Building model reconstruction method and device based on adaptive graph convolutional network

Abstract

The invention discloses a building model reconstruction method and device based on an adaptive graph convolutional network, and the method comprises the steps: obtaining three-dimensional point cloud data; processing the three-dimensional point cloud data by adopting an adaptive binary space division method to obtain a plurality of polyhedrons, and dynamically constructing a binary tree in the processing process to analyze the adjacency between the polyhedrons to obtain an adjacency matrix; obtaining a global implicit code according to the three-dimensional point cloud data, sampling a group of query points in each polyhedron, and obtaining polyhedron features according to coordinates of the query points and the global implicit code; constructing an adaptive graph convolutional network, and updating polyhedral features by combining the adjacent matrix and an attention adjacent matrix obtained through attention mechanism learning; performing three-dimensional building model reconstruction according to the updated polyhedral features; therefore, efficient and accurate conversion from the point cloud data to the three-dimensional building model is realized.

Description

Technical Field

[0001] The present invention relates to the field of intelligent construction technology, and in particular to a building model reconstruction method based on an adaptive graph convolutional network, a computer-readable storage medium, a computer device, and a building model reconstruction device based on an adaptive graph convolutional network. Background Art

[0002] 3D building model reconstruction is an important task in computer vision and graphics for smart construction sites. It is currently widely used in many fields such as urban planning, building information modeling (BIM), virtual reality (VR), augmented reality (AR), and geographic information systems (GIS). Traditional 3D reconstruction methods mainly rely on professional software and technicians to manually or semi-automatically reconstruct 3D models from point cloud data. These methods are not only time-consuming and labor-intensive, but also difficult to adapt to large-scale urban scenes.

[0003] To reconstruct compact polygonal building models, three categories of methods are commonly used in practice, including constrained reconstruction methods, geometric simplification methods, and primitive assembly methods. Constrained reconstruction methods represent buildings with predefined templates or specific topologies. However, the limited variety of available templates or topologies limits the expressiveness of these methods. Geometric simplification methods aim to obtain compact surfaces by simplifying dense triangles. However, these techniques require the input models to be both geometrically and topologically accurate to ensure a faithful approximation. Primitive assembly methods produce polygonal surface models by pursuing the optimal assembly of a series of geometric primitives. However, these methods usually require hand-crafted features and therefore have limited expressive capabilities. Despite their success in other applications, learning-based methods for compact building modeling remain largely unexplored.

[0004] With the development of deep learning technology, 3D reconstruction methods based on machine learning have gradually become a hot topic of research. Graph neural networks (GNNs) are widely used in point cloud processing and 3D reconstruction tasks due to their advantages in processing graph structured data. Adaptive graph convolutional networks combine traditional graph convolutions with attention mechanisms. By learning the dynamic connection strength between nodes, they can more flexibly and effectively capture the relationship between points in the point cloud, thereby improving the accuracy and efficiency of 3D reconstruction.

[0005] Although adaptive graph convolutional networks have shown great potential in processing graph-structured data, how to effectively apply them to 3D building reconstruction from point clouds, especially when processing point cloud data of large-scale and complex scenes, still faces a series of challenges:

[0006] Sparsity and unevenness of point cloud data: Point cloud data collected in the real world often have problems of sparseness and uneven distribution, which puts higher requirements on the robustness of the model.

[0007] Processing capabilities for large-scale scenarios: The amount of point cloud data at the city level is huge. How to design an efficient adaptive graph convolutional network model to process large-scale scenarios is the key to realizing practical applications.

[0008] Reconstruction of fine structures: The fine structures of buildings, such as windows and roofs, require a high degree of model precision, and traditional adaptive graph convolutional networks may be insufficient in capturing details.

[0009] Real-time requirements: In some application scenarios, such as autonomous driving and drone inspections, there are strict requirements for the real-time performance of 3D reconstruction, and the model needs to have a faster processing speed. Summary of the invention

[0010] 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 object of the present invention is to propose a building model reconstruction method based on an adaptive graph convolutional network, which can process point cloud data of large-scale and complex scenes, capture the subtle structure of the building, and meet the real-time requirements, thereby realizing efficient and accurate conversion from point cloud data to three-dimensional building models.

[0011] A second object of the present invention is to provide a computer-readable storage medium.

[0012] A third object of the present invention is to provide a computer device.

[0013] The fourth object of the present invention is to propose a building model reconstruction device based on an adaptive graph convolutional network.

[0014] To achieve the above-mentioned purpose, the first embodiment of the present invention proposes a method for reconstructing a building model based on an adaptive graph convolutional network, comprising acquiring three-dimensional point cloud data; processing the three-dimensional point cloud data using an adaptive binary space partitioning method to obtain a plurality of polyhedrons, and dynamically constructing a binary tree during the processing to analyze the adjacency between each polyhedron to obtain an adjacency matrix; obtaining a global latent code based on the three-dimensional point cloud data, and sampling a set of query points inside each polyhedron, and obtaining polyhedron features based on the coordinates of the query points and the global latent code; constructing an adaptive graph convolutional network, and updating the polyhedron features in combination with the adjacency matrix and the attention adjacency matrix obtained by learning the attention mechanism; reconstructing a three-dimensional building model based on the updated polyhedron features; thereby, it is possible to process point cloud data of large-scale and complex scenes, capture the subtle structure of the building, and meet real-time requirements, thereby achieving efficient and accurate conversion from point cloud data to three-dimensional building models.

[0015] In addition, the architectural model reconstruction method based on the adaptive graph convolutional network proposed in the above embodiment of the present invention may also have the following additional technical features:

[0016] Optionally, obtaining three-dimensional point cloud data includes: receiving original point cloud data of the building; and denoising, downsampling and normalizing the original point cloud data to obtain three-dimensional point cloud data.

[0017] Optionally, obtaining a global hidden code based on the three-dimensional point cloud data includes: converting the three-dimensional point cloud data into a neural feature representation using a convolutional encoder; and projecting the point-by-point neural feature representation onto three orthogonal planes to obtain a global hidden code.

[0018] Optionally, a group of query points are sampled inside each polyhedron, including: creating an empty set to store query points; obtaining a sampling number k, and determining whether the adopted number k is less than or equal to the number of vertices V of the polyhedron; if so, directly randomly selecting k vertices from all vertices as query points corresponding to the polyhedron, and storing them in the empty set; if not, calculating the number km of times each vertex is selected at least according to the adopted number and the number of vertices, and calculating the number kl of points that need to be additionally sampled after each vertex is evenly distributed, sampling km points for each vertex, and these points are along the vector direction from the vertex to the center of mass of the polyhedron, and for the last vertex, sampling km+kl points to obtain k query points, and storing them in the empty set.

[0019] Optionally, obtaining the polyhedral feature according to the coordinates of the query point and the global implicit code includes: performing bilinear interpolation on the global implicit code at the coordinates of the query point to obtain the polyhedral feature.

[0020] Optionally, an adaptive graph convolution network is constructed, and the adjacency matrix and the polyhedron features are input into the adaptive graph convolution network to update the polyhedron features, including: constructing an adaptive graph convolution network, wherein the adaptive graph convolution network includes a graph convolution layer and an attention mechanism layer; inputting the adjacency matrix and the polyhedron features into the adaptive graph convolution network to obtain an attention adjacency matrix by learning with the attention mechanism layer; obtaining a weighted adjacency matrix based on the adjacency matrix and the attention adjacency matrix; and updating the polyhedron features based on the weighted adjacency matrix.

[0021] Optionally, after reconstructing the three-dimensional building model according to the updated polyhedral features, the method further includes: smoothing, denoising and thinning the reconstructed three-dimensional building model to obtain an optimized three-dimensional building model.

[0022] To achieve the above-mentioned objectives, the second aspect of the present invention proposes a computer-readable storage medium, on which a building model reconstruction program based on an adaptive graph convolutional network is stored. When the building model reconstruction program based on an adaptive graph convolutional network is executed by a processor, the building model reconstruction method based on an adaptive graph convolutional network as described above is implemented.

[0023] To achieve the above 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 executable on the processor. When the processor executes the program, the architectural model reconstruction method based on the adaptive graph convolutional network as described above is implemented.

[0024] To achieve the above-mentioned purpose, the fourth aspect of the present invention proposes a building model reconstruction device based on an adaptive graph convolutional network, including an acquisition module for acquiring three-dimensional point cloud data; a graph construction module for processing the three-dimensional point cloud data using an adaptive binary space partitioning method to obtain multiple polyhedrons, and dynamically constructing a binary tree during the processing to analyze the adjacency between each polyhedron to obtain an adjacency matrix; a feature extraction module for obtaining a global latent code based on the three-dimensional point cloud data, and sampling a set of query points inside each polyhedron, and obtaining polyhedron features based on the coordinates of the query points and the global latent code; a network construction processing module for constructing an adaptive graph convolutional network, and updating the polyhedron features in combination with the adjacency matrix and the attention adjacency matrix obtained by learning the attention mechanism; a model reconstruction module for reconstructing a three-dimensional building model based on the updated polyhedron features; thereby, it is possible to process point cloud data of large-scale and complex scenes, capture the subtle structure of the building, and meet real-time requirements, thereby achieving efficient and accurate conversion from point cloud data to three-dimensional building models. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A schematic diagram of a process of a building model reconstruction method based on an adaptive graph convolutional network according to an embodiment of the present invention;

[0026] Figure 2 is a schematic diagram of an adaptive binary space partitioning process according to an embodiment of the present invention;

[0027] Figure 3 A schematic diagram of a framework for reconstructing a three-dimensional point cloud building model according to an embodiment of the present invention;

[0028] Figure 4 Schematic diagram of a block diagram of a building model reconstruction device based on an adaptive graph convolutional network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0029] Embodiments of the present invention are described in detail below, 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 should not be construed as limiting the present invention.

[0030] In order to better understand the above technical solution, 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. On the contrary, 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.

[0031] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0032] Figure 1 Schematic diagram of the process of the building model reconstruction method based on the adaptive graph convolutional network according to an embodiment of the present invention. Figure 1 As shown, the building model reconstruction method based on the adaptive graph convolutional network includes the following steps:

[0033] S101, obtaining three-dimensional point cloud data.

[0034] As an embodiment, obtaining three-dimensional point cloud data includes: receiving original point cloud data of a building; and denoising, downsampling, and normalizing the original point cloud data to obtain three-dimensional point cloud data.

[0035] As a specific embodiment, the original point cloud data of the building is first acquired through a laser radar (LiDAR) or other three-dimensional scanning equipment, and then the original point cloud data is preprocessed; the preprocessing includes denoising, downsampling and normalization, wherein the denoising can eliminate outliers and noise points; performing downsampling can reduce the number of points and increase the processing speed while retaining key geometric features; normalizing the downsampled original point cloud data can adapt to the input range of the model; thereby improving the efficiency and accuracy of subsequent processing.

[0036] S102, using an adaptive binary space partitioning method to process the three-dimensional point cloud data to obtain a plurality of polyhedrons, and dynamically constructing a binary tree during the processing to analyze the adjacency between each polyhedron to obtain an adjacency matrix.

[0037] That is, the three-dimensional point cloud data obtained after preprocessing is constructed into a graph structure. Specifically, each polyhedron is regarded as a node in the graph, and the edges between the nodes are established according to whether the polyhedrons are adjacent, so as to obtain the adjacency matrix A.

[0038] As a specific embodiment, an adaptive binary space partitioning method is used to obtain a polyhedron from a three-dimensional point cloud, and then a graph is constructed. Nodes of the graph represents a polyhedron, and the edges ε={ε1,ε2,…,ε n Specifically, a set of planar primitives are first identified from the input point cloud including buildings, and then the environment 3D space is partitioned to generate a linear unit complex of non-overlapping polyhedra that conform to the primitives. Figure 2 As shown, during the partitioning process, the adaptive binary space partitioning method dynamically constructs a binary tree to analyze the adjacency between polyhedrons. If two polyhedrons are adjacent (i.e., they share a face or vertex), the corresponding position in the adjacency matrix is ​​marked as 1 (or any value indicating connection), otherwise it is marked as 0.

[0039] S103, obtaining a global hidden code according to the three-dimensional point cloud data, sampling a set of query points inside each polyhedron, and obtaining polyhedron features according to the coordinates of the query points and the global hidden code.

[0040] That is, a set of representative points (query points) are sampled inside each polyhedron; the coordinates of these query points are connected with the global latent code to form polyhedron features, including but not limited to the point's position coordinates, normal vector, color, intensity, etc.

[0041] As an embodiment, a global hidden code is obtained based on three-dimensional point cloud data, including: using a convolutional encoder to convert the three-dimensional point cloud data into a neural feature representation; projecting the point-by-point neural feature representation onto three orthogonal planes to obtain a global hidden code.

[0042] Specifically, the three-dimensional point cloud data X is converted into a neural feature representation. This embodiment selects a convolutional encoder to implement this process:

[0043]

[0044] Then the point-by-point features are projected onto three feature planes and form a global latent code z, where u∈{XY,XZ,YZ} represents three orthogonal planes. θ U-Net with weights shared between planes, Project u Represents projection onto plane u, then the formula for obtaining the global hidden code z is as follows:

[0045] z=μ θ (projectu (g))

[0046] As an embodiment, a set of query points is sampled inside each polyhedron, including: creating an empty set to store query points; obtaining the sampling number k, and determining whether the adopted number k is less than or equal to the number of vertices V of the polyhedron; if so, directly randomly selecting k vertices from all vertices as query points corresponding to the polyhedron, and storing them in the empty set; if not, calculating the number km of times each vertex is selected at least based on the adopted number and the number of vertices, and calculating the number kl of points that need to be additionally sampled after each vertex is evenly distributed, sampling km points for each vertex, and these points are along the vector direction from the vertex to the center of mass of the polyhedron, and for the last vertex, sampling km+kl points to obtain k query points, and storing them in the empty set.

[0047] It should be noted that one of the challenges in encoding polyhedra of arbitrary shapes is to consistently describe the geometry of heterogeneous polyhedra. To solve this problem, this application samples representative points from the interior of the polyhedron and forces the geometry to be converted into a fixed-length query s of size k, s = {s1, s2, …, s k}. Obviously, the more representative the sample points are, the more information they convey about the polyhedron. When the k value is low, vertices are preferred over points along the axes because they have an outstanding advantage in describing sharp geometric shapes. This application selects skeleton sampling from vertices and main axes, and the specific implementation method is as follows:

[0048] 1. Initialize the output collection:

[0049] Create an empty set S to store the final query points.

[0050] 2. Determine the relationship between the number of samples and the number of vertices:

[0051] Check whether the input sample number k is less than or equal to the number of vertices V of the polyhedron.

[0052] 3. Sampling when the number of vertices is sufficient:

[0053] If k ≤ |V|, k vertices are randomly selected from all vertices as query points.

[0054] The selected vertex set is recorded as V s , and assign it to S.

[0055] 4. Sampling when the number of vertices is insufficient:

[0056] If k>|V|, a more complex sampling strategy is needed to ensure that the number of representative points meets the requirements.

[0057] calculate Indicates the number of times each vertex is selected at least.

[0058] Calculate kl = k mod |V|, which represents the number of points that need additional sampling after average distribution.

[0059] 5. Skeleton point sampling:

[0060] For each vertex V i (where i = 1, 2, ..., |V|-1), sample km points along the vertex V i The direction of the vector to the center of mass C of the polyhedron.

[0061] For the last vertex V |V| , sample km+kl points to ensure that the total number of sampled points is k.

[0062] 6. Merge sampling points:

[0063] All sampled points are merged into a set S, which will be used as query points of the polyhedron.

[0064] 7. Return results:

[0065] Returns the query point set S, which will be used for subsequent adaptive graph convolutional network processing.

[0066] Through the above steps, the skeleton sampling algorithm can extract points that can represent its geometric features from the polyhedron. These points are optimized in quantity and distribution to ensure that these key information can be effectively used in the subsequent 3D building reconstruction process. This sampling method is particularly suitable for polyhedrons with complex geometric shapes and can improve the accuracy and efficiency of the reconstruction process.

[0067] As an embodiment, obtaining the polyhedral feature according to the coordinates of the query point and the global latent code includes: performing bilinear interpolation on the global latent code at the coordinates of the query point to obtain the polyhedral feature.

[0068] Specifically, for the global implicit code z, the conditional implicit representation z is formed by bilinear interpolation of z at the coordinates of the query point s. s :

[0069]

[0070] Where bi represents a bilinear interpolation operation, which is used to estimate the value of the implicit code at the query point; this feature representation can be viewed as a discrete occupancy function, which describes the occupancy of the polyhedron based on the underlying building instance. This representation can be interpreted as a spatial classifier, whose decision boundary is the surface of the building. Unlike the method of approximating continuous implicit functions by exhaustive enumeration, the discretized representation of this application takes into account the geometric priors of a single polyhedron, significantly reducing the computational complexity and reducing the ambiguity of the solution.

[0071] S104, construct an adaptive graph convolutional network, and update the polyhedron features by combining the adjacency matrix and the attention adjacency matrix learned through the attention mechanism.

[0072] That is to say, the graph structure is processed using an adaptive graph convolutional network, in which the adjacency matrix A and the attention adjacency matrix A' learned through the attention mechanism fuse the learning graph structure information. Through end-to-end learning, the attention adjacency matrix A' reflects the importance or similarity between nodes; then, through the adaptive graph convolutional network, the original adjacency matrix A and the attention adjacency matrix A' are combined to update the polyhedron features (node ​​features) to achieve effective propagation of information in the graph.

[0073] As an embodiment, an adaptive graph convolution network is constructed, and the adjacency matrix and polyhedron features are input into the adaptive graph convolution network to update the polyhedron features, including: constructing an adaptive graph convolution network, wherein the adaptive graph convolution network includes a graph convolution layer and an attention mechanism layer; inputting the adjacency matrix and polyhedron features into the adaptive graph convolution network to obtain an attention adjacency matrix by learning the attention mechanism layer; obtaining a weighted adjacency matrix based on the adjacency matrix and the attention adjacency matrix; and updating the polyhedron features based on the weighted adjacency matrix.

[0074] Specifically, in the adaptive graph convolutional network, the adjacency matrix A is combined with the attention adjacency matrix A′=(v′,ε′) learned through the attention mechanism to form a weighted adjacency matrix E=A+A′. A′ dynamically adjusts the connection strength between nodes. The attention mechanism allows the model to assign different weights to different edges based on the characteristics of the nodes and the context. The formula principle of A′ is as follows:

[0075]

[0076] Where a represents a learnable weight vector, W represents a learnable weight matrix, [·∥·] represents a vector connection, and N(i) represents a node v i The neighbor node set of ,σ(.) represents the LeakyReLU activation function. The adaptive graph convolutional network can be formulated as:

[0077]

[0078] in, D = diag[d] represents the degree matrix, d(i) = ∑ j A i,j , I represents the unit matrix; K is the number of filters, which means that K steps of diffusion graph convolution are performed, which can superimpose neighbor information in multiple layers. Through multiple layers of adaptive graph convolution layers, multi-hop propagation of information in the graph is realized.

[0079] S105, reconstructing the three-dimensional building model according to the updated polyhedron features.

[0080] That is to say, the updated node features are used to reconstruct the three-dimensional building model, including surface reconstruction and structure prediction.

[0081] Specifically, based on the updated node features, surface reconstruction is performed to form a preliminary 3D building surface. Through the structure prediction step, key building structures such as roofs, walls, windows, etc. are identified and reconstructed.

[0082] In order to evaluate the performance of the 3D reconstruction method, this application can comprehensively evaluate and compare the effectiveness and practicality of different 3D reconstruction methods through multi-dimensional evaluations such as classification accuracy, fidelity, and success rate. Classification accuracy directly affects the fidelity of reconstruction and is therefore used as one of the evaluation indicators. This usually involves the model's ability to recognize different categories (such as the interior and exterior of a building).

[0083] The Hausdorff distance (H) is used to quantify the difference between the reconstructed surface and the true surface (ground truth). The Hausdorff distance is a measure that describes the maximum difference between two point sets and is defined as:

[0084] H=max{sup a∈A inf b∈B d(a,b),d(a,b)}

[0085] Where d(a,b) represents the distance between points a and b. This application uses RMSE to evaluate the fidelity of reconstruction on real-world data.

[0086] As an embodiment, after reconstructing the three-dimensional building model according to the updated polyhedral features, the method further includes: smoothing, denoising and thinning the reconstructed three-dimensional building model to obtain an optimized three-dimensional building model.

[0087] That is, by performing model post-processing optimization steps, including smoothing, denoising, and refinement, the reconstructed 3D building model is further optimized to improve the quality of the reconstructed model.

[0088] Therefore, the framework structure of the three-dimensional point cloud building reconstruction of the present application is as follows: Figure 3 As shown in the figure. Given an input point cloud, a graph topology is constructed by polyhedron decomposition, where polyhedra are graph nodes. Node features are formed by polyhedron queries against shape latent codes. Using the encoded node features and inter-polyhedron adjacency, graph nodes are classified to estimate building occupancy; a post-processing step is performed on the reconstructed 3D building model to remove possible artifacts and irregularities. A mesh simplification algorithm is applied to optimize the rendering efficiency and storage requirements of the model.

[0089] In summary, according to the method for reconstructing a building model based on an adaptive graph convolutional network in an embodiment of the present invention, first, three-dimensional point cloud data is acquired; then, an adaptive binary space partitioning method is used to process the three-dimensional point cloud data to obtain a plurality of polyhedrons, and a binary tree is dynamically constructed during the processing to analyze the adjacency between each polyhedron to obtain an adjacency matrix; then, a global latent code is obtained according to the three-dimensional point cloud data, and a set of query points are sampled inside each polyhedron, and polyhedron features are obtained according to the coordinates of the query points and the global latent code; then, an adaptive graph convolutional network is constructed, and the adjacency matrix is ​​dynamically constructed to obtain the adjacency matrix; The polyhedron features are updated by the adjacency matrix and the attention adjacency matrix learned through the attention mechanism; finally, the three-dimensional building model is reconstructed according to the updated polyhedron features; thus, the adaptive graph convolutional network can efficiently process large-scale point cloud data to meet real-time or near real-time application requirements, dynamically adjust the weights between nodes through the attention mechanism, capture the subtle structure of the building, and improve the reconstruction accuracy; the model can handle sparse and unevenly distributed point cloud data, has strong robustness, and the adaptive graph convolutional network structure allows the model to be adjusted and optimized according to different application scenarios and data characteristics.

[0090] In order to implement the above-mentioned embodiment, the embodiment of the present invention further proposes a computer-readable storage medium, on which a building model reconstruction program based on an adaptive graph convolutional network is stored. When the building model reconstruction program based on the adaptive graph convolutional network is executed by a processor, the building model reconstruction method based on the adaptive graph convolutional network as described above is implemented.

[0091] According to the computer-readable storage medium of an embodiment of the present invention, by storing a building model reconstruction program based on an adaptive graph convolutional network, the processor implements the building model reconstruction method based on an adaptive graph convolutional network as described above when executing the building model reconstruction program based on an adaptive graph convolutional network. As a result, it is possible to process point cloud data of large-scale and complex scenes, capture the subtle structure of the building, and meet real-time requirements, thereby achieving efficient and accurate conversion from point cloud data to a three-dimensional building model.

[0092] In order to implement the above embodiment, an embodiment of the present invention proposes a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the architectural model reconstruction method based on the adaptive graph convolutional network as described above is implemented.

[0093] According to the computer device of an embodiment of the present invention, the building model reconstruction program based on the adaptive graph convolutional network is stored in the memory, so that when the processor executes the building model reconstruction program based on the adaptive graph convolutional network, the building model reconstruction method based on the adaptive graph convolutional network as described above is implemented. As a result, it is possible to process point cloud data of large-scale and complex scenes, capture the subtle structure of the building, and meet the real-time requirements, thereby realizing efficient and accurate conversion from point cloud data to three-dimensional building models.

[0094] In order to implement the above embodiment, the embodiment of the present invention proposes a building model reconstruction device based on an adaptive graph convolutional network, such as Figure 4 As shown, the building model reconstruction device based on the adaptive graph convolutional network includes: an acquisition module 10, a graph construction module 20, a feature extraction module 30, a network construction processing module 40 and a model reconstruction module 60.

[0095] Among them, the acquisition module 10 is used to acquire three-dimensional point cloud data; the graph construction module 20 is used to process the three-dimensional point cloud data using an adaptive binary space partitioning method to obtain multiple polyhedrons, and dynamically construct a binary tree during the processing to analyze the adjacency between each polyhedron to obtain an adjacency matrix; the feature extraction module 30 is used to obtain a global latent code based on the three-dimensional point cloud data, and to sample a set of query points inside each polyhedron, and obtain polyhedron features based on the coordinates of the query points and the global latent code; the network construction processing module 40 is used to construct an adaptive graph convolutional network, and update the polyhedron features in combination with the adjacency matrix and the attention adjacency matrix obtained by learning the attention mechanism; the model reconstruction module 50 is used to reconstruct a three-dimensional building model based on the updated polyhedron features.

[0096] It should be noted that the above Figure 1 The description and explanation of the embodiment of the method for reconstructing a building model based on an adaptive graph convolutional network are also applicable to the device for reconstructing a building model based on an adaptive graph convolutional network, and will not be repeated here.

[0097] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Moreover, 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 disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0098] 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 flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, 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 flowchart and / or block diagram. 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.

[0099] These computer program instructions may also be stored in a computer readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture including an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0100] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process in the computer or other programmable device. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0101] It should be noted that in the claims, any reference signs placed between brackets 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 may 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 the same item of hardware. The use of the words first, second, and third etc. does not indicate any order. These words may be interpreted as names.

[0102] Although the preferred embodiments of the present invention have been described, those skilled in the art may make other 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.

[0103] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

[0104] 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 as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "plurality" is two or more, unless otherwise clearly and specifically defined.

[0105] In the present invention, unless otherwise clearly specified and limited, the terms "installed", "connected", "connected", "fixed" and the like should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium, it can be the internal connection of two elements or the interaction relationship between two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0106] In the present invention, unless otherwise clearly specified and limited, a first feature being "above" or "below" a second feature may mean that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. Moreover, a first feature being "above", "above" or "above" a second feature may mean that the first feature is directly above or obliquely above the second feature, or simply means that the first feature is higher in level than the second feature. A first feature being "below", "below" or "below" a second feature may mean that the first feature is directly below or obliquely below the second feature, or simply means that the first feature is lower in level than the second feature.

[0107] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. means 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 representation 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 may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples, unless they are contradictory.

[0108] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A building model reconstruction method based on an adaptive graph convolutional network, characterized in that: The following steps are involved: Obtain 3D point cloud data; The three-dimensional point cloud data is processed by an adaptive binary space partitioning method to obtain a plurality of polyhedrons, and a binary tree is dynamically constructed during the processing to analyze the adjacency between each polyhedron to obtain an adjacency matrix; Obtaining a global hidden code according to the three-dimensional point cloud data, sampling a set of query points inside each polyhedron, and obtaining polyhedron features according to the coordinates of the query points and the global hidden code; Constructing an adaptive graph convolutional network, and updating the polyhedron features by combining the adjacency matrix and the attention adjacency matrix learned by the attention mechanism; The three-dimensional building model is reconstructed according to the updated polyhedron features.

2. The architectural model reconstruction method based on the adaptive graph convolutional network according to claim 1, characterized in that: Get 3D point cloud data, including: Receive raw point cloud data of the building; The original point cloud data is subjected to denoising, downsampling and normalization processing to obtain three-dimensional point cloud data.

3. The architectural model reconstruction method based on the adaptive graph convolutional network according to claim 1, characterized in that: A global hidden code is obtained according to the three-dimensional point cloud data, including: Using a convolutional encoder to convert the three-dimensional point cloud data into a neural feature representation; The point-by-point neural feature representation is projected onto three orthogonal planes to obtain the global latent code.

4. The architectural model reconstruction method based on the adaptive graph convolutional network according to claim 1, characterized in that: A set of query points are sampled inside each polyhedron, including: Create an empty collection to store query points; Obtaining a sampling number k, and determining whether the sampling number k is less than or equal to the number of vertices V of the polyhedron; If yes, directly randomly select k vertices from all vertices as query points corresponding to the polyhedron and store them in the empty set; If not, then calculate the at least number of times km that each vertex is selected based on the number of adoptions and the number of vertices, as well as the number of points kl that need to be additionally sampled after each vertex is evenly distributed. For each vertex, km points are sampled along the vector direction from the vertex to the center of mass of the polyhedron. For the last vertex, km+kl points are sampled to obtain k query points, and stored in the empty set.

5. The architectural model reconstruction method based on adaptive graph convolutional network according to claim 1, characterized in that: Obtaining polyhedral features according to the coordinates of the query point and the global implicit code, including: Bilinear interpolation is performed on the global latent code at the coordinates of the query point to obtain a polyhedral feature.

6. The architectural model reconstruction method based on the adaptive graph convolutional network according to claim 1, characterized in that: Constructing an adaptive graph convolutional network, and inputting the adjacency matrix and the polyhedron features into the adaptive graph convolutional network to update the polyhedron features, including: Constructing an adaptive graph convolutional network, wherein the adaptive graph convolutional network includes a graph convolution layer and an attention mechanism layer; Inputting the adjacency matrix and the polyhedron features into the adaptive graph convolutional network to obtain an attention adjacency matrix using an attention mechanism layer; Obtaining a weighted adjacency matrix according to the adjacency matrix and the attention adjacency matrix; The polyhedron features are updated according to the weighted adjacency matrix.

7. The architectural model reconstruction method based on adaptive graph convolutional network according to claim 1, characterized in that: After the 3D building model is reconstructed based on the updated polyhedral features, it also includes: The reconstructed 3D building model is smoothed, denoised and refined to obtain an optimized 3D building model.

8. A computer-readable storage medium, characterized in that: An architectural model reconstruction program based on an adaptive graph convolutional network is stored thereon, and when the architectural model reconstruction program based on an adaptive graph convolutional network is executed by a processor, an architectural model reconstruction method based on an adaptive graph convolutional network as described in any one of claims 1-7 is implemented.

9. A computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the program, the architectural model reconstruction method based on the adaptive graph convolutional network as described in any one of claims 1 to 7 is implemented.

10. A building model reconstruction device based on an adaptive graph convolutional network, characterized in that: include An acquisition module, used to acquire three-dimensional point cloud data; A graph construction module, for processing the three-dimensional point cloud data using an adaptive binary space partitioning method to obtain a plurality of polyhedrons, and dynamically constructing a binary tree during the processing to analyze the adjacency between each polyhedron to obtain an adjacency matrix; A feature extraction module, used to obtain a global hidden code according to the three-dimensional point cloud data, and to sample a set of query points inside each polyhedron, and to obtain polyhedron features according to the coordinates of the query points and the global hidden code; A network construction processing module, used to construct an adaptive graph convolutional network, and update the polyhedron features in combination with the adjacency matrix and the attention adjacency matrix learned through the attention mechanism; The model reconstruction module is used to reconstruct the three-dimensional building model according to the updated polyhedral features.

Images