manufacturing systems in tasks such as quality inspection and production line optimization. Summary of the Invention

[0004] The present invention aims to solve at least one of the technical problems in the above-mentioned technologies to a certain extent. To this end, one purpose of the present invention is to propose a 3D point cloud reconstruction method based on a graph cross-attention network, which does not rely on a fixed parameterized model and directly uses GCN to regress the coordinates of mesh vertices to adaptively capture the complex geometric shapes of industrial parts; by introducing geometric modes and spatial modes to construct a learnable graph structure, it breaks through the bottleneck of traditional static adjacency matrix modeling capabilities; combining spectral domain graph convolution with the cross-attention mechanism, it improves the structural expression capability with global attention relationships in the GCAN structure, and significantly improves the robustness and accuracy in the reconstruction of complex part topology.

[0005] A second object of the present invention is to provide a computer-readable storage medium.

[0006] A third object of the present invention is to provide a computer device.

[0007] The fourth object of the present invention is to propose a three-dimensional point cloud reconstruction device based on graph cross-attention network.

[0008] To achieve the above-mentioned purpose, the first embodiment of the present invention proposes a three-dimensional point cloud reconstruction method based on a graph cross-attention network, which includes the following steps: acquiring three-dimensional point cloud data; preprocessing the three-dimensional point cloud data to obtain preprocessed three-dimensional point cloud data; dividing the preprocessed three-dimensional point cloud data into local polyhedron areas to obtain a polyhedron node set, and calculating the edge weights between polyhedrons through geometric modality and spatial modality information based on the polyhedron node set to obtain an edge set; constructing a polyhedron graph structure based on the polyhedron node set and the edge set; and using a convolution encoder to extract the preprocessed three-dimensional point cloud data. The point cloud features are projected onto three orthogonal viewing planes (XY, XZ, and YZ) to generate potential features, and a skeleton sampling algorithm is used to sample from the inside of a polyhedron to obtain query points; the query points are interpolated on the potential features using a bilinear interpolation method to obtain polyhedron features; a graph cross-attention network is used to fuse the spatial structure modeling of graph convolution and the dynamic feature fusion capability of cross-attention, the connection relationship of the polyhedron graph structure is dynamically adjusted according to the polyhedron features to obtain graph node features, and three-dimensional point cloud reconstruction is performed according to the graph node features, thereby improving the robustness and accuracy in the reconstruction of complex part topology.

[0009] In addition, the 3D point cloud reconstruction method based on the graph cross attention network proposed in the above embodiment of the present invention may also have the following additional technical features:

[0010] Optionally, the three-dimensional point cloud data is preprocessed to obtain preprocessed three-dimensional point cloud data, including: performing point cloud denoising, downsampling and normalization on the preprocessed three-dimensional point cloud data; wherein, in the point cloud denoising process, a statistical filtering algorithm is used to adaptively remove abnormal points of the three-dimensional point cloud data through dynamic threshold adjustment; in the downsampling process, the original point density is maintained in the high curvature area to retain the detail features, while the point density is reduced by voxel rasterization in the flat area to achieve regional adaptive sampling rate control; in the normalization process, the point cloud is mapped as a whole to a unit space range to obtain the preprocessed three-dimensional point cloud data.

[0011] Optionally, the preprocessed three-dimensional point cloud data is divided into local polyhedral areas to obtain a polyhedral node set, including: using an octree segmentation algorithm to recursively segment the preprocessed three-dimensional point cloud data into multiple child nodes according to a preset termination condition; extracting a boundary point set for each child node after segmentation, generating a convex polyhedron through constrained Delaunay triangulation, and applying the Alpha Shapes algorithm to the concave area to generate polyhedral units to obtain a polyhedral node set.

[0012] Optionally, the edge weights between polyhedrons are calculated based on the polyhedron node set through geometric modal and spatial modal information to obtain an edge set, including: calculating the geometric modal weights and spatial modal weights of two adjacent polyhedron nodes; fusing the geometric modal weights and the spatial modal weights to obtain a fusion weight, and adjusting the fusion weight through a dynamic gating mechanism to obtain an edge set.

[0013] Optionally, calculating the geometric modal weights of two adjacent polyhedron nodes includes: obtaining the normal vectors corresponding to the two adjacent polyhedron nodes to calculate the normal vector consistency corresponding to the two adjacent polyhedron nodes; obtaining the center of mass corresponding to the two adjacent polyhedron nodes to calculate the center of mass distance attenuation corresponding to the two adjacent polyhedron nodes; and obtaining the geometric modal weights based on the normal vector consistency and the center of mass distance attenuation.

[0014] Optionally, calculating the spatial modal weights of two adjacent polyhedral nodes includes: calculating the spatial modal weights of the two adjacent polyhedral nodes according to spatial contact areas of the two adjacent polyhedral nodes.

[0015] To achieve the above-mentioned objectives, the second aspect of the present invention proposes a computer-readable storage medium on which a three-dimensional point cloud reconstruction program based on a graph cross-attention network is stored. When the three-dimensional point cloud reconstruction program based on a graph cross-attention network is executed by a processor, the three-dimensional point cloud reconstruction method based on a graph cross-attention network as described above is implemented.

[0016] To achieve the above-mentioned objectives, the third aspect of the present invention proposes a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements the three-dimensional point cloud reconstruction method based on the graph cross-attention network as described above.

[0017] To achieve the above-mentioned purpose, the fourth embodiment of the present invention proposes a three-dimensional point cloud reconstruction device based on a graph cross-attention network, comprising: an acquisition module for acquiring three-dimensional point cloud data; a preprocessing module for preprocessing the three-dimensional point cloud data to obtain preprocessed three-dimensional point cloud data; a calculation module for dividing the preprocessed three-dimensional point cloud data into local polyhedron areas to obtain a polyhedron node set, and calculating the edge weights between polyhedrons according to the polyhedron node set through geometric modality and spatial modality information to obtain an edge set; a graph topology construction module for constructing a polyhedron graph structure according to the polyhedron node set and the edge set; a feature sampling module for extracting point cloud features of the preprocessed three-dimensional point cloud data using a convolutional encoder, projecting the point cloud features to three orthogonal viewing planes of XY, XZ, and YZ to generate potential features, and sampling from the inside of the polyhedron using a skeleton sampling algorithm to obtain a query point;

[0018] A feature extraction module is used to interpolate the query point on the potential feature using a bilinear interpolation method to obtain a polyhedron feature; a point cloud reconstruction module is used to use a graph cross-attention network to fuse the spatial structure modeling of graph convolution and the dynamic feature fusion capability of cross-attention, dynamically adjust the connection relationship of the polyhedron graph structure according to the polyhedron feature to obtain graph node features, and perform three-dimensional point cloud reconstruction according to the graph node features, thereby improving the robustness and accuracy in the reconstruction of complex part topology. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 2. Schematic diagram of a flow chart of a three-dimensional point cloud reconstruction method based on a graph cross attention network according to an embodiment of the present invention;

[0020] Figure 2 2 is a network structure diagram of a three-dimensional point cloud reconstruction method based on a graph cross attention network according to an embodiment of the present invention;

[0021] Figure 3 Schematic diagram of a block diagram of a three-dimensional point cloud reconstruction device based on a graph cross-attention network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] The following describes embodiments of the present invention in detail, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and are not to be construed as limiting the present invention.

[0023] To better understand the above technical solutions, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. Instead, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0024] In order to better understand the above technical solution, the above technical solution will be described in detail below with reference to the accompanying drawings and specific implementation methods.

[0025] refer to Figure 1 As shown, the three-dimensional point cloud reconstruction method based on the graph cross attention network of an embodiment of the present invention includes the following steps:

[0026] S101, obtaining three-dimensional point cloud data.

[0027] As an embodiment, three-dimensional point cloud data of industrial parts can be acquired through various three-dimensional scanning devices such as laser scanners or structured light cameras.

[0028] S102: Preprocess the three-dimensional point cloud data to obtain preprocessed three-dimensional point cloud data.

[0029] It should be noted that in order to improve the quality of 3D point cloud data and the efficiency of subsequent processing, an adaptive density control algorithm is used to preprocess the input 3D point cloud data.

[0030] As an embodiment, three-dimensional point cloud data is preprocessed to obtain preprocessed three-dimensional point cloud data, including: performing point cloud denoising, downsampling and normalization processing on the preprocessed three-dimensional point cloud data; wherein, in the point cloud denoising process, a statistical filtering algorithm is used to adaptively remove abnormal points in the three-dimensional point cloud data through dynamic threshold adjustment; in the downsampling process, the original point density is maintained in the high curvature area to retain the detail features, while the point density is reduced by voxel rasterization in the flat area to achieve regional adaptive sampling rate control; in the normalization process, the point cloud is mapped as a whole to a unit space range to obtain the preprocessed three-dimensional point cloud data.

[0031] In other words, an improved statistical filtering method is used to remove outliers, retain the original density in high-curvature areas, and compress the sampling density in flat areas to balance data volume and information retention. At the same time, a learnable spatial transformation matrix is ​​introduced to achieve geometric normalization of the part point cloud to ensure that it adapts to the input specifications of the subsequent model. After the above processing, the point cloud data can be represented as: Among them, p o represents the oth point.

[0032] In addition, for any part sample, this application scales it to the unit scale space through a normalization operation and generates a set of polyhedral units with spatial adjacency. Ray tracing is used to determine whether each polyhedral unit is occupied by a part entity, thereby obtaining the true occupancy label of each polyhedral unit: Among them, y m =1 means the polyhedron is inside, y m =0 means it is located outside, providing a basis for subsequent supervision.

[0033] S103 , dividing the pre-processed three-dimensional point cloud data into local polyhedron regions to obtain a polyhedron node set, and calculating edge weights between polyhedrons based on the polyhedron node set using geometric modal and spatial modal information to obtain an edge set.

[0034] That is to say, an improved octree partitioning strategy is used to divide the preprocessed three-dimensional point cloud into local polyhedral regions, and multi-dimensional edge features are generated based on geometric modes (such as normal vectors and curvature) and spatial modes (such as spatial position relationships and contact areas). A learnable edge weight graph is constructed through multi-layer linear mapping to comprehensively characterize the spatial connectivity and geometric continuity of industrial parts.

[0035] As an embodiment, the preprocessed three-dimensional point cloud data is divided into local polyhedron areas to obtain a polyhedron node set, including: using an octree segmentation algorithm to recursively segment the preprocessed three-dimensional point cloud data into multiple child nodes according to a preset termination condition; extracting a boundary point set for each child node after segmentation, generating a convex polyhedron through constrained Delaunay triangulation, and applying the Alpha Shapes algorithm to the concave area to generate polyhedron units to obtain a polyhedron node set.

[0036] It should be noted that the segmentation threshold is dynamically adjusted based on the local density of the point cloud data. In high-density areas, a smaller segmentation threshold can be set to more finely divide the space; in low-density areas, a larger segmentation threshold can be set to reduce unnecessary segmentation. For example, if the point cloud distribution within a node is relatively uniform and the density is low, the segmentation can be terminated early.

[0037] As an embodiment, the edge weights between polyhedrons are calculated based on the geometric modal and spatial modal information of the polyhedron node set to obtain an edge set, including: calculating the geometric modal weights and spatial modal weights of two adjacent polyhedron nodes; fusing the geometric modal weights and spatial modal weights to obtain a fusion weight, and adjusting the fusion weight through a dynamic gating mechanism to obtain an edge set.

[0038] As an embodiment, the geometric modal weights of two adjacent polyhedron nodes are calculated, including: obtaining the normal vectors corresponding to the two adjacent polyhedron nodes to calculate the normal vector consistency corresponding to the two adjacent polyhedron nodes; obtaining the center of mass corresponding to the two adjacent polyhedron nodes to calculate the center of mass distance attenuation corresponding to the two adjacent polyhedron nodes; and obtaining the geometric modal weights based on the normal vector consistency and the center of mass distance attenuation.

[0039] As an embodiment, calculating the spatial modal weights of two adjacent polyhedral nodes includes: calculating the spatial modal weights of the two adjacent polyhedral nodes according to the spatial contact areas of the two adjacent polyhedral nodes.

[0040] The detailed polyhedron node generation process and edge weight calculation process are shown in Algorithm 1 and Algorithm 2 below:

[0041]

[0042] The curvature calculation formula is as follows:

[0043]

[0044] Among them, N(p o ) is the neighborhood point set with a radius of 0.3m, and the point cloud p o The normal vector represents n o .

[0045]

[0046]

[0047] S104: Construct a polyhedral graph structure according to the polyhedral node set and edge set.

[0048] Specifically, the edge weights between polyhedra are calculated based on the geometric modal information and spatial modal information, and then the graph structure is constructed. Among them, the node set V={v1,v2,…,v m}, representing each polyhedral unit, the edge set ε={ε1,ε2,…,ε n} represents the similarity relationship between polyhedrons; the calculation of edge weights comprehensively considers multiple factors, including the consistency of normal vectors between polyhedrons, the attenuation function of the distance between center of mass, and the size of the actual contact area, to comprehensively describe its spatial and structural correlation.

[0049] S105, using a convolution encoder to extract point cloud features of the preprocessed three-dimensional point cloud data, and projecting the point cloud features to three orthogonal viewing planes XY, XZ, and YZ to generate latent features, and using a skeleton sampling algorithm to sample from the inside of the polyhedron to obtain query points.

[0050] That is to say, after using the convolutional encoder to extract point cloud features, they are projected onto three orthogonal viewing planes, XY, XZ, and YZ, to generate the potential feature representation z. Then, a skeleton sampling strategy is executed inside the polyhedron to select representative sampling points from the inside, giving priority to retaining the vertex information of sharp geometric parts, effectively meeting the encoding needs of industrial parts of arbitrary shapes.

[0051] Specifically, the point cloud data Encoded as a neural network feature representation, using a convolutional encoder h θ Calculation features:

[0052]

[0053] The feature is projected onto three orthogonal planes XY, XZ, and YZ, and the U-Net network μ with shared weights is used to θ The potential feature z = μ θ (project u (g)). To address the geometric encoding challenge of arbitrary-shaped industrial parts, a skeleton sampling algorithm is used to sample representative points from the interior of the polyhedron and transform the geometric features into a fixed-length query set s = {s1, s2, ..., s k The more geometric structure information a representative sampling point contains, the more completely it can express the part shape. When the k value is small, vertices on sharp edges are more descriptive and outperform other axially distributed points. Skeleton sampling effectively extracts a point set that represents the core geometric features of industrial parts.

[0054] S106 , interpolating the query point on the potential feature using a bilinear interpolation method to obtain a polyhedron feature.

[0055] That is, the implicit geometric encoding of each query point is estimated through bilinear interpolation, forming a discrete binary occupancy function, which serves as an implicit representation of the spatial distribution state inside the modeling polyhedron. This implicit function is used to guide the generation of the 3D reconstruction boundary.

[0056] Specifically, through the bilinear interpolation method, the application interpolates the query point s on the potential feature z to obtain the implicit representation of the polyhedron:

[0057]

[0058] Here, bi represents the bilinear interpolation operation, which is used to estimate the implicit code of the query point. This feature can be regarded as a discrete occupancy function, which characterizes the spatial occupancy of industrial parts. It is a spatial classifier whose decision boundary is the part surface.

[0059] S107 uses a graph cross-attention network to fuse the spatial structure modeling of graph convolution and the dynamic feature fusion capability of cross-attention, dynamically adjusts the connection relationship of the polyhedron graph structure according to the polyhedron features to obtain graph node features, and performs three-dimensional point cloud reconstruction based on the graph node features.

[0060] That is to say, if Figure 2 As shown in the figure, a graph cross-attention network (GCAN) is constructed to integrate the local structure modeling capability of graph convolution with the global dynamic perception capability of the attention mechanism, alleviating the "over-smoothing" problem of traditional graph convolution in processing complex boundaries of industrial parts. In the attention mechanism, the query is derived from the current node feature, and the key / value comes from the adjacency propagation feature. The effective connection is screened through the Sparsify mechanism to achieve dynamic modeling of unstructured relationships between polyhedrons and enhance the robust expression of complex structural parts (such as holes, chamfers, rib positions, etc.). Preliminary mesh reconstruction is performed based on the deep features of the graph nodes, and key structures (such as flanges, screw holes, contour edges, grooves, etc.) are extracted and identified, and reconstructed into a regular 3D polyhedron representation; then post-processing steps such as Laplacian smoothing, edge collapse, and normal optimization are performed to improve mesh quality, reduce pseudo-structures, and enhance the applicability of the reconstructed model in downstream inspection, design analysis and other tasks.

[0061] Specifically, we propose a graph cross-attention network (GCAN), which combines the spatial structure modeling of graph convolution with the dynamic feature fusion capability of cross-attention to adjust the connection strength between nodes in the graph. Three linear transformation layers are used to construct the Query, Key, Value triple: Q = L q (H),K=L k (AH) and V=L v (AH). The attention matrix is ​​dynamically generated using the following formula: Where A′ does not depend on the original adjacency matrix (correlation tensor A) and is applicable to unstructured graphs. The adjusted complete adjacency matrix is:

[0062] A full =A+α·Sparsify(A′)

[0063] Sparsify includes Top-k screening and threshold truncation (A'[A'<β]=0). The final attention mechanism is:

[0064] Attention(Q,K,V)=softmax(A′⊙log(A full +∈))V

[0065] The update after combining graph convolution and cross attention is expressed as

[0066] H′=σ(AHW)+LayerNorm(Attention(Q,K,V))

[0067] During training, A full It can be learned and optimized; during inference, it remains frozen to improve efficiency. It uses logarithmic priors to enforce adjacency constraints, improving perception of the complex structures of industrial parts.

[0068] This paper considers the 3D reconstruction of industrial parts as a node classification problem. Nodes are polyhedrons, and classification is performed using coded features and adjacency information to estimate the probability of occupancy within the part. Multi-layer graph convolution is used to extract high-level semantics, and then the MLP classifier outputs the prediction:

[0069] y i =softmax(λ Θ (G i ))

[0070] Among them, λ Θ Represents the MLP network. The cross entropy loss function is used during training to supervise and optimize the prediction results.

[0071] In terms of reconstruction quality assessment, the Hausdorff distance is introduced to measure the maximum difference between the reconstructed surface and the real surface:

[0072] H=max{sup a∈A inf b∈B d(a,b),d(a,b)}

[0073] In addition, RMSE is used to measure the fidelity of the reconstructed parts in real-world coordinates to ensure the geometric accuracy and engineering suitability of the reconstruction results.

[0074] This method performs 3D reconstruction based on updated graph node features, reconstructing key geometric structures of industrial parts, such as flanges, holes, and contour edges. During the structural prediction process, these key part structures are identified and restored, and post-processing, including smoothing, denoising, and boundary refinement, is performed to improve model quality and practical usability. The reconstructed model is suitable for part inspection, design backtracking, and digital twin scenarios.

[0075] In summary, the three-dimensional point cloud reconstruction method based on the graph cross-attention network according to the embodiment of the present invention does not rely on a fixed parameterized model (such as 3DMM), and directly uses GCN to regress the coordinates of mesh vertices to adaptively capture the complex geometric shapes of industrial parts; introduces geometric modes (such as normal consistency and center of gravity distance) and spatial modes (such as contact area) to construct a learnable graph structure, and adjusts the connection strength through a dynamic gating mechanism (such as Sigmoid+MLP) to break through the bottleneck of traditional static adjacency matrix modeling capabilities; innovatively combines spectral domain graph convolution and cross-attention mechanism, and uses global attention relationship in the GCAN structure to enhance the structural expression ability, significantly improving the robustness and accuracy in the reconstruction of complex part topology.

[0076] In addition, the present invention also proposes a computer-readable storage medium, on which a three-dimensional point cloud reconstruction program based on a graph cross-attention network is stored. When the three-dimensional point cloud reconstruction program based on a graph cross-attention network is executed by a processor, it implements the three-dimensional point cloud reconstruction method based on a graph cross-attention network as described above.

[0077] In addition, an embodiment of the present invention also proposes a computer device, including a memory, a processor, and a computer program stored in the memory and runnable on the processor. When the processor executes the program, it implements the three-dimensional point cloud reconstruction method based on the graph cross-attention network as described above.

[0078] Figure 3 FIG is a block diagram of a three-dimensional point cloud reconstruction device based on a graph cross attention network according to an embodiment of the present invention. Figure 3 As shown, the three-dimensional point cloud reconstruction device includes: an acquisition module 10, a preprocessing module 20, a calculation module 30, a graph topology construction module 40, a feature sampling module 50, a feature extraction module 60 and a point cloud reconstruction module 70;

[0079] Among them, the acquisition module 10 is used to acquire three-dimensional point cloud data; the preprocessing module 20 is used to preprocess the three-dimensional point cloud data to obtain preprocessed three-dimensional point cloud data; the calculation module 30 is used to divide the preprocessed three-dimensional point cloud data into local polyhedron areas to obtain a polyhedron node set, and calculate the edge weights between polyhedrons based on the polyhedron node set through geometric mode and spatial mode information to obtain an edge set; the graph topology construction module 40 is used to construct a polyhedron graph structure based on the polyhedron node set and the edge set; the feature sampling module 50 is used to extract the preprocessed three-dimensional point cloud data using a convolution encoder Point cloud features are projected onto three orthogonal viewing planes, XY, XZ, and YZ, to generate latent features, and a skeleton sampling algorithm is used to sample from the inside of the polyhedron to obtain query points; a feature extraction module 60 is used to interpolate query points on the latent features using a bilinear interpolation method to obtain polyhedron features; a point cloud reconstruction module 70 is used to use a graph cross-attention network to fuse the spatial structure modeling of graph convolution and the dynamic feature fusion capability of cross-attention, dynamically adjust the connection relationship of the polyhedron graph structure according to the polyhedron features, to obtain graph node features, and perform three-dimensional point cloud reconstruction based on the graph node features.

[0080] It should be noted that the above explanations of the embodiment of the three-dimensional point cloud reconstruction method based on the graph cross-attention network are also applicable to the three-dimensional point cloud reconstruction device based on the graph cross-attention network of this embodiment, and will not be repeated here.

[0081] It will be understood by those skilled in the art that embodiments of the present invention may be provided as methods, systems, or computer program products. Thus, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0082] The present invention is described with reference to flowcharts and / or block diagrams of methods, devices (systems), and computer program products according to embodiments of the present invention. It should be understood that each process and / or block in the flowcharts and / or block diagrams, as well as combinations of processes and / or blocks in the flowcharts and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowcharts and / or block diagrams. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0083] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0084] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 The steps for the function specified in one or more boxes.

[0085] It should be noted that in the claims, any reference signs placed between parentheses shall not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claim. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, third etc. does not indicate any order. These words may be interpreted as names.

[0086] Although the preferred embodiments of the present invention have been described, those skilled in the art may make additional changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention.

[0087] Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if such changes and modifications fall within the scope of the claims and their equivalents, the present invention is intended to include such changes and modifications.

[0088] In the description of the present invention, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0089] In the present invention, unless otherwise expressly specified or limited, the terms "mounted," "connected," "connect," "fixed," etc. should be understood broadly. For example, they may refer to fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; internal communication between two components or interaction between two components. Those skilled in the art will understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0090] In the present invention, unless otherwise expressly specified or limited, when a first feature is "above" or "below" a second feature, it may mean that the first and second features are in direct contact, or that the first and second features are in indirect contact through an intermediary. Furthermore, when a first feature is "above," "above," or "above" a second feature, it may mean that the first feature is directly above or diagonally above the second feature, or simply means that the first feature is at a higher level than the second feature. When a first feature is "below," "below," or "below" a second feature, it may mean that the first feature is directly below or diagonally below the second feature, or simply means that the first feature is at a lower level than the second feature.

[0091] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms should not be understood as necessarily referring to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0092] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A 3D point cloud reconstruction method based on graph cross attention network, characterized in that: The following steps are involved: Obtain 3D point cloud data; Preprocessing the three-dimensional point cloud data to obtain preprocessed three-dimensional point cloud data; Dividing the preprocessed three-dimensional point cloud data into local polyhedron regions to obtain a polyhedron node set, and calculating edge weights between polyhedrons based on the polyhedron node set using geometric modal and spatial modal information to obtain an edge set; Constructing a polyhedral graph structure according to the polyhedral node set and the edge set; extracting point cloud features from the preprocessed three-dimensional point cloud data using a convolutional encoder, projecting the point cloud features onto three orthogonal viewing planes (XY, XZ, and YZ) to generate latent features, and sampling from the interior of a polyhedron using a skeleton sampling algorithm to obtain query points; interpolating the query points on the potential features using a bilinear interpolation method to obtain polyhedral features; A graph cross-attention network is adopted to fuse the spatial structure modeling of graph convolution and the dynamic feature fusion capability of cross-attention, and the connection relationship of the polyhedron graph structure is dynamically adjusted according to the polyhedron features to obtain graph node features, and three-dimensional point cloud reconstruction is performed based on the graph node features.

2. The 3D point cloud reconstruction method based on graph cross attention network according to claim 1, characterized in that: Preprocessing the three-dimensional point cloud data to obtain preprocessed three-dimensional point cloud data includes: Performing point cloud denoising, downsampling and normalization processing on the preprocessed three-dimensional point cloud data; In the point cloud denoising process, a statistical filtering algorithm is used to adaptively remove abnormal points in the three-dimensional point cloud data through dynamic threshold adjustment; During the downsampling process, the original point density is maintained in high-curvature areas to preserve detailed features, while voxel rasterization is used to reduce the point density in flat areas to achieve regional adaptive sampling rate control. During the normalization process, the point cloud is mapped as a whole to the unit space range to obtain the preprocessed three-dimensional point cloud data.

3. The 3D point cloud reconstruction method based on graph cross attention network according to claim 1, characterized in that: Dividing the preprocessed three-dimensional point cloud data into local polyhedron regions to obtain a polyhedron node set, including: An octree segmentation algorithm is used to recursively segment the pre-processed three-dimensional point cloud data into a plurality of child nodes according to a preset termination condition; For each child node after segmentation, a boundary point set is extracted, and a convex polyhedron is generated by constrained Delaunay triangulation. The Alpha Shapes algorithm is applied to the concave area to generate polyhedral units to obtain a polyhedral node set.

4. The 3D point cloud reconstruction method based on graph cross attention network according to claim 3, characterized in that: Calculating edge weights between polyhedrons using geometric modal and spatial modal information according to the polyhedron node set to obtain an edge set, including: Calculate the geometric modal weights and spatial modal weights of two adjacent polyhedron nodes; The geometric modal weight and the spatial modal weight are fused to obtain a fusion weight, and the fusion weight is adjusted through a dynamic gating mechanism to obtain an edge set.

5. The three-dimensional point cloud reconstruction method based on graph cross attention network according to claim 4, characterized in that Calculate the geometric modal weights of two adjacent polyhedral nodes, including: Obtaining normal vectors corresponding to two adjacent polyhedron nodes to calculate the consistency of the normal vectors corresponding to the two adjacent polyhedron nodes; Obtaining the centroids corresponding to two adjacent polyhedron nodes to calculate the centroid distance attenuation corresponding to the two adjacent polyhedron nodes; A geometric modal weight is obtained according to the normal vector consistency and the centroid distance attenuation.

6. The 3D point cloud reconstruction method based on graph cross attention network according to claim 5, characterized in that: Calculating the spatial modal weights of two adjacent polyhedron nodes includes: calculating the spatial modal weights of the two adjacent polyhedron nodes according to the spatial contact areas of the two adjacent polyhedron nodes.

7. A computer-readable storage medium, characterized in that A three-dimensional point cloud reconstruction program based on a graph cross-attention network is stored thereon, and when the three-dimensional point cloud reconstruction program based on a graph cross-attention network is executed by a processor, a three-dimensional point cloud reconstruction method based on a graph cross-attention network as described in any one of claims 1-6 is implemented.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, it implements the three-dimensional point cloud reconstruction method based on the graph cross attention network as described in any one of claims 1 to 6.

9. A 3D point cloud reconstruction device based on graph cross attention network, characterized in that: include: Acquisition module, used to acquire three-dimensional point cloud data; A preprocessing module, configured to preprocess the three-dimensional point cloud data to obtain preprocessed three-dimensional point cloud data; a calculation module, configured to divide the preprocessed three-dimensional point cloud data into local polyhedral regions to obtain a polyhedral node set, and calculate edge weights between polyhedrons based on the polyhedral node set using geometric modal and spatial modal information to obtain an edge set; A graph topology construction module, configured to construct a polyhedral graph structure according to the polyhedral node set and the edge set; a feature sampling module for extracting point cloud features from the preprocessed three-dimensional point cloud data using a convolutional encoder, projecting the point cloud features onto three orthogonal viewing planes (XY, XZ, and YZ) to generate latent features, and sampling from within a polyhedron using a skeleton sampling algorithm to obtain query points; a feature extraction module, configured to interpolate the query points on the potential features using a bilinear interpolation method to obtain polyhedral features; The point cloud reconstruction module is used to adopt the graph cross-attention network to fuse the spatial structure modeling of graph convolution and the dynamic feature fusion capability of cross-attention, dynamically adjust the connection relationship of the polyhedron graph structure according to the polyhedron features to obtain graph node features, and perform three-dimensional point cloud reconstruction according to the graph node features.

Images