System for boundary adaptive urban block layout and method of operation thereof

Abstract

A method of operation for a city block layout generation system operated by at least one processor, wherein a city block layout in which a plurality of buildings are arranged within a city block having arbitrary city block boundaries is obtained as training data, and a boundary graph for the city block boundaries and a building graph in which buildings are building nodes are obtained. A graph integration model is trained to infer spatial adjacency between the boundary nodes constituting the boundary graph and the building nodes to generate node connection information between the boundary nodes and the building nodes in a city block graph in which the boundary graph and the building graph are merged, and a building feature generation model is trained to encode the node features of the boundary graph and the building graph included in the city block layout, and to decode the encoding result to reconstruct the city block layout from the output building features.

Description

Technology Field

[0001] The present invention relates to a boundary-adaptive city block layout generation system and a method of operation thereof. Background Technology

[0002] Automatically generating cities using computers has been a long-standing research issue in academia and industry. Past computer-based city generation technologies exhibit the characteristics of rule-based generation, creating predefined rules based on expert knowledge and applying these rules sequentially to generate cities.

[0003] When generating cities using these conventional methods, the participation of experts with specialized knowledge in urban engineering and computer-based generation is essential. Furthermore, using a method that generates cities based on predefined rules has the disadvantage that the resulting output is relatively monotonous.

[0004] With the recent advancement of AI-based deep learning generative models, technology has emerged to generate city block layouts using generative models. While generating city block layouts using generative models can solve problems that require expert knowledge, it has the disadvantage that it is difficult to generate realistic city block layouts for arbitrary city blocks.

[0005] Furthermore, conventionally, data from real-world buildings was used as input. Consequently, if the building graph or city block boundaries are not defined, the generative model cannot generate the layout of the city blocks. Additionally, the generative model has the disadvantage that it can only create a city block layout if the city block boundaries are consistently defined. The problem to be solved

[0006] Accordingly, the present invention provides a boundary-adaptive city block layout generation system and a method of operation thereof that can generate a city block layout in which buildings are placed inside the boundaries, even when various types of block boundary graphs are input. means of solving the problem

[0007] A method of operation of a city block layout generation system operated by at least one processor for achieving the technical problem of the present invention comprises: a step of acquiring a city block layout as training data in which a plurality of buildings are arranged within a city block having an arbitrary city block boundary; a step of acquiring a boundary graph for the city block boundary and a building graph in which the buildings are building nodes; a step of training a graph integration model to infer spatial adjacency between the boundary nodes constituting the boundary graph and the building nodes, and to generate node connection information between the boundary nodes and the building nodes in a city block graph in which the boundary graph and the building graph are merged; and a step of encoding the node characteristics of the boundary graph and the building graph included in the city block layout, and training a building characteristic generation model to restore the city block layout from the building characteristics output by decoding the encoding result.

[0008] The method may further include the step of preprocessing the above city block layout to generate a city block graph, and obtaining node characteristics for building nodes and boundary nodes constituting the city block graph.

[0009] The above building feature generation model can be implemented as a variational autoencoder structure including an encoder and a decoder.

[0010] The step of training the building feature generation model can involve inputting the encoding result output from the encoder and latent variables for the boundary image mask transformed by the boundary graph into the decoder, and training the building feature generation model to restore the city block layout from the node features output from the decoder.

[0011] The step of training the graph integration model can infer spatial adjacency by connecting the arbitrary building node and the arbitrary boundary node if the distance between an arbitrary building node and an arbitrary boundary node is less than or equal to a preset reference distance and no other node exists between the arbitrary building node and the arbitrary boundary node.

[0012] A method of operation of a city block layout generation system operated by at least one processor, which is another feature of the present invention for achieving the technical problem of the present invention, comprises the steps of: obtaining a boundary graph and a building graph for a city block boundary of any shape; inferring spatial adjacency between a boundary node constituting the boundary graph and a building node constituting the building graph to obtain node connection information in which the nodes of the boundary graph and the building graph are connected; obtaining building characteristics of buildings to be located in a city block within the city block boundary from the node connection information; and generating a city block layout by placing buildings corresponding to the building characteristics in the city block.

[0013] The step of obtaining the above node connection information may include the step of generating a city block graph by merging the boundary graph and the building graph based on the spatial adjacency between the boundary nodes constituting the boundary graph and the building nodes constituting the building graph.

[0014] The step of generating the city block graph above allows obtaining node connection information based on spatial adjacency by connecting the arbitrary building node and the arbitrary boundary node if the distance between an arbitrary building node and an arbitrary boundary node is less than or equal to a preset reference distance and no other node exists between the arbitrary building node and the arbitrary boundary node.

[0015] After the step of obtaining the building characteristics, the method may further include the step of obtaining latent variables from the boundary image mask in which the boundary graph is transformed, and the step of generating the city block layout using the latent variables and the building characteristics.

[0016] As another feature of the present invention for achieving the technical problem of the present invention, a city block layout generation system comprises a processor that drives a building characteristic generation model and a graph integration model, wherein the processor acquires a city block layout in which a plurality of buildings are arranged within a city block having an arbitrary city block boundary as training data, and when the city block layout is input, trains an encoder and a decoder of the building characteristic generation model to restore the city block layout, and generates node connection information between the boundary node and the building node in a city block graph in which the boundary graph and the building graph are merged using a boundary graph for the city block boundary and a building graph in which the buildings are building nodes.

[0017] When the city block layout is input, the processor can encode the node characteristics of the building graph and the boundary graph through the encoder, and decode the encoding result through the decoder to restore the city block layout from the output building characteristics.

[0018] The above processor, wherein the building feature generation model further includes a convolutional neural network, and can obtain latent variables from a boundary image mask in which a boundary graph constituting the city block graph is transformed through the convolutional neural network.

[0019] The processor can infer spatial adjacency by connecting the arbitrary building node and the arbitrary boundary node if the distance between the arbitrary building node and the arbitrary boundary node is less than or equal to a preset reference distance and no other node exists between the arbitrary building node and the arbitrary boundary node.

[0020] The processor obtains a boundary graph and a building graph for a city block boundary of an arbitrary shape, infers spatial adjacency between a boundary node constituting the boundary graph and a building node constituting the building graph, obtains node connection information in which the nodes of the boundary graph and the building graph are connected, and obtains building characteristics of buildings to be located in the city block within the city block boundary from the node connection information, thereby generating a city block layout by placing buildings corresponding to the building characteristics in the city block.

[0021] When the boundary image mask, in which the boundary graph has been transformed, is input to the processor, the processor can obtain a latent variable from the boundary image mask. Effects of the invention

[0022] According to the present invention, even if various forms of city block boundaries are input, a city block layout can be created in which buildings are placed inside the boundaries in a specific form.

[0023] In addition, a boundary-adaptive city block layout generation system can efficiently generate realistic city environments in game or simulation settings, which can reduce content development costs and enhance the realism and immersion of the user experience.

[0024] In addition, realistic virtual city data generated through a boundary-adaptive city block layout generation system can be utilized as a dataset required for training autonomous vehicle technology, thereby having a direct impact on improving the quality of autonomous driving technology. Brief explanation of the drawing

[0025] Figure 1 is an example of a general city block layout generation technique. FIGS. 2 and FIGS. 3 are exemplary diagrams of a boundary-adaptive city block layout generation system according to an embodiment of the present invention. Figure 4 is an example diagram of a graph integration model according to an embodiment of the present invention. FIGS. 5 and 6 are exemplary diagrams of a building characteristic generation model according to an embodiment of the present invention. FIG. 7 is a flowchart of the operation method of a city block layout generation system according to an embodiment of the present invention. FIG. 8 is a flowchart of a learning method for a city block layout generation system according to an embodiment of the present invention. FIG. 9 is an exemplary diagram of a city block layout generation system according to an embodiment of the present invention. FIG. 10 is a flowchart of a method for generating a city block layout according to an embodiment of the present invention. FIG. 11 is an example diagram of an adjacency matrix of a boundary graph according to an embodiment of the present invention. FIG. 12 is a structural diagram of a computing system according to an embodiment of the present invention. Specific details for implementing the invention

[0026] Embodiments of the present invention are described below with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present invention may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present invention in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0027] Throughout the specification, when a part is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components.

[0028] A boundary-adaptive city block layout generation system and a method of operation thereof according to an embodiment of the present invention will be described below with reference to the drawings. Before describing the embodiment of the present invention, a conventional city block layout generation technology will first be described with reference to FIG. 1.

[0029] Figure 1 is an example of a general city block layout generation technique.

[0030] Figure 1 (a) is an example of a city block layout generated using BlockPlanner, and Figure 1 (b) is an example of a city block layout generated using GlobalMapper.

[0031] The block planner illustrated in Fig. 1(a) is designed so that it cannot receive information about city block boundaries into the city block layout generation model. Therefore, the city block layout generation model of the block planner can only process building graphs in the form of ring graphs.

[0032] Furthermore, the dataset used to train the Block Planner's urban block layout generation model consists solely of urban blocks from the real city of New York. Since all urban blocks in New York are rectangular, they fit well with the ring graph shape of the building graph assumed by the Block Planner. However, because the Block Planner does not receive information about urban block boundaries, it cannot generate various types of boundaries and the corresponding building distributions.

[0033] In addition, the global mapper shown in Fig. 1(b) uses a non-parametric algorithm called canonical spatial transformation to transform various shapes of city block boundaries into rectangular boundaries. Then, the global mapper places a building within the transformed rectangular boundary and performs an inverse canonical spatial transformation on the building and the boundary to restore the boundary to its original shape.

[0034] However, there are a significant number of boundaries to which standard spatial transformations do not apply, and since the Global Mapper places buildings of a consistent shape on all boundaries by default, it is difficult to say that the placement of buildings along city block boundaries is affected. Therefore, the Global Mapper has the disadvantage of being unable to place buildings suitable for various types of boundaries.

[0035] Accordingly, in an embodiment of the present invention, a boundary-adaptive city block layout generation system and a method of operation thereof are described, which use a graph merging model that represents buildings within a city block layout as graphs and infers the spatial density between nodes of two graphs to add edges between close nodes.

[0036] FIGS. 2 and FIGS. 3 are exemplary diagrams of a boundary-adaptive city block layout generation system according to an embodiment of the present invention.

[0037] FIG. 2 is an example diagram showing a boundary-adaptive city block layout generation system (hereinafter referred to as the "city block layout generation system" for convenience of explanation) (100) training an artificial intelligence model mounted therein. FIG. 3 is an example diagram showing that when a city block graph and a building graph are input using the trained artificial intelligence model, the city block layout generation system (100) provides a city block layout.

[0038] In the embodiments of the present invention, for convenience of explanation, the city block layout generation system (100) is described as training an artificial intelligence model. However, a separate artificial intelligence model training device (not shown in the drawing) may train an artificial intelligence model using training data, and the trained artificial intelligence model may be loaded into the city block layout generation system (100) to be used for generating a city block layout.

[0039] First, as illustrated in FIG. 2, the city block layout generation system (100) trains an artificial intelligence model to output a city block layout when a city block layout, which is training data, is input. Additionally, the city block layout generation system (100) trains an artificial intelligence model to output connection information between boundaries and buildings when a building graph and a boundary graph are input.

[0040] That is, as illustrated in FIG. 2, when a building graph and a boundary graph are input, a graph integration model (110), which is a transformer-based deep neural network, infers the spatial adjacency between the building nodes constituting the building graph and the boundary nodes constituting the boundary graph. Then, based on the inferred spatial adjacency, the graph integration model (110) connects adjacent nodes with edges to merge the building graph and the boundary graph, and outputs connection information between the boundary nodes and building nodes of the merged graph (hereinafter referred to as 'node connection information' for convenience of explanation).

[0041] Here, the building graph and boundary graph are explained using the example that building graphs and boundary graphs for each of the actual major cities around the world have already been formed. Since the graph integration model (110) can obtain building graphs and boundary graphs in various ways, the embodiment of the present invention is not limited to any one method.

[0042] The graph integration model (110) connects each boundary node constituting the boundary graph to form a single sequence and outputs it. Then, the graph integration model (110) is trained to output node connection information based on the single sequence output and the building graph extracted from the city block graph.

[0043] That is, the graph integration model (110) performs cross attention on the building graph and the sequence to integrate the building graph and the boundary graph to generate a city block graph. The integrated city block graph consists of building nodes, boundary nodes, and node connection information between the building nodes and the boundary nodes.

[0044] At this time, the edges connecting boundary nodes and the edges connecting building nodes in the node connection information of the city block graph use the information included in the input boundary graph and building graph as is. That is, the graph integration model (110) is explained by taking the example of learning the connection information between boundary nodes and building nodes and providing it as node connection information.

[0045] Additionally, when urban block layouts for actual major cities are input as training data, a building feature generation model (120) composed of an encoder (121) and a decoder (122) and a variational autoencoder model is trained to output the urban block layout.

[0046] In this case, the city block layout includes multiple buildings located within the city block and the city block boundary, and the explanation is given using examples of city block layouts for major cities around the world that have already been created.

[0047] When a city block layout is input to the building characteristic generation model (120), the encoder (121) encodes the boundary graph and the node characteristics of the building graph and outputs them as a 512-dimensional tensor. Here, the encoding result is generated with the same configuration or form as the node connection information generated in the graph integration model (110), and the configuration or form is not limited to any one.

[0048] Then, the decoder (122) receives the encoding result from the encoder (121), decodes the received encoding result to identify building characteristics, and then outputs the city block layout.

[0049] To this end, the building feature generation model (120) preprocesses the input city block layout to extract a city block graph. Here, the city block graph consists of building nodes representing buildings located inside the city block, boundary nodes representing the boundaries surrounding the city block, and edges connecting two nodes when the distance between a building node and a boundary node, between a building node and a building node, or between a boundary node and a boundary node is less than or equal to a certain distance. Boundary nodes refer to boundary segments that divide the city block boundaries of the city block layout into equal lengths.

[0050] When the building characteristic generation model (120) inputs the node characteristics of each building node and boundary node constituting the city block graph and the city block graph to the encoder (121), the encoder (121) encodes the node characteristics of the building node and boundary node and outputs an encoding result. Then, the decoder (122) decodes the encoding result and outputs building characteristics, and based on the output building characteristics, the building characteristic generation model (120) restores the city block layout received.

[0051] Here, building characteristics include building node characteristics and boundary node characteristics, and multiple nodes constituting the city block graph each have node characteristics. In the embodiment of the present invention, the explanation is given by exemplifying that each node has six characteristics (c, x, y, h, w, θ).

[0052] To explain each characteristic, category (c) indicates whether each node is a building node or a boundary node. x and y represent the x-coordinate (x) and y-coordinate (y) of each node in a two-dimensional space. h and w represent the height (h) and width (w) of the building bounding box or boundary node representing the building, and θ represents the slope (θ) of the building bounding box and boundary node with respect to the x-axis. In the embodiment of the present invention, the node characteristics are described as including six characteristics, but other characteristics may be included.

[0053] At this time, the building characteristic generation model (120) may receive as input a boundary image mask formed by transforming a boundary graph included in the city block graph into an image form. The building characteristic generation model (120) processes the boundary image mask to obtain a boundary image latent variable. Since the method of the building characteristic generation model (120) receiving the boundary image mask as input or processing it to obtain the boundary image latent variable can be executed in various ways, the embodiment of the present invention is not limited to any one method.

[0054] In this way, the decoder (122), which receives the encoding result or the encoding result and the boundary image latent variable as input, decodes the encoding result or the encoding result and the boundary image latent variable and outputs building node characteristics. Then, the decoder (122) uses the building node characteristics to create and output a city block layout.

[0055] After the graph integration model (110) and building characteristic generation model (120) mounted on the city block layout generation system (100) are learned, as shown in FIG. 3, the city block layout generation system (100) receives a building graph and a boundary graph as input and generates a city block layout based on them.

[0056] That is, the graph integration model (110) of the city block layout generation system (100) generates node connection information when a boundary graph and a building graph for an arbitrary city block boundary are input. The node connection information generated by the graph integration model (110) is input to the decoder (122) of the building characteristic generation model (120), and the decoder (122) generates a city block layout by placing buildings within an arbitrary city block boundary.

[0057] The graph integration model (110) and building characteristic generation model (120) constituting the city block layout generation system (100) will be explained with reference to FIGS. 4 to 6.

[0058] Figure 4 is an example diagram of a graph integration model according to an embodiment of the present invention.

[0059] As illustrated in FIG. 4, the graph integration model (110) learns the relationship between the building graph and the boundary graph, which are input graphs. In an embodiment of the present invention, the graph integration model (110) is described as being implemented as a transformer network. The structure of the transformer network is already known, so a detailed description is omitted in the embodiment of the present invention.

[0060] When a boundary graph is input, the transformer encoder (111) of the graph integration model (110) outputs a sequence for the boundary graph as a tensor. In an embodiment of the present invention, the transformer encoder (111) is described as an example of outputting a sequence for the boundary graph as a 512-dimensional tensor.

[0061] When a sequence generated by connecting information for each boundary node is input, the transformer encoder (111) encodes the sequence and outputs it. The method by which the transformer encoder (111) encodes the sequence is a known technique, and the embodiment of the present invention is not limited to any one of these methods.

[0062] And, the transformer decoder (112) of the graph integration model (110) receives the sequence and building graph, which are the outputs of the transformer encoder (111), as inputs and outputs node connection information.

[0063] That is, the transformer decoder (112) performs cross attention between the building graph and the sequence which is the output of the transformer encoder (111) to generate node connection information of the city block graph in which the building graph and the boundary graph are integrated. The technique of the transformer decoder (112) performing cross attention between the building graph and the encoded sequence to generate node connection information is a known technique, and therefore, a detailed description is omitted in the embodiment of the present invention.

[0064] As such, when a building graph and a boundary graph are input to the graph integration model (110), it infers the spatial adjacency between the two graphs through a transformer-based deep neural network. Then, based on the inferred spatial adjacency, it outputs node connection information of a city block graph formed by merging the two graphs.

[0065] FIGS. 5 and 6 are exemplary diagrams of a building characteristic generation model according to an embodiment of the present invention.

[0066] In an embodiment of the present invention, when generating building characteristics for positioning a building inside a city block, node characteristics obtained from a block graph may be used, or boundary image latent variables obtained from a boundary image mask may be used together with node characteristics. To this end, in an embodiment of the present invention, FIG. 5 shows a first embodiment using only node characteristics, and FIG. 6 shows a second embodiment using boundary image latent variables together with node characteristics.

[0067] First, as illustrated in FIG. 5, the building characteristic generation model (120) extracts a city block graph and node characteristics from the input city block layout.

[0068] Building nodes and boundary nodes constituting the city block graph each have node characteristics, and in the embodiment of the present invention, six characteristics (c, x, y, h, w, θ) are described as examples. Accordingly, the building characteristic generation model (120) can identify node characteristics from a city block graph topology composed of building nodes representing buildings located inside the city block, boundary nodes for the boundary surrounding the city block, and edges connecting between building nodes and boundary nodes, between building nodes and building nodes, or between boundary nodes and boundary nodes.

[0069] The building feature generation model (120) maps node features into a latent space and then restores the input city block layout. To this end, the building feature generation model (120) can be implemented as a variational autoencoder that learns by mapping the input node features and the city block graph into a high-dimensional space and receiving the result from the latent space to restore the city block layout.

[0070] Here, the variational autoencoder model is trained to receive real data as input to the encoder during training, map it to a latent space, and restore it through the decoder, while simultaneously training the distribution of the latent space to approximate a normal distribution. Ideally, the variational autoencoder model has an encoder (121) capable of mapping real data to a latent distribution that matches a normal distribution, and a decoder (122) capable of restoring a vector on the latent distribution to be identical to the input data. Therefore, during actual inference, even if an arbitrary vector is sampled from the normal distribution, the decoder (122) can generate a new data sample that belongs to the real data distribution.

[0071] Accordingly, the building feature generation model (120) inputs the node features and the city block graph into the encoder (121). The encoder (121) operates only during training and does not operate when inferring node features.

[0072] When the encoder (121) receives the city block graph topology and node characteristics of the city block graph as input, it maps the node characteristics to a 512-dimensional latent space. Then, the encoder (121) outputs an approximation value that approximates the latent space to a normal distribution. In an embodiment of the present invention, the approximation value is referred to as connection relationship information, which is the connection relationship between building nodes and boundary nodes.

[0073] The decoder (122) of the building feature generation model (120) operates during both the learning process and the actual inference process.

[0074] During learning, the decoder (122) receives the encoding result, which is the output of the encoder (121), as input, decodes it, and outputs the building characteristics {x, y, w, h, θ} of the city block graph. Then, based on the building characteristics output from the decoder (122), it restores a city block layout identical to the input city block layout.

[0075] Thus, in an embodiment of the present invention, the building characteristic generation model (120) utilizes spatial information of the city block boundary to determine the spatial distribution of buildings within the city block. Through this, the city block layout generation system (100) can generate a realistic city block layout.

[0076] That is, the city block layout generation system (100) does not simply use only the geometric characteristics of the building, but infers the city block boundaries where the building is likely to be spatially adjacent based on the building characteristic generation model (120). And the building characteristic generation model (120) can infer the building characteristics of each building based on the inferred positional relationships.

[0077] Meanwhile, as illustrated in FIG. 6, the decoder (122) of the building feature generation model (120) can output building features by decoding the latent variable output from the convolutional neural network (Resnet34) (123) that processes the boundary image mask and the encoding result of encoding the boundary graph and the node features of the building graph in the encoder (121).

[0078] The convolutional neural network (123) receives a boundary image generated from a boundary graph as input and generates a latent variable. Here, the convolutional neural network (123) generates a boundary image latent variable by processing a boundary image mask with a known structure, and a detailed description is omitted in the embodiment of the present invention.

[0079] Although spatial information is included in the boundary graph, the purpose is to recognize spatial characteristics more accurately by additionally using a convolutional neural network (123).

[0080] Next, the operation method of the city block layout generation system (100) is described with reference to FIG. 7. Then, the method of the city block layout generation system (100) learning artificial intelligence models is described with reference to FIG. 8 and FIG. 9.

[0081] FIG. 7 is a flowchart of the operation method of a city block layout generation system according to an embodiment of the present invention.

[0082] As illustrated in FIG. 7, the city block layout generation system (100) acquires the city block layouts of actual major cities as training data. When the city block layout is input into the building characteristic model (120), the city block layout generation system (100) trains the building characteristic model (120) to restore the city block layout (S100). The city block layout generation system (100) can collect the city block layouts in various ways, and the embodiments of the present invention are not limited to any one method.

[0083] Additionally, the city block layout generation system (100) acquires boundary graphs and building graphs of actual major cities, and when the acquired graphs are input into the graph integration model (110), the graph integration model (110) is trained to output node connection information from the graph integration model (110) (S110). Here, the node connection information refers to information representing the boundary nodes, building nodes, and edges connecting the building nodes and boundary nodes of the city block graph generated by integrating the building graph and the boundary graph.

[0084] After training the building characteristic generation model (120) and the graph integration model (110) in this manner, when an arbitrary city block boundary graph and a building graph are input to the city block layout generation system (100), the graph integration model (110) generates node connection information.

[0085] Then, the generated node connection information is input into the decoder (122) of the building characteristic generation model (120), and a city block layout in which buildings are distributed within an arbitrary city block is generated from the decoder (122) (S120).

[0086] At this time, with reference to FIGS. 8 and 9, a method for the city block layout generation system (100) to train the graph integration model (110) and the building characteristic generation model (120) using training data in step S100 is described.

[0087] FIG. 8 is a flowchart of a learning method for a graph integration model according to an embodiment of the present invention.

[0088] As illustrated in FIG. 8, when the city block layout generation system (100) receives a building graph and a boundary graph (S201), it inputs the received building graph and boundary graph into a graph integration model (110).

[0089] The graph integration model (110) infers spatial adjacency between the building graph and the boundary graph (S202). In an embodiment of the present invention, if the distance between a building node included in the building graph and a boundary node included in the boundary graph is shorter than a preset distance and there is no other node between the building node and the boundary node, spatial adjacency is inferred by connecting the building node and the boundary node with an edge, but is not necessarily limited to this.

[0090] In this way, the graph integration model (110) is trained to infer spatial adjacency between two graphs and then output node connection information of a city block graph that integrates the building graph and the boundary graph (S203).

[0091] FIG. 9 is a flowchart of a learning method for a building characteristic generation model according to an embodiment of the present invention.

[0092] As illustrated in FIG. 9, when the city block layout generation system (100) receives a city block layout as training data (S301), it preprocesses the city block layout to extract a city block graph (S302). The extracted city block graph is input into a building characteristic generation model (120).

[0093] At this time, the building characteristic generation model (120) also receives node characteristics extracted from the city block graph as input (S303). The building nodes and boundary nodes constituting the city block graph each have node characteristics, and in the embodiment of the present invention, it is explained by using the example that there are six characteristics (c, x, y, h, w, θ).

[0094] The building characteristic generation model (120) maps the node characteristics into a latent space (S304), and then encodes the city block graph and the node characteristics to obtain connection relationship information, which is the connection relationship between the building nodes and the boundary nodes (S305).

[0095] Meanwhile, the building characteristic generation model (120) can receive a modified boundary image mask from the boundary graph (S306). The building characteristic generation model (120) extracts boundary image latent variables from the received boundary image mask (S307). Since there are various methods for the building characteristic generation model (120) to extract boundary image latent variables from the boundary image mask, the embodiment of the present invention is not limited to any one method.

[0096] At this time, the building feature generation model (120) may omit the step of extracting boundary image latent variables from a boundary image mask generated from a boundary graph during training and using the extracted boundary image latent variables in training.

[0097] The building characteristic generation model (120) obtains building characteristics by decoding the encoding result obtained in step S305 (S308). If boundary image latent variables are extracted from the boundary image, the building characteristic generation model (120) may additionally use the latent variables extracted in step S307 when obtaining building characteristics in step S308.

[0098] The building characteristic generation model (120) obtains building characteristics in step S308 and then restores the city block layout received in step S301 (S309). The method by which the building characteristic generation model (120) restores the city block layout is

[0099] The following describes an example of a city block layout generation system (100) including a learned graph integration model (110) and a building property generation model (120) with reference to FIG. 10.

[0100] FIG. 10 is an exemplary diagram of a city block layout generation system according to an embodiment of the present invention.

[0101] FIG. 10 (a) shows an example of generating a city block layout based on a boundary graph and a building graph when input to a city block layout generation system (100). FIG. 10 (b) shows an example of generating a city block layout based on a boundary image latent variable generated from a boundary image mask generated based on a city block boundary, a city block boundary, and building generation request information when input to a city block layout generation system (100) when a boundary graph and a building graph are input.

[0102] First, as illustrated in FIG. 10 (a), the graph integration model (110) receives a boundary graph and a building graph of any shape, merges the two graphs, and then generates node connection information.

[0103] Node connection information generated by the graph integration model (110) is input to the decoder (122) of the building characteristic generation model (120). The learned decoder (122) identifies building characteristics based on the node connection information and generates a city block layout in which buildings are positioned within a city block of an arbitrary shape based on the identified building characteristics.

[0104] Meanwhile, as illustrated in Fig. 10 (b), the graph integration model (110) receives a boundary graph and a building graph of any shape, merges the two graphs, and then generates node connection information.

[0105] The convolutional neural network (123) of the building feature generation model (120) receives a boundary image mask, which is a boundary graph received from the graph integration model (110) transformed into an image, as input and generates a boundary image latent variable. Since the boundary graph contains spatial information, the performance of the city block layout generation system (100) can be improved by generating a boundary image mask that transforms this information into an image form and generating a boundary image latent variable from the generated boundary image mask.

[0106] Node connection information generated by the graph integration model (110) is input to the decoder (122) of the building characteristic generation model (120). The learned decoder (122) identifies building characteristics based on the node connection information and generates a city block layout in which buildings are positioned within a city block of an arbitrary shape based on the building characteristics and boundary image latent variables.

[0107] Next, a method for generating a city block layout using a city block layout generation system (100) trained with an artificial intelligence model is described with reference to FIG. 11.

[0108] FIG. 11 is a flowchart of a method for generating a city block layout according to an embodiment of the present invention.

[0109] As illustrated in FIG. 11, when the graph integration model (110) of the city block layout generation system (100) receives a boundary graph and a building graph of any shape (S400), it integrates the boundary graph and the building graph (S401). To do this, the graph integration model (110) uses the building nodes included in the building graph and the boundary nodes included in the boundary graph to check whether the distance between the building nodes and the boundary nodes is shorter than a pre-set distance.

[0110] And, the graph integration model (110) connects an edge between two nodes, that is, between a building node and a boundary node, or between two building nodes, if no other node exists between them. Through this spatial adjacency inference procedure of two graphs, the graph integration model (110) obtains node connection information (S402).

[0111] The node connection information obtained by the graph integration model (110) is transmitted to the decoder (122) of the building characteristic generation model (120). The decoder (122) checks the building characteristics based on the obtained node connection information (S403), and uses the building characteristics to generate a city block layout by arranging buildings within an arbitrary city block according to the building characteristics (S404).

[0112] FIG. 12 is a structural diagram of a computing system according to an embodiment of the present invention.

[0113] Referring to FIG. 12, a city block layout generation system (100) operated by at least one processor may be implemented as a computing system (200) and executes a program containing instructions described to execute the operation of the present invention in the computing system (200). The program may be stored in a computer-readable storage medium and may be distributed.

[0114] The hardware of the computing system (200) may include at least one processor (210), memory (220), storage (230), and communication interface (240), and may be connected via a bus. In addition, hardware such as input devices and output devices may be included. The computing system (200) may be equipped with various software, including an operating system capable of running programs.

[0115] The processor (210) is a device that controls the operation of the computing system (200) and may be a various type of processor that processes instructions included in a program, for example, a CPU (Central Processing Unit), MPU (Micro Processor Unit), MCU (Micro Controller Unit), GPU (Graphic Processing Unit), etc.

[0116] The memory (220) loads the corresponding program so that instructions described for executing the operation of the present invention are processed by the processor (210). The memory (220) may be, for example, ROM (read only memory), RAM (random access memory), etc. The storage (230) stores various data, programs, etc. required to execute the operation of the present invention. The communication interface (240) may be a wired / wireless communication module.

[0117] Although embodiments of the present invention have been described in detail above, the scope of the present invention is not limited thereto, and various modifications and improvements by those skilled in the art using the basic concept of the present invention as defined in the following claims also fall within the scope of the present invention.

Claims

Claim 1 A method of operation of a city block layout generation system operated by at least one processor, comprising: acquiring a city block layout as training data in which a plurality of buildings are arranged within a city block having an arbitrary city block boundary; acquiring a boundary graph for the city block boundary and a building graph in which the buildings are building nodes; training a graph integration model to infer spatial adjacency between a boundary node constituting the boundary graph and a building node to generate node connection information between the boundary node and the building node in a city block graph in which the boundary graph and the building graph are merged; and training a building feature generation model to encode the node features of the boundary graph and the building graph included in the city block layout, and to decode the encoding result to restore the city block layout from the output building features. Claim 2 A method of operation according to claim 1, further comprising the step of preprocessing the city block layout to generate a city block graph, and obtaining node characteristics for building nodes and boundary nodes constituting the city block graph. Claim 3 A method of operation according to claim 1, wherein the building characteristic generation model is implemented as a variational autoencoder structure including an encoder and a decoder. Claim 4 In paragraph 3, the step of training the building feature generation model comprises inputting the encoding result output from the encoder and latent variables for the boundary image mask converted from the boundary graph into the decoder, and training the building feature generation model to restore the city block layout from the node features output from the decoder. Claim 5 In claim 1, the step of training the graph integration model is a method of operation in which, if the distance between an arbitrary building node and an arbitrary boundary node is less than or equal to a preset reference distance and no other node exists between the arbitrary building node and the arbitrary boundary node, the spatial adjacency is inferred by connecting the arbitrary building node and the arbitrary boundary node. Claim 6 A method of operation of a city block layout generation system operated by at least one processor, comprising: a step of obtaining a boundary graph and a building graph for a city block boundary of any shape; a step of inferring spatial adjacency between a boundary node constituting the boundary graph and a building node constituting the building graph to obtain node connection information in which the nodes of the boundary graph and the building graph are connected; a step of obtaining building characteristics of buildings to be located in a city block within the city block boundary from the node connection information; and a step of generating a city block layout by placing buildings corresponding to the building characteristics in the city block. Claim 7 A method of operation according to claim 6, wherein the step of obtaining node connection information includes the step of generating a city block graph by merging the boundary graph and the building graph based on the spatial adjacency between the boundary nodes constituting the boundary graph and the building nodes constituting the building graph. Claim 8 In claim 7, the step of generating the city block graph is a method of operation in which, if the distance between an arbitrary building node and an arbitrary boundary node is less than or equal to a preset reference distance and no other node exists between the arbitrary building node and the arbitrary boundary node, the arbitrary building node and the arbitrary boundary node are connected to obtain the node connection information according to the spatial adjacency. Claim 9 A method of operation according to claim 6, further comprising, after the step of obtaining the building characteristics, the step of obtaining a latent variable from a boundary image mask in which the boundary graph is transformed, and the step of generating the city block layout using the latent variable and the building characteristics. Claim 10 A city block layout generation system comprising a processor for driving a building feature generation model and a graph integration model, wherein the processor acquires a city block layout in which a plurality of buildings are arranged within a city block having an arbitrary city block boundary as training data, acquires a boundary graph for the city block boundary and a building graph in which the buildings are building nodes, infers spatial adjacency between the boundary nodes constituting the boundary graph and the building nodes, trains a graph integration model to generate node connection information between the boundary nodes and the building nodes in a city block graph in which the boundary graph and the building graph are merged, encodes the node features of the boundary graph and the building graph included in the city block layout, and trains a building feature generation model to restore the city block layout from the building features output by decoding the encoding result. Claim 11 In claim 10, the above processor preprocesses the city block layout to generate a city block graph and obtains node characteristics for building nodes and boundary nodes constituting the city block graph, a city block layout generation system. Claim 12 In claim 11, the processor is a city block layout generation system in which the building characteristic generation model is implemented as a variational autoencoder structure including an encoder and a decoder. Claim 13 A city block layout generation system according to claim 12, wherein the processor inputs the encoding result output from the encoder and latent variables for the boundary image mask converted from the boundary graph into the decoder, and trains the building feature generation model to restore the city block layout from the node features output from the decoder. Claim 14 A city block layout generation system according to claim 13, wherein the processor infers spatial adjacency by connecting the arbitrary building node and the arbitrary boundary node when the distance between the arbitrary building node and the arbitrary boundary node is less than or equal to a preset reference distance and no other node exists between the arbitrary building node and the arbitrary boundary node. Claim 15 A city block layout generation system according to claim 14, wherein the processor obtains a boundary graph and a building graph for a city block boundary of any shape, infers spatial adjacency between a boundary node constituting the boundary graph and a building node constituting the building graph, obtains node connection information in which the nodes of the boundary graph and the building graph are connected, obtains building characteristics of buildings to be located in the city block within the city block boundary from the node connection information, and then generates a city block layout by placing buildings corresponding to the building characteristics in the city block.

Images