Building individualization information extraction method, computer device, and storage medium

Abstract

The application discloses a kind of building monomerization information extraction method, computer device and storage medium, obtain the building two-dimensional orthographic image with geographic coordinate information and corresponding three-dimensional point cloud data;The instance segmentation model based on deep learning obtains the approximate outline surface of building in two-dimensional orthographic image;The approximate outline surface is converted into the outline surface of real geographic coordinates;The three-dimensional point cloud set is obtained by cutting the three-dimensional point cloud data based on the outline surface obtained in step S3;Separate and save multiple buildings existing in the three-dimensional point cloud set, based on the spatial geometric structure of building, carry out fine processing to house, extract the attribute information of building.The application breaks through the technical bottleneck of the use of oblique photography data, realizes the rapid and intelligent physical monomerization of building in oblique photography model and point cloud model, greatly reduces the labor cost, improves the utilization rate of oblique photography data, improves production efficiency.

Description

Technical Field

[0001] This invention relates to the field of oblique photogrammetry three-dimensional modeling, and in particular to a method, computer device and storage medium for extracting individual building information. Background Technology

[0002] In projects such as land acquisition and resettlement, smart cities, smart construction sites, land planning, and engineering quantity surveying, the statistics on building height, area, and material structure mainly rely on extensive on-site surveys and measurements by numerous field researchers, or on manual delineation using satellite imagery, making statistical data extremely difficult. Therefore, building-specific analysis based on oblique photogrammetry is one of the core technological directions in current engineering.

[0003] Oblique photogrammetry 3D modeling has become an important tool for large-scale, large-scene 3D reconstruction due to its advantages of high efficiency, high realism, and low production cost. However, due to limitations in the oblique photogrammetry data production process and its data protocol barriers, the oblique photogrammetry model is a continuous TIN triangulation network with no attribute information, whose surface is covered by high-resolution imagery. Surface objects in the data cannot be selected, manipulated, or managed individually, and GIS operations such as attribute querying, spatial querying, and thematic map creation cannot be implemented, reducing the value and practicality of the model data.

[0004] Current oblique photogrammetry techniques for individualization can be mainly divided into the following three categories:

[0005] (1) Three-dimensional individualization techniques based on dynamic individualization, ID individualization, and segmented individualization. Among them, ID individualization involves manually attaching model attributes and selecting regions during the oblique photogrammetry model generation process, so that the generated model can be selected and queried when used. Dynamic individualization and segmented individualization involve manually segmenting individuals in the oblique photogrammetry model after data production through methods such as region selection and 3D clipping, so as to further realize the calculation and query of individual attributes. However, three-dimensional individualization techniques require a lot of manual participation, with a lot of repetitive work, low efficiency, and high error rate, and cannot perform autonomous segmentation and attribute calculation for individuals of interest.

[0006] (2) Artificial intelligence-based individual building technology based on 2D orthophotos. This type of method has problems such as poor classification effect, difficulty in distinguishing between ground and buildings, difficulty in identifying individual buildings in dense building groups, and inability to obtain 3D attribute information of buildings when facing complex scenes such as blurred ground boundaries, dense buildings, and large variations in building size. This type of method cannot effectively utilize 3D structural information. For example, it can only extract the area of ​​the top of the eaves of a house, but not the area of ​​the bottom of the house; the height information of the house cannot be obtained; it is difficult to distinguish between cement roofs and cement grounds when their features are similar. Moreover, traditional algorithms will result in the segmentation and recognition of dense and messy houses in rural areas, making it difficult to individualize them.

[0007] (3) Artificial intelligence-based individualization technology based on 3D point clouds. Currently, point cloud semantic segmentation technologies such as RandLA-Net, which are applicable to large-scale scenarios, have emerged. However, due to current technological limitations, the accuracy of 3D semantic segmentation is lower than that of 2D image semantic segmentation. This type of method is relatively applicable to regular urban buildings, but has poor robustness to reservoir areas and rural areas. The accuracy of 3D point cloud semantic segmentation algorithms in large-scale scenarios is low, and the individualization effect of buildings is poor.

[0008] CN109492606A, "Multispectral Vector Image Acquisition Method and System, 3D Individualization Method and System," utilizes deep learning to identify remote sensing images and extract vector contour maps of various land features. However, remote sensing images typically have low accuracy, and the extracted building edge contours can only approximate the data. When cutting 3D oblique photogrammetry models, redundant noise information is easily introduced, resulting in low individualization accuracy. While individualization is achieved by cutting the oblique photogrammetry model using the obtained contour vector images, each land feature attribute requires manual entry into the database, thus failing to achieve true automatic individualization. Summary of the Invention

[0009] The technical problem to be solved by the present invention is to provide a method, computer device and storage medium for extracting individual building information, thereby improving the accuracy of automatic individual building extraction, in order to address the shortcomings of the existing technology.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for extracting individual building information, comprising the following steps:

[0011] S1. Obtain two-dimensional orthophotos of buildings with geographic coordinate information and corresponding three-dimensional point cloud data;

[0012] S2. A deep learning-based instance segmentation model obtains the approximate outline of a building in a two-dimensional orthophoto.

[0013] S3. Convert the approximate outline surface into an outline surface with real geographic coordinates.

[0014] S4. Based on the contour surface obtained in step S3, cut the three-dimensional point cloud data to obtain a three-dimensional point cloud set;

[0015] S5. Separate and save multiple buildings existing in the three-dimensional point cloud set, and refine the buildings based on their spatial geometry to extract their attribute information.

[0016] This invention, based on 2D orthophotos and corresponding 3D point cloud data, designs a 2D / 3D coupled unitization technique to achieve efficient and accurate positioning and unitization of buildings in a 3D point cloud model. The invention employs an instance segmentation algorithm for building positioning and contour extraction, fully utilizing the refined representation information of building contours in the orthophoto data. The high efficiency and high precision of the 2D instance segmentation algorithm ensure the accuracy of building positioning and contour extraction. Addressing the limitation of 2D orthophotos lacking spatial structural information, this invention uses 3D techniques to deeply mine the structural information of the 3D point cloud set of the building obtained through coordinate mapping of the 2D contour surface, completing the attribute interpretation of the building in 3D space. This invention considers the characteristics of both 2D and 3D data, integrating the advantages of both 2D and 3D technologies to achieve accurate unitization of buildings and improve the accuracy of automatic unitization.

[0017] The specific implementation process of step S2 includes:

[0018] S2.1. Cropping the two-dimensional orthophoto into sub-images of size M*M; where M is a constant;

[0019] S2.2. Import the cropped sub-images obtained in S2.1 into the labelme software, add label information to each image, and construct a building instance segmentation dataset; based on the deep learning framework, construct the Mask_Rcnn instance segmentation model;

[0020] S2.3. Train the constructed instance segmentation model using the instance segmentation dataset to obtain the building instance segmentation model;

[0021] S2.4. Use the building instance segmentation model to perform instance segmentation of all buildings in the two-dimensional orthophoto, and obtain the instance segmentation binary image of the buildings.

[0022] S2.5. Obtain the outline of the binary image of the building, fit the building outline using the minimum rectangle method, and further enlarge the building outline by a set number of pixels at equal intervals to obtain the approximate outline surface of the building, so that the building is completely enclosed within the rectangular outline surface.

[0023] Two-dimensional instance segmentation algorithms can accurately locate building regions in two-dimensional orthophotos, providing a good data foundation for the subsequent extraction of building point cloud sets.

[0024] In step S3, the approximate outline pixel contour coordinates (X... pixel ,Y pixel Convert to real geographic coordinates (X) geo ,Y geo The formula for ) is: [X lt ,R x ,0,Yrd ,0,R y ] is an affine matrix containing the approximate outline of the image, with the matrix format being [x-coordinate of the top-left pixel, resolution of the pixel in the width direction, rotation parameter, y-coordinate of the top-left pixel, rotation parameter, resolution of the pixel in the height direction]; the resolution unit is m.

[0025] The aforementioned transformation from pixel coordinates to geographic coordinates enables interactive conversion between two-dimensional and three-dimensional data, achieving the coupling of two-dimensional and three-dimensional technical means.

[0026] The specific implementation process of step S4 includes:

[0027] S4.1. The contour point data is read using a dynamic loading method, and the contour polygon is generated according to the loading order of the contour point data.

[0028] S4.2 Extract the points inside the outline polygon to obtain a 3D point cloud set;

[0029] S4.3 Calculate the difference between the maximum and minimum heights in the 3D point cloud set, and save the 3D point cloud sets with height differences greater than a set threshold as a single point cloud file.

[0030] The building outline obtained by two-dimensional means is used to extract the corresponding three-dimensional point cloud set of the building, which ensures the integrity of the three-dimensional structure of the building and provides a guarantee for subsequent attribute calculation.

[0031] In step S5, the specific implementation process of separating and saving the multiple buildings existing in the three-dimensional point cloud set described in step S4.3 includes:

[0032] S5.1 Load the three-dimensional point cloud set, and downsample the three-dimensional point cloud set in a uniform downsampling manner, and calculate the normal vector set of the downsampled three-dimensional point cloud set;

[0033] S5.2. Based on the clustering algorithm, perform cluster analysis on the normal vector set to obtain the coordinates of the center points of each class and the point cloud with class labels;

[0034] S5.3. Based on the category labels, the downsampled 3D point cloud set is segmented to separate point clouds of different categories and non-adjacent point clouds of the same category, and these are then combined into new point cloud sets to obtain the 3D point cloud set of each component of the building.

[0035] S5.4. Fit a smooth plane to the three-dimensional point cloud set of each component based on the SVR algorithm, and calculate the smooth planes that are closest to the class center point in the five directions of front, back, left, right and top according to the class center obtained by clustering in step S5.2. Combine the point clouds contained in the smooth planes in the five directions to form a new regular building point cloud, i.e., regular point cloud.

[0036] S5.5. Based on deep learning algorithms, identify the surface categories of regular point clouds, eliminate noise information, and assign structure category names to each smooth plane; at the same time, identify buildings in regular point clouds and obtain the outline regions of buildings.

[0037] S5.6. Combining the outline region and structural category attributes of the building, the point cloud set of a single building is segmented from the three-dimensional point cloud set and saved separately to realize the physical unitization of a single building.

[0038] By clustering buildings into different components, interference point cloud data that does not belong to the building part can be efficiently removed from the point cloud set extracted from the contour surface. Building reorganization based on plane fitting and cluster centers can achieve accurate separation of multiple buildings.

[0039] In step S5, the specific implementation process of refining the building's spatial geometry and extracting its attribute information includes:

[0040] 1) Based on the building's structural category attributes, identify the top components of the building to recognize the shape of the building's top; calculate the height difference between the highest and lowest points of the building to obtain the building's height; extract the components of the four walls of the building according to the building's structural category attributes, fit a two-dimensional closed polygon enclosed by the four walls, and calculate the building's bottom area based on the two-dimensional closed polygon.

[0041] 2) Based on the building's height, roof geometry and color texture as the judgment criteria, (1) the building's top is a gable roof with a height of 3-6m, and it is judged as a brick-wood structure; (2) the building's top is a flat roof with a height of 3-9m, and it is judged as a brick-concrete structure; (3) the building's height is more than 9m, and it is judged as a frame structure. Based on the building's orthophoto data, the building's roof geometry category (gable roof, flat roof) is obtained using a neural network classification model. Combined with the building's height attribute obtained in step S5, the identification of the building's material type is realized.

[0042] The present invention extracts building attributes using judgment criteria, fully combining multi-dimensional information to achieve accurate and efficient calculation of building attributes.

[0043] In step 2) above, the judgment criteria include: if the building has a gable roof and a height of 3-6m, it is judged to be a brick-wood structure; if the building has a flat roof and a height of 3-9m, it is judged to be a brick-concrete structure; if the building is more than 9m high, it is judged to be a frame structure.

[0044] In this invention, the judgment criteria are used as the basis for neural network judgment. The results predicted by the neural network are compared with the judgment criteria, and those that meet the requirements are the corresponding material types.

[0045] As an inventive concept, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory; the processor executes the computer program to implement the steps of the method described in the present invention.

[0046] As an inventive concept, the present invention also provides a computer program product, including a computer program / instructions; when the computer program / instructions are executed by a processor, they implement the steps of the method described in the present invention.

[0047] As an inventive concept, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon; when the computer program / instructions are executed by a processor, they implement the steps of the method described in the present invention.

[0048] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0049] (1) This invention breaks through the technical bottleneck of utilizing oblique photogrammetry data, realizes the rapid and intelligent physical individualization of buildings in oblique photogrammetry models and point cloud models, greatly reduces labor costs, improves the utilization rate of oblique photogrammetry data, and improves production efficiency.

[0050] (2) This invention breaks through the technical bottlenecks of existing single-unit technology, such as the inability to identify materials, difficulty in extracting area, and inaccurate height extraction when extracting buildings, and greatly improves the extraction accuracy of building attributes, so that the results can meet the requirements of engineering applications.

[0051] (3) The present invention uses a two-dimensional coupling method to automatically identify the building type, realizes the automatic extraction and assignment of attributes such as building area, height, number of floors, and material, and solves the problem that the existing single-unit technology requires manual addition of attributes and has a cumbersome process. Attached Figure Description

[0052] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of point cloud contour extraction according to an embodiment of the present invention;

[0054] Figure 3(a) shows the original image; Figure 3(b) shows the result after the original image has been singled out. Detailed Implementation

[0055] like Figure 1 The method of this invention mainly includes the following steps:

[0056] S1. Obtain two-dimensional orthophotos with geographic coordinate information and corresponding three-dimensional point cloud data.

[0057] S2. A deep learning-based instance segmentation model is used to obtain the approximate outline of a building in a two-dimensional orthophoto.

[0058] S3: Convert the approximate outline of the building in S2 into an outline with actual geographic coordinates.

[0059] S4. Based on the contour surface of S3, cut the three-dimensional point cloud data to obtain a three-dimensional point cloud set (point cloud model), and realize the initial physical unitization of the building.

[0060] S5. Based on the individual building model, eliminate noise points such as ground and trees, and separate and preserve multiple buildings that may exist in the S4 point cloud model.

[0061] S6. Based on the spatial geometry of the building, refine the building's structure and extract its attribute information.

[0062] S7. Complete the physical unitization of buildings based on two-dimensional and three-dimensional coupling.

[0063] In step S1, the two-dimensional orthophoto with geographic coordinate information and the corresponding three-dimensional point cloud data refer to the fact that the geographic coordinates of each pixel in the two-dimensional orthophoto are known.

[0064] In step S2, a deep learning-based instance segmentation model is used to obtain the approximate outline of the building in the 2D orthophoto. Obtaining the approximate outline of the building based on this model further includes:

[0065] S2.1. The 2D orthophoto image containing buildings is cropped into 2048×2048 sub-images. Specifically, in this embodiment, point cloud data and orthophoto images (DOM) are generated based on the oblique photogrammetry model, and the DOM is divided into blocks using the gdal library, so that each DOM block is 2048*2048 in size.

[0066] S2.2. Perform instance segmentation annotation on the sub-images cropped in S2.1 to construct a building instance segmentation dataset.

[0067] S2.3. Construct the Mask_Rcnn instance segmentation model based on deep learning frameworks such as PyTorch or TensorFlow.

[0068] S2.4. Use the instance segmentation dataset obtained in S2.2 to train the instance segmentation model constructed in S2.3 to obtain the building instance segmentation model.

[0069] S2.5. Using the building instance segmentation model in S2.4, perform instance segmentation on all buildings in the two-dimensional orthophoto image to be identified in S1 with geographic coordinate information, and obtain the binary image of the building instance segmentation.

[0070] S2.6 Obtain the outline of the building's binary image obtained in S2.5, fit the building outline using the minimum rectangle method, and further enlarge it by a certain number of pixels at equal intervals to obtain the approximate outline surface of the building, ensuring that the building is completely enveloped within the rectangular outline surface.

[0071] In step S3, the approximate outline of the building is converted into an outline with actual geographic coordinates, which further includes:

[0072] The two-dimensional orthophoto with geographic coordinate information mentioned in S3.1 and S1 refers to a two-dimensional orthophoto whose pixel geographic transformation affine matrix is ​​known. The matrix format is [x-coordinate of the top left pixel, resolution of the pixel in the width direction (in meters), rotation parameter, y-coordinate of the top left pixel, rotation parameter, resolution of the pixel in the height direction (in meters)].

[0073] The pixel contour coordinates (X) obtained by S3.2 and S26 pixel ,Y pixel Convert to real geographic coordinates (X) geo ,Y geo The conversion formula is as follows:

[0074] Given that the affine matrix of each image is [X lt ,R x ,0,Y rd ,0,R y ],but:

[0075]

[0076] Based on the above formula, the actual geographic outline surface corresponding to the pixel outline surface of the building can be obtained.

[0077] Step S4 is a method for extracting buildings within a region based on a two-dimensional segmented contour surface, which further includes:

[0078] S4.1 uses a dynamic loading method to read two-dimensional contour point data and generates contour polygons according to the loading order of contour points.

[0079] S4.2 uses Open3D to extract points inside the outline polygon;

[0080] S4.3 calculates the difference between the maximum and minimum heights in the point set, saves the point cloud of building objects with a height difference greater than 1.2m as a single point cloud file, and realizes the initial segmentation of buildings and the removal of point clouds of vehicles, small objects, etc.

[0081] S5 is a 3D point cloud denoising and separation technology for buildings, and it achieves intelligent extraction of various parts of buildings. It solves problems such as buildings easily merging into one area in dense areas, difficulty in separating buildings from the ground, and occlusion of target buildings by tall buildings and trees. Further features include:

[0082] S5.1 loads the point cloud set of a single building separated in S4.3 (referred to as point set 0), calculates its normal vector set, and downsamples the point cloud set in a uniform downsampling manner to obtain point set 1, reducing the number of points to 50% of the original number of points.

[0083] S5.2 performs cluster analysis on the normal vectors based on clustering algorithms (such as k-means, mean shift, Random Sample Consensus, etc.) to obtain the coordinates of the center points of each class and the point cloud with class labels.

[0084] S5.3 segments the point cloud set based on the class labels in S5.2, separating point cloud sets of different categories from point cloud sets of non-adjacent points within the same category, thus obtaining the point sets of each component of the building.

[0085] S5.4 fits a smooth plane to the point set of each component obtained in S5.3 based on the SVR algorithm and S5.3. Based on the cluster centers obtained, it calculates the smooth planes that are closest to the cluster center points in the five directions of front, back, left, right and top. The point clouds contained in the smooth planes in the five directions are combined to form a new regular building point cloud (hereinafter referred to as regular point cloud).

[0086] S5.5 uses deep learning algorithms (such as PointNet++, RandLa-Net, etc.) to identify the surface categories of regular point clouds, eliminate noise information such as ground and trees, and assign structure category names to each smooth surface; at the same time, it identifies buildings in regular point clouds to obtain the outline regions of buildings.

[0087] S5.6 combines the building outline region and structural category attributes obtained in S5.5 to further segment the point cloud set of individual buildings from point set 0 and save them separately, thus realizing the physical individualization of individual buildings.

[0088] S6 involves extracting attribute information such as building height, material structure, and building base area, which further includes:

[0089] S6.3, based on the structural category attributes of building components in S5, identifies the top components of the building, thereby recognizing the shape of the building's top.

[0090] S6.4 calculates the height difference between the highest and lowest points of a building to extract the building's height.

[0091] S6.5 Based on the building's structural category attributes, extract the components of the four walls of the building, fit a two-dimensional closed polygon enclosed by the four walls, and calculate the building's bottom area based on the two-dimensional closed polygon.

[0092] S6.6 extracts the material structure of buildings based on their height, geometry, and color texture.

[0093] S6.7 integrates the recognition results from S6.3 to S6.6 and saves them as a dictionary in a file to save building attributes.

[0094] S7 integrates the point cloud of a single building in S6 and its corresponding attribute information, and outputs the result, thus realizing the physical individualization of the building.

[0095] Taking the actual application in the Toba land acquisition and resettlement project as an example, the following is the specific implementation process of the application of this invention.

[0096] S1. Generate point cloud data and orthophoto images (DOM) based on the oblique photogrammetry model. Use the gdal library to divide the DOM into blocks, so that the size of each DOM block is 2048*2048.

[0097] S2. Separate the sliced ​​DOM data and use the gdal library to process the DOM to obtain RGB image data and coordinate affine matrix;

[0098] S3. Based on the Mask_Rcnn model, the RGB image is recognized and segmented to obtain the building segmentation mask; based on the mask information, the corresponding RGB original image is cut to obtain the building miniature image;

[0099] S4. Analyze building thumbnails based on machine vision algorithms, extract building boundary contours, and map them to a three-dimensional coordinate system;

[0100] S5. Based on the results of S4, segment the 3D point cloud model and delete buildings with a height of less than 1.2m;

[0101] S6. Separate individual buildings from the results of S5 and generate a set of point clouds for each individual building.

[0102] S7. Refine the segmentation results and extract building attributes;

[0103] S8. Output the results.

[0104] Furthermore, S4 specifically includes:

[0105] S4.1 Build the Canny algorithm based on the OpenCV library and calculate the corresponding binary images of buildings;

[0106] S4.2. The building outline is fitted using the minimum rectangle method, and the geographic coordinates of the building outline are calculated based on the affine matrix and mapped to the three-dimensional point cloud coordinate system.

[0107] Furthermore, S6 specifically includes:

[0108] S6.1 loads the point cloud set of a single building separated in S5 (referred to as point set 0), calculates its normal vector set, and downsamples the point cloud set in a uniform downsampling manner to obtain point set 1, reducing the number of points to 50% of the original number of points.

[0109] S6.2 uses the Random Sample Consensus algorithm to perform cluster analysis on the normal vectors, obtaining the coordinates of the center points of each class and the point cloud with class labels.

[0110] S6.3 segments the point cloud set based on the class labels in S6.2, separating point cloud sets of different categories from point cloud sets of non-adjacent points within the same category, thus obtaining the point sets of each component of the building.

[0111] S6.4 fits a smooth plane to the point set of each component obtained in S6.3 based on the SVR algorithm and S6.3. Based on the cluster centers obtained, it calculates the smooth planes that are closest to the cluster center points in the five directions of front, back, left, right and top. The point clouds contained in the smooth planes in the five directions are combined to form a new regular building point cloud, namely the regular point cloud (hereinafter referred to as the regular point cloud).

[0112] S6.5 uses the PointNet++ deep learning algorithm to identify the surface categories of regular point clouds, removes point information of ground features such as trees from the point cloud, and assigns the identified categories to each smooth surface; simultaneously, it calculates the contour regions of buildings in the regular point cloud based on z-axis compression, such as... Figure 2 As shown.

[0113] S6.6 combines the building outline region and structural category attributes obtained in S6.5 to further segment the point cloud set of individual buildings from point set 0 and save them separately, thus realizing the physical individualization of individual buildings.

[0114] Furthermore, S7 specifically refers to:

[0115] Based on the identification results of building components in S6, S7.1 extracts the information of the top of the building, calculates the included angle parameters of each face of the top, and realizes the judgment of the shape of the top of the building (pointed roof, flat roof, gable roof, etc.).

[0116] S7.2 calculates the height difference between the highest and lowest points of a building, thus extracting the building's height.

[0117] S7.3 Based on the building's structural category attributes, extract the components of the four walls of the building, fit a two-dimensional closed polygon enclosed by the four walls, and calculate the building's bottom area based on the two-dimensional closed polygon.

[0118] S7.4 uses the building's height, geometry, and color texture, combined with known building material structure information maps, to achieve a rough identification of the building's material structure.

[0119] Figures 3(a) and 3(b) illustrate the effect of the method of the present invention on house extraction. The method of the present invention automatically extracts the attribute information such as height, number of floors, area, and material of houses through two-dimensional and three-dimensional coupling. The identified house outline and corresponding attribute information are used to generate a shape surface file. The shape file is then overlaid on the three-dimensional oblique photogrammetry model (Figure 3(a) is the original oblique photogrammetry model). By selecting a detached house, the attribute query of the individual house can be realized. The houses highlighted in Figure 3(b) are the attribute queries.

Claims

1. A building individualization information extraction method characterized by, The method comprises the following steps: S1, obtaining a building two-dimensional orthographic image with geographic coordinate information and corresponding three-dimensional point cloud data; S2, obtaining a rough outline surface of the building in the two-dimensional orthographic image based on an instance segmentation model of deep learning; S3, converting the rough outline surface into an outline surface of real geographic coordinates; S4, cutting the three-dimensional point cloud data based on the outline surface obtained in step S3 to obtain a three-dimensional point cloud set; S5, separating and saving a plurality of buildings existing in the three-dimensional point cloud set, and performing fine processing on the buildings based on the spatial geometric structure of the buildings to extract attribute information of the buildings; The specific implementation process of step S4 comprises: S4.1, reading the outline surface point data in a dynamic loading manner, and generating an outline polygon according to the loading sequence of the outline surface point data; S4.2, extracting points inside the outline polygon to obtain a three-dimensional point cloud set; S4.3, calculating the difference between the maximum height and the minimum height in the three-dimensional point cloud set, and saving the three-dimensional point cloud set with a height difference greater than a set threshold as a single point cloud file; In step S5, the specific implementation process of separating and saving a plurality of buildings existing in the three-dimensional point cloud set in step S4.3 comprises: S5.1, loading the three-dimensional point cloud set, and performing down-sampling on the three-dimensional point cloud set in a uniform down-sampling manner, and calculating a normal vector set of the down-sampled three-dimensional point cloud set; S5.2, performing clustering analysis on the normal vector set based on a clustering algorithm to obtain a class center point coordinate and a point cloud with a class label; S5.3, based on the class label, segmenting the down-sampled three-dimensional point cloud set, separating point clouds of different categories and point clouds that are not adjacent in the same category, and grouping them into new point cloud sets respectively, so as to obtain a three-dimensional point cloud set of each component of the building; S5.4, fitting a smooth plane to each component of the three-dimensional point cloud set based on an SVR algorithm, and calculating the smooth plane closest to the class center point in the front, rear, left, right and upper five directions according to the class center obtained in step S5.2, and grouping the point clouds contained in the five direction smooth planes to form a new regular building point cloud, i.e. a regular point cloud; S5.5, identifying each face category of the regular point cloud based on a deep learning algorithm, eliminating noise information and assigning a structure category name to each smooth plane; at the same time, identifying the building in the regular point cloud to obtain the outline area of the building; S5.6, combining the outline area and the structure category attribute of the building, segmenting a single building point cloud set from the three-dimensional point cloud set and saving it respectively to realize physical monomerization of a single building.

2. The building individualization information extraction method according to claim 1, characterized by, The specific implementation process of step S2 comprises: S2.1, cropping the two-dimensional orthographic image into a sub-image with a size of M*M; wherein M is a constant; S2.2, attaching label information to each sub-image cropped in S2.1 to construct a building instance segmentation dataset; based on a deep learning framework, constructing a Mask_Rcnn instance segmentation model; S2.3, training the constructed Mask_Rcnn instance segmentation model using the instance segmentation dataset to obtain a building instance segmentation model; S2.4, instance segmentation of all buildings in the two-dimensional orthographic image is completed by using the building instance segmentation model, to obtain the instance segmentation binary image of the building; S2.5, the contour of the building binary image is obtained, the minimum rectangle method is used to fit the building contour, the building contour is further enlarged by a certain number of pixels, and the approximate contour surface of the building is obtained, so that the building is completely enveloped in the rectangular contour surface.

3. The building individualization information extraction method according to claim 1, characterized by, In step S3, the formula for converting the outline pixel profile coordinates (X pixel ,Y pixel ) of the approximate profile into real geographical coordinates (X geo ,Y geo ) is: [X lt ,R x ,0,Y rd ,0,R y ] is an affine matrix containing the image of the approximate profile, the matrix format being [x coordinate of the upper left pixel, resolution of the pixel in the width direction, rotation parameter, y coordinate of the upper left pixel, rotation parameter, resolution of the pixel in the height direction]; the resolution unit is m.

4. The building individualization information extraction method according to claim 1, characterized by, In step S5, based on the spatial geometric structure of the building, the house is finely processed, and the specific implementation process of extracting the attribute information of the building includes: 1) Based on the building structure category attribute, the top component of the building is distinguished, and the identification of the top shape of the building is realized; The height difference between the highest point and the lowest point of the building is calculated to obtain the height of the building; According to the building structure category attribute, the four wall components of the building are extracted, a two-dimensional closed polygon surrounded by the four walls is fitted, and the building floor area is calculated based on the two-dimensional closed polygon; 2) Taking the height, roof geometric structure and color texture of the building as the judgment criteria, based on the building orthographic image data, the building roof geometric structure category is obtained by using the neural network classification model, and the building material type is identified by combining the building height attribute obtained in step S5.

5. The building individualization information extraction method according to claim 4, characterized by, In step 2), the judgment criteria include: if the top of the building is a hipped roof, the height is 3-6m, it is judged as a brick-wood structure; If the top of the building is a flat roof, the height is 3-9m, it is judged as a brick-concrete structure; If the height of the building is more than 9m, it is judged as a frame structure.

6. A computer apparatus comprising a memory, a processor, and a computer program stored on the memory; characterized in that, The processor executes the computer program to realize the steps of the method of any one of claims 1-5.

7. A computer program product comprising computer programs / instructions; characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of any one of claims 1-5.

8. A computer readable storage medium having stored thereon computer programs / instructions; characterized in that, The computer program / instructions are executed by the processor to realize the steps of the method of any one of claims 1-5. The computer program / instructions are executed by the processor to realize the steps of the method of any one of claims 1-5.

Images