A method and system for three-dimensional reconstruction of urban space based on big data

Abstract

The present invention relates to the technical field of three-dimensional reconstruction, and discloses a method and system for three-dimensional reconstruction of urban space based on big data, comprising the following steps: step S101: collecting shooting information of a plurality of multi-angle spectral images of a target urban space and corresponding spectral images; step S102: preprocessing the photographed spectral images to obtain the preprocessed spectral images; step S103: collecting wide-area data to train an improved adversarial neural network; step S104: using a graph network of the target space to replace a preset random occlusion module, and reversely updating the graph network of the target space; step S105: modeling the target space based on point cloud data of the target space; the present invention realizes the generation of a high-quality and structurally reasonable three-dimensional model when the target space data is incomplete through the collaborative work of the improved adversarial neural network and the graph network.

Description

Technical Field

[0001] The present invention relates to the field of three-dimensional reconstruction, and more specifically, to a method and system for three-dimensional reconstruction of urban space based on big data. Background Art

[0002] With the development of big data technology, 3D reconstruction of urban space has become a key technology in urban planning, architectural design, and environmental monitoring. 3D reconstruction collects a large amount of image data and point cloud data to accurately restore elements such as buildings and roads in the city into digital models.

[0003] Existing 3D reconstruction technology usually relies on a large amount of point cloud data or a large number of high-definition images to achieve better modeling effects. However, the modeling method based on data volume has obvious marginal effects. The data collected to improve the modeling effect is not only difficult to align, but also the improved modeling effect is very limited and the cost is very high. However, not collecting such data will lead to problems with modeling accuracy and poor results. Summary of the invention

[0004] The present invention provides a method and system for three-dimensional reconstruction of urban space based on big data, which solves the problems raised in the background technology.

[0005] In a first aspect, the present invention provides a method for three-dimensional reconstruction of urban space based on big data, comprising:

[0006] Step S101: collecting a number of multi-angle spectral images of the target city space and the shooting information of the corresponding spectral images; the shooting information includes: camera principal point coordinates, optical coordinates and focal length;

[0007] Step S102: preprocessing the spectral image to obtain a preprocessed spectral image; extracting feature points in the preprocessed spectral image, mapping the feature points to nodes in a graph network, and constructing a graph network based on the nodes;

[0008] Step S103: Collect wide-area data to train an improved adversarial neural network, the adversarial neural network includes a discriminator A, a generator A, a discriminator B and a random occlusion module; the wide-area data includes a graph network of the training space and a graph network and point cloud data of the pan-space; wherein the point cloud data of the pan-space is used as a real sample A of the discriminator A, and the graph network of the pan-space is used as a real sample B of the discriminator B; part of the graph network of the training space is shielded by the random occlusion module, and the shielded graph network is respectively input into the discriminator B and the generator A, and the discriminator B outputs an evaluation value of the graph network; wherein the random occlusion module randomly changes the shielded area and position of the graph network of the training space; the generator A inputs the shielded graph network of the training space, and outputs the pseudo point cloud data of the training space, and the discriminator A outputs the evaluation value of the pseudo point cloud data;

[0009] Step S104: The preset random occlusion module is replaced by the graph network of the target space, and the graph network of the target space is reversely updated according to the evaluation value of the graph network output by the discriminator B. The updated graph network of the target space is input into the generator A to obtain the point cloud data of the target space;

[0010] Step S105: Modeling the target space based on the point cloud data of the target space.

[0011] Further, the spectral image is preprocessed to obtain a preprocessed spectral image, including:

[0012] A spectral image taken vertically from above is selected as a reference, a number of registration points are set in the spectral image, a non-reference spectral image is matched with the reference spectral image through the registration points, and the non-reference spectral image is corrected to obtain a multi-view spectral image set;

[0013] In each spectral image, feature edges are extracted, feature points are detected and extracted, and the two-dimensional coordinates of the feature points are obtained; the number of times the feature points appear in the multi-view spectral image set is counted, and feature points with a number below a threshold are removed;

[0014] The disparity information is obtained through the position of the two-dimensional coordinates of the feature points in different spectral images, and the two-dimensional coordinates of the feature points are converted into three-dimensional coordinates based on the two-dimensional coordinates of the feature points, the disparity information and the shooting information of the spectral images.

[0015] Furthermore, a graph network is constructed based on nodes, including:

[0016] Feature points are mapped to nodes of the graph network; if there is a feature edge between two nodes, the feature edge is mapped to the edge of the graph network; the initial feature of the node is the feature vector of the feature point, which includes the coordinates, color and reflectivity data of the feature point.

[0017] Furthermore, the graph network of the target space is reversely updated according to the evaluation value of the output graph network of the discriminator B, including:

[0018] Calculate the weight of each node to other nodes and construct a weight matrix;

[0019] Perform cluster analysis on the weight matrix, distinguish feature points into high-weight feature points and low-weight feature points based on the weight factor threshold, and divide the graph network into high-weight areas and low-weight areas based on the high-weight feature points and low-weight feature points;

[0020] Insert a new node in the low-weight area and build edges between the new node and the high-weight feature points close to the new node;

[0021] Discriminator B outputs an evaluation value for the graph network in the target space. If the evaluation value is lower than the evaluation threshold, the position of the node is adjusted.

[0022] Further, adjust the position of the node, including:

[0023] Define the reverse optimization factor of the graph network according to the evaluation value of the discriminator B on the graph network;

[0024] Calculate the partial derivative of the inverse optimization factor with respect to the position of each new node;

[0025] According to the direction of the partial derivative, move the new node in the direction that reduces the evaluation value of the discriminator B on the graph network;

[0026] The position of the new node is adjusted iteratively until the evaluation value of the discriminator B on the graph network is greater than the evaluation threshold.

[0027] Furthermore, the nodes in the graph network are labeled and a generator matrix is ​​constructed; the size of the generator matrix is ​​D×K; wherein each row of the generator matrix represents the coordinates and features of a feature point;

[0028] Count the number of edges of feature points and set the dimension of each row of the generator matrix to K;

[0029] Fill in the coordinates, color, reflectivity and edge relationship of the feature points in sequence;

[0030] Set multiple identifiers, and use multiple identifiers to identify coordinates, colors, reflectivity and edge relationships respectively;

[0031] If the coordinates, color, reflectivity and edge relationship filling dimensions of the feature point are less than K dimensions, the feature point label and feature point weight are filled with masks, the identifier remains unchanged, and the edge relationship is expanded to K dimensions;

[0032] If the number of feature points is ≤D, the number of rows of the generator matrix is ​​expanded to D dimensions, and the expanded number of rows is filled with masks.

[0033] Furthermore, the random occlusion module includes:

[0034] The random occlusion module randomly selects one or more regions with random positions and areas in the graph network of the training space for occlusion; the occluded area includes occluding the graph network nodes and edges within the selected area, and the occluded nodes and edges will be removed from the graph network.

[0035] Furthermore, the target space is modeled based on the point cloud data of the target space, including:

[0036] The point cloud data of the target space is converted into a three-dimensional model; the color and reflectivity information in the point cloud data is mapped to the surface of the three-dimensional model; and the three-dimensional model of the target space is obtained.

[0037] In a second aspect, the present invention provides a big data-based three-dimensional urban space reconstruction system, comprising:

[0038] Data collection module 201: collects a number of multi-angle spectral images of the target city space and the shooting information of the corresponding spectral images;

[0039] Graph network construction module 202: preprocessing the spectral image to obtain a preprocessed spectral image; extracting feature points in the preprocessed spectral image, mapping the feature points to nodes in the graph network, and constructing a graph network based on the nodes;

[0040] Model training module 203: collecting wide area data to train an improved adversarial neural network, the adversarial neural network includes a discriminator A, a generator A, a discriminator B and a random occlusion module;

[0041] Point cloud generation module 204: using the graph network of the target space to replace the preset random occlusion module, the graph network of the target space is reversely updated according to the evaluation value of the graph network output by the discriminator B, and the updated graph network of the target space is input into the generator A to obtain the point cloud data of the target space;

[0042] 3D reconstruction module 205: Modeling the target space based on the point cloud data of the target space.

[0043] In a third aspect, the present invention provides a computer device, comprising:

[0044] one or more processors;

[0045] The storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method described.

[0046] The beneficial effects of the present invention are as follows: through the collaborative work of the improved adversarial neural network and the graph network, it is possible to generate a high-quality, structurally sound three-dimensional model when the target space data is incomplete; by introducing a random occlusion module and a reverse update mechanism of the graph network, the data requirements and modeling costs are greatly reduced, making the solution more reasonable and widely adaptable in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of a method for three-dimensional reconstruction of urban space based on big data of the present invention;

[0048] Figure 2 It is an architecture diagram of a big data-based urban space three-dimensional reconstruction system of the present invention.

[0049] In the figure: 201, data acquisition module; 202, graph network construction module; 203, model training module; 204, point cloud generation module; 205, 3D reconstruction module. DETAILED DESCRIPTION

[0050] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that the discussion of these embodiments is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Each example may omit, replace or add various processes or components as needed. In addition, the features described relative to some examples may also be combined in other examples.

[0051] like Figure 1-2 As shown, a method and system for three-dimensional reconstruction of urban space based on big data, comprising:

[0052] Step S101: collecting a number of multi-angle spectral images of the target city space and the shooting information of the corresponding spectral images; the shooting information includes: camera principal point coordinates, optical coordinates and focal length;

[0053] Step S102: preprocessing the spectral image to obtain a preprocessed spectral image; extracting feature points in the preprocessed spectral image, mapping the feature points to nodes in a graph network, and constructing a graph network based on the nodes;

[0054] Step S103: Collect wide-area data to train an improved adversarial neural network, the adversarial neural network includes a discriminator A, a generator A, a discriminator B and a random occlusion module; the wide-area data includes a graph network of the training space and a graph network and point cloud data of the pan-space; wherein the point cloud data of the pan-space is used as a real sample A of the discriminator A, and the graph network of the pan-space is used as a real sample B of the discriminator B; part of the graph network of the training space is shielded by the random occlusion module, and the shielded graph network is respectively input into the discriminator B and the generator A, and the discriminator B outputs an evaluation value of the graph network; wherein the random occlusion module randomly changes the shielded area and position of the graph network of the training space; the generator A inputs the shielded graph network of the training space, and outputs the pseudo point cloud data of the training space, and the discriminator A outputs the evaluation value of the pseudo point cloud data;

[0055] It should be noted that pan-space refers to public spatial structures. For example, the complete point cloud data of a building in a famous scenic spot is open to the Internet. The graph network of the building can be obtained based on the abstraction of the point cloud data, such as retaining only the corner point cloud of the building point cloud data, or it can be independently produced based on public images. Through the complete image set of the building, the corners of the building are defined as nodes of the graph network; thus, a large number of real samples are obtained;

[0056] It should be noted that when modeling the target space, the target space, training space, and pan-space should be consistent in style;

[0057] Step S104: The preset random occlusion module is replaced by the graph network of the target space, and the graph network of the target space is reversely updated according to the evaluation value of the graph network output by the discriminator B. The updated graph network of the target space is input into the generator A to obtain the point cloud data of the target space;

[0058] Step S105: Modeling the target space based on the point cloud data of the target space.

[0059] In one embodiment of the present invention, preprocessing the spectral image to obtain a preprocessed spectral image includes:

[0060] A spectral image taken vertically from above is selected as a reference, a number of registration points are set in the spectral image, a non-reference spectral image is matched with the reference spectral image through the registration points, and the non-reference spectral image is corrected to obtain a multi-view spectral image set;

[0061] For example, define the reference spectral image as , and other spectral images are ;in, ; n represents the reference spectrum image In addition, there are a total of n spectral images; defined in Middle registration point The coordinates of ,exist Middle registration point The coordinates of ; Perform Fourier transform on the reference spectral image and the spectral image to be registered to obtain the frequency domain representation and , the phase correlation matrix is ​​calculated based on the Fourier transform results, and the formula is expressed as: ;in express The complex conjugate of express and The phase correlation matrix of Perform inverse Fourier transform and get and The offset between ;according to For images Perform translation correction to make With the reference image Alignment;

[0062] Extract feature edges and feature points from each spectral image in the multi-view spectral image set, and obtain the two-dimensional coordinates of the feature points; count the number of times each feature point appears in the multi-view spectral image set, and remove feature points below a threshold number of times;

[0063] For example, using the Canny edge detection algorithm, first perform Gaussian smoothing on the spectral image to reduce the impact of noise on edge detection. The formula is expressed as: ;in, The standard deviation is The Gaussian kernel of Represents the convolution operation; the Sobel operator is used to calculate the spectral image in the horizontal direction and vertical direction The gradient of Represents the horizontal gradient of the spectral image, that is, the brightness change rate of the pixel point of the spectral image in the horizontal direction (along the x-axis); Represents the vertical gradient of the spectral image, that is, the brightness change rate of the spectral image pixel in the vertical direction (along the y-axis); and Get the gradient magnitude and direction, suppress non-maximum values, ensure that only the local maximum value in the gradient direction is retained as the edge, use high and low thresholds to distinguish strong edges from weak edges, and connect these edges to get the final feature edge;

[0064] For example, take the Harris corner detection algorithm; calculate the gradient covariance matrix of each pixel in its neighborhood , the formula is:

[0065] ; where u represents the lateral displacement relative to the pixel point x in the neighborhood, and v represents the longitudinal displacement relative to the pixel point y in the neighborhood. and express and Directional spectral image gradient, Represents Gaussian weight; calculate the corner point response for each pixel in the spectral image to obtain , if the pixel Greater than threshold If a feature point is matched in 3 or more spectral images, it is considered as a valid feature point, otherwise it is removed.

[0066] Obtain disparity information through the position of the two-dimensional coordinates of the feature points in different spectral images, and convert the two-dimensional coordinates of the feature points into three-dimensional coordinates based on the two-dimensional coordinates of the feature points, the disparity information and the shooting information of the spectral images;

[0067] For example, the coordinates of the same feature point in two different spectral images are and , calculate the disparity , the formula is: ; According to the camera principal point coordinates, optical coordinates and focal length of each spectral image collected, the three-dimensional coordinates of the feature points are obtained by triangulation, and the formula is expressed as: , , ;in, and Represents the principal point coordinates of the camera; Represents the baseline distance between two spectral images, which is the Euclidean distance of the optical coordinates of the two spectral images; Indicates the focal length for taking spectral images;

[0068] In one embodiment of the present invention, constructing a graph network based on nodes includes:

[0069] Feature points are mapped to nodes of the graph network; if there is a feature edge between two nodes, the feature edge is mapped to the edge of the graph network; the initial feature of the node is the feature vector of the feature point, which includes the coordinates, color and reflectivity data of the feature point;

[0070] It should be noted that the extracted feature edge needs to match two valid feature points. If the feature edge matches the feature point, the feature edge is mapped to an edge. If the feature edge cannot match two valid feature points, the feature edge is removed.

[0071] In one embodiment of the present invention, the graph network of the target space is input into the discriminator B to obtain the evaluation value of the graph network by the discriminator B, and the graph network of the target space is reversely updated according to the evaluation value, including: calculating the weight of each node to other nodes and constructing a weight matrix;

[0072] Calculate the nodes in the first graph network Calculate the sum of any node The attention weight of the node is set to the set H, and the node is calculated For Node The weight of; the calculation formula is as follows:

[0073] ;in, and Respectively represent nodes and nodes The eigenvector of Represents a splicing operation; represents the weight vector; express Activation function; represents any node in the set H;

[0074] Perform cluster analysis on the weight matrix, set a weight factor threshold, distinguish feature points into high-weight feature points and low-weight feature points based on the weight factor threshold, and divide the graph network into high-weight areas and low-weight areas based on the high-weight feature points and low-weight feature points;

[0075] For example, node i is connected to 4 nodes, and the weight factor threshold is 0.5. is 0.6, is 0.7, is 0.6, is 0.4, and the weight factor of node i is 0.75, then node i is a high-weight feature point;

[0076] Insert a new node in the graph network area mapped in the low-weight area, and build edges between the new node and the high-weight feature points close to the new node;

[0077] Specifically, high-weight feature points are more important nodes in a certain area, connecting most of the low-weight feature points in the area;

[0078] Discriminator B outputs an evaluation value for the graph network in the target space. If the evaluation value is lower than the evaluation threshold, the position of the node is adjusted.

[0079] Preferably, the evaluation threshold is a custom threshold, preferably 0.3;

[0080] In one embodiment of the present invention, adjusting the position of a node includes:

[0081] Determine the reverse optimization factor of the graph network according to the evaluation value of the graph network by the discriminator B;

[0082] It should be noted that the output calculation formula of the discriminator A is:

[0083] ;in, Represents the evaluation value of the discriminator A graph network, represents the weight of the fully connected layer of the discriminator A, represents the bias of the fully connected layer of the discriminator A, The feature vector representing the point cloud data generated by the graph network is composed of the current node position Decide, express Activation function;

[0084] The calculation formula of the reverse optimization factor of the graph network is:

[0085] ;in, Represents the reverse optimization factor of the graph network;

[0086] Calculate the partial derivative of the inverse optimization factor with respect to the position of each new node;

[0087] For example, to calculate the partial derivative of the i-th newly inserted node, the calculation formula is as follows:

[0088] ;in, Represents the position of the reverse optimization factor for node i The partial derivative of Used to adjust the size of the partial derivative so that the partial derivative is inversely proportional to the evaluation value. Represents the output of the discriminator A The position of node i The partial derivative of

[0089] Update the position of node i to minimize the reverse optimization factor. The update formula is as follows:

[0090] ;in, represents the updated position of node i, represents the current position of node i, Used to control the size of each update step;

[0091] According to the direction of the partial derivative, move the new node in the direction that reduces the evaluation value of the discriminator B on the graph network;

[0092] Iteratively adjust the position of the new node until the evaluation value of the discriminator B on the graph network is greater than the evaluation threshold;

[0093] It should be noted that node i needs to repeatedly update its position until the evaluation value of the graph network does not change significantly or is greater than the evaluation threshold;

[0094] In one embodiment of the present invention, the nodes in the graph network are numbered and a generator matrix is ​​constructed; the size of the generator matrix is ​​D×K; wherein each row of the generator matrix represents the coordinates and features of a feature point;

[0095] Count the number of edges of feature points and set the dimension of each row of the generator matrix to K;

[0096] Fill in the coordinates, color, reflectivity and edge relationship of the feature points in sequence;

[0097] Set multiple identifiers, and use multiple identifiers to identify coordinates, colors, reflectivity and edge relationships respectively;

[0098] If the coordinates, color, reflectivity and edge relationship filling dimensions of the feature point are less than K dimensions, the feature point label and feature point weight are filled with masks, the identifier remains unchanged, and the edge relationship is expanded to K dimensions;

[0099] If the number of feature points is less than D, the number of rows of the generator matrix is ​​expanded to D dimensions, and the expanded number of rows is filled with masks;

[0100] It should be noted that generator A is composed of D stacked generator units, which jointly process the input graph network information. Each generator unit is responsible for processing one row of the generator matrix, and these generator units are interconnected through a fully connected layer to integrate information from different generator units. The input of generator A is a matrix of size D×K, where D represents the number of rows of the generator matrix, that is, the number of feature points in the graph network; K represents the number of columns of the generator matrix, that is, the dimensions of the coordinates, color, reflectivity, edge relationship and other information of each feature point.

[0101] Furthermore, the size of the generator matrix is ​​determined by the number of feature points D and the dimension K of each feature point information; the information of each feature point includes coordinates, color, reflectivity, and edge relationships with other feature points; if the actual feature point information dimension is not K, mask padding is used. If the number of feature points is less than D, the matrix is ​​also expanded by adding mask rows;

[0102] Exemplarily, any row of the generator matrix is ​​represented as:

[0103] ;in, , , , They are coordinate identifier, color identifier, reflectivity identifier and edge relationship identifier, respectively. Represents a mask, Represents dimension; , and Respectively represent the abscissa, ordinate and height coordinates of the three-dimensional coordinates; , and Respectively represent the red channel value, green channel value and blue channel value, and the value range of each channel value is between 0-255; Represents reflectivity; Both represent edge relationships.

[0104] In one embodiment of the present invention, the random occlusion module includes:

[0105] The random occlusion module randomly selects one or more regions with random positions and areas in the graph network of the training space for occlusion; occlusion includes occluding the nodes and edges of the graph network in the selected region, and the occluded nodes and edges will be removed from the graph network;

[0106] For example, a simple graph network includes:

[0107] Node collection: ;

[0108] Edge collection: ;

[0109] The random occlusion module decides to occlude node B and its related edges, and obtains:

[0110] Node collection: ; Edge set: ;

[0111] It should be noted that the main purpose of the random occlusion module is to simulate the data missing situation in the real world by occluding the complete graph network of the training space; the discriminator B evaluates the input graph network to evaluate the gap between its structure and the real spatial structure, and guides the structural update and optimization of the graph network through the back-propagation mechanism, so that the graph network gradually restores a reasonable structure; the generator A learns the structure of the graph network by inputting the occluded graph network to map high-quality point cloud data;

[0112] Specifically, the random occlusion module randomly selects one or more areas in the graph network for occlusion during the training process, and removes the nodes and edges in these areas; after receiving these incomplete graph networks, the generator A attempts to generate the corresponding point cloud data, and continuously adjusts and optimizes its generation strategy based on the feedback from the discriminator A; the discriminator B evaluates these occluded graph networks, calculates the reverse optimization factor and performs back propagation to promote the adjustment of the positions of the nodes and edges of the graph network, thereby restoring the occluded parts; through this design, the system can generate high-quality and well-structured point cloud data when faced with partially missing input data;

[0113] In one embodiment of the present invention, modeling the target space based on point cloud data of the target space includes:

[0114] Convert the point cloud data of the target space into a three-dimensional model;

[0115] Map the color and reflectivity in the features of the point cloud data to the surface of the three-dimensional model; obtain the three-dimensional model of the target space;

[0116] It should be noted that converting point cloud data into a three-dimensional model is a routine operation for ordinary technicians in this field and will not be described in detail here;

[0117] In one embodiment of the present invention, the present invention provides a big data-based three-dimensional urban space reconstruction system, comprising:

[0118] Data collection module 201: collects a number of multi-angle spectral images of the target city space and the shooting information of the corresponding spectral images;

[0119] Graph network construction module 202: preprocessing the spectral image to obtain a preprocessed spectral image; extracting feature points in the preprocessed spectral image, mapping the feature points to nodes in the graph network, and constructing a graph network based on the nodes;

[0120] Model training module 203: collecting wide area data to train an improved adversarial neural network, the adversarial neural network includes a discriminator A, a generator A, a discriminator B and a random occlusion module;

[0121] Point cloud generation module 204: using the graph network of the target space to replace the preset random occlusion module, the graph network of the target space is reversely updated according to the evaluation value of the graph network output by the discriminator B, and the updated graph network of the target space is input into the generator A to obtain the point cloud data of the target space;

[0122] 3D reconstruction module 205: Modeling the target space based on the point cloud data of the target space;

[0123] In one embodiment of the present invention, to model the target space, it is only necessary to obtain a rough spectral image of the target space, avoiding the acquisition of a large number of spectral images for the detail space, reducing the demand for marginal data for modeling the target space, and saving time and cost;

[0124] In one embodiment of the present invention, a computing device is provided, comprising:

[0125] one or more processors;

[0126] The storage device is used to store one or more programs. When the one or more programs are executed by the one or more processors, the one or more processors implement the method described.

[0127] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.

Claims

1. A method for three-dimensional reconstruction of urban space based on big data, characterized in that: include: Step S101: collecting a plurality of multi-angle spectral images of a target city space and corresponding shooting information of the spectral images; Shooting information includes: camera principal point coordinates, optical coordinates and focal length; Step S102: preprocessing the spectral image to obtain a preprocessed spectral image; extracting feature points in the preprocessed spectral image, mapping the feature points to nodes in a graph network, and constructing a graph network based on the nodes; Step S103: Collect wide-area data to train an improved adversarial neural network, the adversarial neural network includes a discriminator A, a generator A, a discriminator B and a random occlusion module; the wide-area data includes a graph network of the training space and a graph network and point cloud data of the pan-space; wherein the point cloud data of the pan-space is used as the real sample A of the discriminator A, and the graph network of the pan-space is used as the real sample B of the discriminator B; part of the graph network of the training space is shielded by the random occlusion module, and the shielded graph network is respectively input into the discriminator B and the generator A, and the discriminator B outputs an evaluation value of the graph network; wherein the random occlusion module randomly changes the shielded area and position of the graph network of the training space; the shielded graph network of the training space is input into the generator A, and the pseudo point cloud data of the training space is output, and the discriminator A outputs the evaluation value of the pseudo point cloud data; Step S104: The preset random occlusion module is replaced by the graph network of the target city space. The graph network of the target city space is reversely updated according to the evaluation value of the graph network output by the discriminator B. The updated graph network of the target city space is input into the generator A to obtain the point cloud data of the target city space. Step S105: Modeling the target city space based on the point cloud data of the target city space.

2. The method for three-dimensional reconstruction of urban space based on big data according to claim 1, characterized in that: Preprocessing the spectral image to obtain a preprocessed spectral image includes: A spectral image taken vertically from above is selected as a reference, a number of registration points are set in the spectral image, a non-reference spectral image is matched with the reference spectral image through the registration points, and the non-reference spectral image is corrected to obtain a multi-view spectral image set; In each spectral image, feature edges are extracted, feature points are detected and extracted, and the two-dimensional coordinates of the feature points are obtained; the number of times the feature points appear in the multi-view spectral image set is counted, and feature points with a number below a threshold are removed; The disparity information is obtained through the position of the two-dimensional coordinates of the feature points in different spectral images, and the two-dimensional coordinates of the feature points are converted into three-dimensional coordinates based on the two-dimensional coordinates of the feature points, the disparity information and the shooting information of the spectral images.

3. The method for three-dimensional reconstruction of urban space based on big data according to claim 1, characterized in that: Build a graph network based on nodes, including: Feature points are mapped to nodes of the graph network; if there is a feature edge between two nodes, the feature edge is mapped to the edge of the graph network; the initial feature of the node is the feature vector of the feature point, which includes the coordinates, color and reflectivity data of the feature point.

4. The method for three-dimensional reconstruction of urban space based on big data according to claim 1, characterized in that: The graph network of the target city space is reversely updated according to the evaluation value of the output graph network of the discriminator B, including: Calculate the weight of each node to other nodes and construct a weight matrix; Perform cluster analysis on the weight matrix, distinguish feature points into high-weight feature points and low-weight feature points based on the weight factor threshold, and divide the graph network into high-weight areas and low-weight areas based on the high-weight feature points and low-weight feature points; Insert a new node in the low-weight area and build edges between the new node and the high-weight feature points close to the new node; Discriminator B outputs an evaluation value for the graph network of the target city space. If the evaluation value is lower than the evaluation threshold, the position of the node is adjusted.

5. The method for three-dimensional reconstruction of urban space based on big data according to claim 4, characterized in that: Adjust the position of the node, including: Define the reverse optimization factor of the graph network according to the evaluation value of the discriminator B on the graph network; Calculate the partial derivative of the inverse optimization factor with respect to the position of each new node; According to the direction of the partial derivative, move the new node in the direction that reduces the evaluation value of the discriminator B on the graph network; The position of the new node is iteratively adjusted until the evaluation value of the discriminator B on the graph network is greater than the evaluation threshold.

6. The method for three-dimensional reconstruction of urban space based on big data according to claim 4, characterized in that: Label the nodes in the graph network and construct a generator matrix; the size of the generator matrix is ​​D×K; each row of the generator matrix represents the coordinates and features of a feature point; Count the number of edges of feature points and set the dimension of each row of the generator matrix to K; Fill in the coordinates, color, reflectivity and edge relationship of the feature points in sequence; Set multiple identifiers, and use multiple identifiers to identify coordinates, colors, reflectivity and edge relationships respectively; If the coordinates, color, reflectivity and edge relationship filling dimensions of the feature point are less than K dimensions, the feature point label and feature point weight are filled with masks, the identifier remains unchanged, and the edge relationship is expanded to K dimensions; If the number of feature points is ≤D, the number of rows of the generator matrix is ​​expanded to D dimensions, and the expanded number of rows is filled with masks.

7. The method for three-dimensional reconstruction of urban space based on big data according to claim 1, characterized in that: Random occlusion module, including: The random occlusion module randomly selects one or more regions with random positions and areas in the graph network of the training space for occlusion; the occluded area includes occluding the graph network nodes and edges within the selected area, and the occluded nodes and edges will be removed from the graph network.

8. The method for three-dimensional reconstruction of urban space based on big data according to claim 1, characterized in that: Modeling the target urban space based on the point cloud data of the target urban space includes: The point cloud data of the target city space is converted into a three-dimensional model; the color and reflectivity information in the point cloud data is mapped to the surface of the three-dimensional model; and the three-dimensional model of the target city space is obtained.

9. A big data-based 3D urban space reconstruction system, characterized in that: Use of the method according to any one of claims 1 to 8, comprising: Data acquisition module: collects a number of multi-angle spectral images of the target city space and the shooting information of the corresponding spectral images; Graph network construction module: preprocess the spectral image to obtain a preprocessed spectral image; extract feature points in the preprocessed spectral image, map the feature points to nodes in the graph network, and construct a graph network based on the nodes; Model training module: collect wide-area data to train an improved adversarial neural network, which includes discriminator A, generator A, discriminator B and random occlusion module; Point cloud generation module: The preset random occlusion module is replaced by the graph network of the target city space. The graph network of the target city space is reversely updated according to the evaluation value of the graph network output by the discriminator B. The updated graph network of the target city space is input into the generator A to obtain the point cloud data of the target city space. 3D reconstruction module: Model the target city space based on its point cloud data.

10. A computing device, characterized in that include: one or more processors; A storage device for storing one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method as claimed in any one of claims 1 to 8.

Images