Complex shape three-dimensional building model generation method based on frame field deep learning

Abstract

The invention relates to a method for generating a complex-shape three-dimensional building model based on frame field deep learning. The problems that in the prior art, when a three-dimensional building model is generated, training difficulty is large, the three-dimensional building model is easily affected by smooth corners, and artifacts are generated are solved. The method comprises the following steps: S1, acquiring remote sensing data and point cloud data; s2, preprocessing the acquired image data; s3, carrying out data normalization processing; s4, designing a frame field and optimizing details; s5, optimizing the activity skeleton model; and S6, generating a multi-shape contour and a three-dimensional model. The method has the advantages that the generation precision and the generation efficiency of the three-dimensional building model are improved, buildings with complex shapes can be processed, and the method has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the field of architectural technology, and in particular to a method for generating a three-dimensional architectural model with a complex shape based on frame field deep learning. Background Art

[0002] With the rapid development of remote sensing technology and computer vision, there is an increasing demand for automatically extracting building information and generating 3D models from satellite images, drone aerial images or ground laser scanning data. Traditional methods usually rely on manual modeling or the use of rule-based automation tools, which are inefficient and inaccurate when dealing with complex-shaped buildings.

[0003] In recent years, deep learning techniques, especially convolutional neural networks, have made significant progress in the field of image processing and segmentation. However, most existing deep image segmentation models output segmentation results in raster form, which is not ideal in geographic information system applications because GIS usually requires data in vector polygon format. In addition, it is still a challenge to generate 3D building models with complex shapes directly from image or point cloud data, especially when dealing with complex situations such as sharp corners, internal holes, or shared walls between adjacent buildings.

[0004] In response to the above problems, some polygonization methods based on deep learning have emerged in recent years, such as Curve-GCN and Poly Mapper. However, these methods draw three-dimensional models based on the appearance information of the building, which requires the collection of a large amount of drawing design data in the early stage, involving communication among multiple departments, and a large workload. There is no guarantee that the required architectural drawings can be obtained, and it requires a long engineering production cycle and a large investment cost. In addition, the existing three-dimensional building models are difficult to train and can only handle simple polygons, and cannot effectively cope with complex urban building scenes. In addition, some methods generate polygons through contour detection and post-processing, but they are easily affected by smooth corners and produce artifacts.

[0005] In order to solve the deficiencies of the existing technology, people have conducted long-term exploration and proposed various solutions. For example, a Chinese patent document discloses a method and system for generating a virtual building model [CN202211550542.7], which includes step S1: obtaining a building plane image, parsing the building plane image, obtaining the final contour line of the plane, and locating the contour centroid of the target building based on the final contour line of the plane; step S2: estimating the height of the target building; step S3: obtaining the second shooting point of the shooting device based on the contour centroid of the target building and the vertical shooting range of the shooting device; step S4: shooting towards the target building at each of the second shooting points to obtain multiple building facade images; step S5: generating a rectangular model in the virtual space, folding the rectangular model, and obtaining an initial virtual building model; step S6: correcting the initial virtual building model based on the building facade image to obtain the final virtual building model.

[0006] The above scheme solves to a certain extent the problem of large initial workload and long production cycle of three-dimensional models drawn based on building appearance information in the existing technology. However, the scheme still has many shortcomings, such as: the model training is difficult and is easily affected by smooth corners, resulting in artifacts. Summary of the invention

[0007] The purpose of the present invention is to provide a method for generating a complex three-dimensional building model based on frame field deep learning in order to solve the above problems.

[0008] To achieve the above object, the present invention adopts the following technical scheme: a method for generating a complex-shaped three-dimensional building model based on frame field deep learning, the method comprising the following steps:

[0009] S1. Remote sensing data and point cloud data acquisition;

[0010] S2, preprocessing the acquired image data;

[0011] S3, data normalization processing;

[0012] S4, frame field design and detail optimization;

[0013] S5, active skeleton model optimization;

[0014] S6. Polymorphic contour and three-dimensional model generation.

[0015] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, step S1 includes the following steps:

[0016] S01. Equipment configuration and calibration: Before collecting data, the drone or LiDAR device needs to be calibrated to ensure that the position, angle and sensor accuracy meet the requirements;

[0017] S02. Area planning and path setting: Determine the aerial photography or scanning area according to the scale of the building and set the best flight or driving path to cover the entire target area;

[0018] S03, Data Collection: The equipment runs along the set path and altitude, collects remote sensing images or point cloud data in real time, and records GPS information to ensure the spatial consistency of the data;

[0019] S04. Data export: After the acquisition is completed, the image or point cloud data is imported into the computing platform for subsequent processing.

[0020] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, in step S1, the remote sensing image acquisition step includes:

[0021] S1-1. Obtain high-resolution images of buildings through satellite remote sensing platforms;

[0022] S1-11. Use drones equipped with high-resolution cameras to take aerial images of buildings;

[0023] The steps of acquiring point cloud data include:

[0024] S1-2. Use LiDAR equipment to scan buildings. LiDAR uses the principle of laser ranging to generate high-precision three-dimensional point cloud data to obtain the geometric features of the building's shape;

[0025] LiDAR equipment mainly scans buildings through airborne, vehicle-mounted or ground-based LiDAR systems. During the LiDAR data acquisition process, the equipment is usually installed on drones, helicopters or ground vehicles. By setting a route or path, it automatically scans the designated area and generates point cloud data.

[0026] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, step S2 includes the following steps:

[0027] Point cloud data denoising:

[0028] S2-1, Bilateral Filter: It removes noise while retaining the edge features of the point cloud. The filter reduces the influence of noise points by considering the spatial neighborhood and the surface normal vector of the point cloud.

[0029] S2-11, Static Outlier Removal: Removes outliers that deviate from the normal distribution based on the neighborhood distribution of points in the point cloud. It is suitable for processing accidental noise in large-scale point cloud data.

[0030] Remote sensing image denoising:

[0031] S2-2, Gaussian filter: removes noise by blurring the high-frequency part of the image, and processes remote sensing images with large changes in light intensity;

[0032] S2-21, Total Variation Regularization: retains the main edge features in the image while reducing random noise;

[0033] S2-22. Data Retention: After noise removal, key features of buildings are retained in the data, such as wall edges, roof lines, and details of building facades.

[0034] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, step S3 includes the following steps:

[0035] Point cloud data normalization:

[0036] S3-1. Scale normalization: Ensure that the model can handle buildings of different sizes by mapping the coordinates of the point cloud data to the unit cube [0,1] range or scaling according to known building sizes;

[0037] S3-11, Centering: Align the geometric center of the point cloud to the origin of the coordinate system to ensure that the model can input building geometry data in a standard posture;

[0038] Image data normalization:

[0039] S3-2, pixel value normalization: normalize the grayscale or color value of the image to the [0,1] interval to reduce the impact of different image brightness or contrast on the deep learning model;

[0040] S3-21. Spatial alignment: Align multiple remote sensing images to ensure that the positions of buildings are consistent from different viewing angles and reduce deviations caused by differences in image angles and shooting heights.

[0041] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, step S4 includes the following steps:

[0042] Frame field design and optimization details:

[0043] S4-1, frame field structure: Each pixel in the frame field contains four direction vectors, which point to the main direction of the building outline;

[0044] S4-11, Corner point enhancement: Design a special corner point detection mechanism in the model to ensure that the frame field generates high-precision corner points at the corners of the building to avoid the contour being too smooth;

[0045] Frame field training details:

[0046] S4-2, Direction consistency loss: guide the model by calculating the consistency between the frame field vector and the actual building edge direction to optimize the accuracy of the frame field;

[0047] S4-21. Corner enhancement loss: For the corner areas of buildings, a loss function is designed to enhance the model’s ability to detect corners and ensure that the frame field accurately reflects the directional changes of these locations.

[0048] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, step S5 includes the following steps:

[0049] The active skeleton model optimizes the outline to align with the frame field and the frame field detects corners:

[0050] S5-1, Active Skeleton Model (ASM): gradually align the building outline with the frame field output by moving and adjusting the skeleton points;

[0051] Optimize the contour steps:

[0052] S5-2, initial skeleton generation: based on the four-vector field output by the frame field, the active skeleton of the building is preliminarily generated; S5-21, skeleton adjustment: using the direction vector information in the frame field, the position of the skeleton points is optimized by gradient descent to make it consistent with the acute angle and main direction of the building edge;

[0053] S5-22, corner point detection: detect the corner positions of buildings and mark them through the corner point information of the frame field. For these corner points, use the local optimization algorithm to adjust their positions to ensure that the generated polygonal contour has high accuracy;

[0054] Skeleton and frame field alignment details:

[0055] S5-3, Dynamically adjust skeleton points: When aligning skeleton points with frame fields, adjust the displacement of skeleton points through iterative optimization to ensure that they are highly consistent with the actual building boundary;

[0056] S5-31. Error control: The difference between the skeleton and the frame field is calculated through the error function, and a regularization term is introduced to avoid generating contours that are too complex or smooth.

[0057] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, in step S6, when the polygonal outline is generated, the skeleton of the building is converted into a polygonal outline by detecting the turning points and edge points of the skeleton. In order to avoid too many vertices, a sparseness algorithm is used to simplify non-critical vertices, thereby generating a simplified and accurate polygonal outline.

[0058] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, in step S6, the three-dimensional model generation is divided into three-dimensional model establishment and three-dimensional model deepening processing, wherein the three-dimensional model establishment is divided into the following steps:

[0059] S6-1. Acquisition of height information: The height information of buildings can be inferred from the shadow length in remote sensing images or directly obtained through GIS data. For complex buildings, the height of multiple layers can be determined in combination with the existing building outline map;

[0060] S6-11, Contour-based 3D model construction: Generate a 2D plane based on polygonal contours and expand it into a 3D model using building height information;

[0061] S6-12. Multi-storey building modeling: For high-rise buildings, the model will generate multiple floors according to the number of floors of the building, and adjust the building shape of each floor to fit the actual structure;

[0062] Facade details:

[0063] S6-13, facade texture: Enhance the visual effect of the building by fitting the facade texture of the actual building on the generated 3D model;

[0064] S6-14, Complex geometry processing: For complex geometric forms in buildings, the details of polygon outlines are adjusted according to the directional information of the frame field to ensure the accuracy of the building's geometric details.

[0065] In the above-mentioned method for generating a complex-shaped three-dimensional building model based on frame field deep learning, in step S6, the three-dimensional model deepening process includes the following steps:

[0066] S6-2, Smoothing: Apply the Laplace smoothing algorithm on the generated 3D model to eliminate sharp edges in the building shape while retaining key geometric features;

[0067] S6-21, Denoising: Remove artifacts and redundant vertices in the model through wavelet transform to optimize the surface smoothness of the model;

[0068] S6-22, Texture mapping: Map the actual image of the building or the corresponding material onto the surface of the 3D model to improve its realism and visual effect. Depending on the different parts of the building, multi-layer texture mapping technology is used;

[0069] Model format conversion:

[0070] S6-23. Model export: The generated 3D building model can be exported to common 3D formats.

[0071] Compared with the prior art, the advantages of the present invention are:

[0072] 1. Improve accuracy: By introducing frame field learning, it is possible to capture the sharp angles and directional information of the building outline, reduce the ambiguity caused by discrete probability maps, and thus improve the generation accuracy of the 3D building model.

[0073] 2. Improve efficiency: Frame field learning hardly increases the training and reasoning time cost, simplifies the polygonization process, and improves the generation efficiency of 3D building models.

[0074] 3. Strong adaptability: This method is able to handle buildings with complex shapes, including large buildings, buildings with internal holes, and situations where adjacent buildings share walls.

[0075] 4. Wide application: The generated 3D building model can be used in many fields such as urban planning, architectural design, GIS, etc., and has broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] Figure 1 is a flow chart of the method of the present invention;

[0077] Figure 2 It is a schematic diagram of a frame field in the present invention;

[0078] Figure 3 It is a framework diagram of the active skeleton model algorithm in the present invention;

[0079] Figure 4 It is a diagram of the building outline extraction process in the present invention;

[0080] Figure 5 The present invention is based on the nonlinear fitting weight map of elevation;

[0081] Figure 6 It is the building elevation prediction map; DETAILED DESCRIPTION

[0082] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0083] like Figure 1-6 As shown, a method for generating a complex-shaped three-dimensional building model based on frame field deep learning includes the following steps:

[0084] S1. Remote sensing data and point cloud data acquisition;

[0085] S2, preprocessing the acquired image data;

[0086] S3, data normalization processing;

[0087] S4, frame field design and detail optimization;

[0088] S5, active skeleton model optimization;

[0089] S6. Polymorphic contour and three-dimensional model generation.

[0090] Wherein, step S1 comprises the following steps:

[0091] S01. Equipment configuration and calibration: Before collecting data, the drone or LiDAR device needs to be calibrated to ensure that the position, angle and sensor accuracy meet the requirements;

[0092] S02. Area planning and path setting: Determine the aerial photography or scanning area according to the scale of the building and set the best flight or driving path to cover the entire target area;

[0093] S03, Data Collection: The equipment runs along the set path and altitude, collects remote sensing images or point cloud data in real time, and records GPS information to ensure the spatial consistency of the data;

[0094] S04. Data export: After the acquisition is completed, the image or point cloud data is imported into the computing platform for subsequent processing.

[0095] In step S1, the remote sensing image acquisition step includes:

[0096] S1-1. Obtain high-resolution images of buildings through satellite remote sensing platforms;

[0097] Satellite remote sensing platforms include satellites such as Landsat, Sentinel or World View, which provide multispectral data that can be used to extract geometric information about buildings.

[0098] S1-11. Use drones equipped with high-resolution cameras to take aerial images of buildings;

[0099] It is particularly suitable for low-altitude aerial photography and can be used for small area or local fine modeling.

[0100] The steps of acquiring point cloud data include:

[0101] S1-2. Use LiDAR equipment to scan buildings. LiDAR uses the principle of laser ranging to generate high-precision three-dimensional point cloud data to obtain the geometric features of the building's shape;

[0102] LiDAR equipment mainly scans buildings through airborne, vehicle-mounted or ground-based LiDAR systems. During the LiDAR data acquisition process, the equipment is usually installed on drones, helicopters or ground vehicles. By setting a route or path, it automatically scans the designated area and generates point cloud data.

[0103] Step S2 includes the following steps:

[0104] Point cloud data denoising:

[0105] S2-1, Bilateral Filter: It removes noise while retaining the edge features of the point cloud. The filter reduces the influence of noise points by considering the spatial neighborhood and the surface normal vector of the point cloud.

[0106] S2-11, Static Outlier Removal: Removes outliers that deviate from the normal distribution based on the neighborhood distribution of points in the point cloud. It is suitable for processing accidental noise in large-scale point cloud data.

[0107] Denoising is an important step in data preprocessing, which aims to reduce interference signals introduced by environmental factors or equipment noise and improve the quality and accuracy of data.

[0108] Remote sensing image denoising:

[0109] S2-2, Gaussian filter: removes noise by blurring the high-frequency part of the image, and processes remote sensing images with large changes in light intensity;

[0110] S2-21, Total Variation Regularization: retains the main edge features in the image while reducing random noise;

[0111] S2-22. Data Retention: After noise removal, key features of buildings are retained in the data, such as wall edges, roof lines, and details of building facades.

[0112] Step S3 includes the following steps:

[0113] Point cloud data normalization:

[0114] S3-1. Scale normalization: Ensure that the model can handle buildings of different sizes by mapping the coordinates of the point cloud data to the unit cube [0,1] range or scaling according to known building sizes;

[0115] S3-11, Centering: Align the geometric center of the point cloud to the origin of the coordinate system to ensure that the model can input building geometry data in a standard posture;

[0116] Image data normalization:

[0117] S3-2, pixel value normalization: normalize the grayscale or color value of the image to the [0,1] interval to reduce the impact of different image brightness or contrast on the deep learning model;

[0118] S3-21. Spatial alignment: Align multiple remote sensing images to ensure that the positions of buildings are consistent from different viewing angles and reduce deviations caused by differences in image angles and shooting heights.

[0119] The purpose of normalization is to convert data from different sources and scales to a unified scale or spatial range so that subsequent deep learning models can accurately process and learn consistent building features.

[0120] Step S4 includes the following steps:

[0121] Frame field design and optimization details:

[0122] S4-1, frame field structure: Each pixel in the frame field contains four direction vectors, which point to the main direction of the building outline;

[0123] For buildings with sharp angles and complex structures, frame fields can accurately capture these local geometric features.

[0124] S4-11, Corner point enhancement: Design a special corner point detection mechanism in the model to ensure that the frame field generates high-precision corner points at the corners of the building to avoid the contour being too smooth;

[0125] Frame field training details:

[0126] S4-2, Direction consistency loss: guide the model by calculating the consistency between the frame field vector and the actual building edge direction to optimize the accuracy of the frame field;

[0127] S4-21. Corner enhancement loss: For the corner areas of buildings, a loss function is designed to enhance the model’s ability to detect corners and ensure that the frame field accurately reflects the directional changes of these locations.

[0128] Step S5 includes the following steps:

[0129] The active skeleton model optimizes the outline to align with the frame field and the frame field detects corners:

[0130] S5-1, Active Skeleton Model ASM: gradually align the building outline with the frame field output by moving and adjusting the skeleton points;

[0131] ASM is a shape-based modeling technique that can generate the outline of a building based on the skeleton information of the input data.

[0132] Optimize the contour steps:

[0133] S5-2. Initial skeleton generation: Based on the four-vector field output by the frame field, the active skeleton of the building is preliminarily generated; the skeleton points represent the main structure of the building.

[0134] S5-21, skeleton adjustment: using the direction vector information in the frame field, the position of the skeleton points is optimized by gradient descent to make it consistent with the sharp angles and main directions of the building edges;

[0135] S5-22, corner point detection: detect the corner positions of buildings and mark them through the corner point information of the frame field. For these corner points, use the local optimization algorithm to adjust their positions to ensure that the generated polygonal contour has high accuracy;

[0136] Skeleton and frame field alignment details:

[0137] S5-3, Dynamically adjust skeleton points: When aligning skeleton points with frame fields, adjust the displacement of skeleton points through iterative optimization to ensure that they are highly consistent with the actual building boundary;

[0138] S5-31. Error control: The difference between the skeleton and the frame field is calculated through the error function, and a regularization term is introduced to avoid generating contours that are too complex or smooth.

[0139] In step S6, when the polygonal outline is generated, the skeleton of the building is converted into a polygonal outline by detecting the turning points and edge points of the skeleton. To avoid too many vertices, a sparseness algorithm is used to simplify non-critical vertices, thereby generating a simplified and accurate polygonal outline.

[0140] In step S6, the 3D model generation is divided into 3D model establishment and 3D model deepening processing, wherein the 3D model establishment is divided into the following steps:

[0141] S6-1. Acquisition of height information: The height information of buildings can be inferred from the shadow length in remote sensing images or directly obtained through GIS data. For complex buildings, the height of multiple layers can be determined in combination with the existing building outline map;

[0142] S6-11, Contour-based 3D model construction: Generate a 2D plane based on polygonal contours and expand it into a 3D model using building height information;

[0143] The facade of each building is extruded according to the shape of the outline.

[0144] S6-12. Multi-storey building modeling: For high-rise buildings, the model will generate multiple floors according to the number of floors of the building, and adjust the building shape of each floor to fit the actual structure;

[0145] Facade details:

[0146] S6-13, facade texture: Enhance the visual effect of the building by fitting the facade texture of the actual building on the generated 3D model;

[0147] S6-14, Complex geometry processing: For complex geometric forms in buildings, such as depressions and internal holes, the details of the polygon outline are adjusted according to the direction information of the frame field to ensure the accuracy of the building's geometric details.

[0148] In step S6, the 3D model deepening process includes the following steps:

[0149] S6-2, Smoothing: Apply the Laplace smoothing algorithm on the generated 3D model to eliminate sharp edges in the building shape while retaining key geometric features;

[0150] S6-21, Denoising: Remove artifacts and redundant vertices in the model through wavelet transform to optimize the surface smoothness of the model;

[0151] S6-22, Texture mapping: Map the actual image of the building or the corresponding material onto the surface of the 3D model to improve its realism and visual effect. Depending on the different parts of the building, multi-layer texture mapping technology is used;

[0152] Model format conversion:

[0153] S6-23. Model export: The generated 3D building model can be exported to common 3D formats, such as .obj, .stl, and .fbx, for easy use in other 3D modeling and visualization tools.

[0154] In summary, the principle of this embodiment is: first, remote sensing images and point cloud data of the building are obtained, and the acquired remote sensing images and point cloud data are preprocessed and normalized to improve data quality and accuracy and convert data from different sources and different scales to a unified scale or spatial range, so that the subsequent deep learning model can accurately process and learn consistent building features, and then frame field design and detail optimization are performed to ensure that the frame field generates high-precision corner points at the corners of the building to avoid the contour being too smooth. By calculating the consistency between the frame field vector and the actual building edge direction, the model is guided to optimize the accuracy of the frame field. For the corner point area of ​​the building, a specific loss function is designed to enhance the model's detection ability for corner points to ensure that the frame field accurately reflects the directional changes of these positions; and the skeleton points are aligned with the frame field during the steps of optimizing the contour, aligning the frame field, and detecting the corner points of the frame field through the active skeleton model, and the displacement of the skeleton points is adjusted through iterative optimization to ensure that it is highly consistent with the actual building boundary. The gap between the skeleton and the frame field is calculated through the error function, and a regularization term is introduced to avoid generating a contour that is too complex or smooth, thereby improving the generation accuracy and efficiency of the three-dimensional building model.

[0155] The specific embodiments described herein are merely examples of the spirit of the present invention. Those skilled in the art may make various modifications or additions to the specific embodiments described or replace them in similar ways, but they will not deviate from the spirit of the present invention or exceed the scope defined by the appended claims.

[0156] Although the hood connecting pipe 1, ventilation hole 11, barrier net 12, air inlet mesh 13, filter net 14, air outlet structure 2, first air outlet group 21, first air outlet 22, second air outlet group 23, second air outlet 24, adjustable positioning clamping assembly 3, annular threaded adjustment part 31, threaded sleeve 32, annular positioning seat 33, clamping positioning column 34, annular adjustment groove 35, threaded adjustment sleeve 36, linkage connecting rod 37, positioning clamping point part 38, annular positioning slot 39, air inlet channel 4, first air inlet channel 41, second air inlet channel 42, marking area 5, strip limiting groove 6, positioning locking rod 61, positioning button 62, positioning locking hole 63, spring loading chamber 64, top pressure spring 65 and other terms are used more frequently in this article, the possibility of using other terms is not excluded. The use of these terms is only for the purpose of more conveniently describing and explaining the essence of the present invention; interpreting them as any additional restrictions is contrary to the spirit of the present invention.

Claims

1. A method for generating a complex three-dimensional building model based on frame field deep learning, characterized in that: The method comprises the following steps: S1. Remote sensing data and point cloud data acquisition; S2, preprocessing the acquired image data; S3, data normalization processing; S4, frame field design and detail optimization; S5, active skeleton model optimization; S6. Polymorphic contour and three-dimensional model generation.

2. According to the method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 1, it is characterized in that: Step S1 includes the following steps: S01. Equipment configuration and calibration: Before collecting data, the drone or LiDAR device needs to be calibrated to ensure that the position, angle and sensor accuracy meet the requirements; S02. Area planning and path setting: Determine the aerial photography or scanning area according to the scale of the building and set the best flight or driving path to cover the entire target area; S03, Data Collection: The equipment runs along the set path and altitude, collects remote sensing images or point cloud data in real time, and records GPS information to ensure the spatial consistency of the data; S04. Data export: After the acquisition is completed, the image or point cloud data is imported into the computing platform for subsequent processing.

3. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 2, characterized in that: In step S1, the remote sensing image acquisition step includes: S1-1. Obtain high-resolution images of buildings through satellite remote sensing platforms; S1-11. Use drones equipped with high-resolution cameras to take aerial images of buildings; The point cloud data acquisition step includes: S1-2. Use LiDAR equipment to scan buildings. LiDAR uses the principle of laser ranging to generate high-precision three-dimensional point cloud data to obtain the geometric features of the building's shape; The laser radar equipment mainly scans buildings through airborne, vehicle-mounted or ground-based LiDAR systems, and in the process of LiDAR data acquisition, the equipment is usually installed on a drone, helicopter or ground vehicle, and automatically scans the designated area and generates point cloud data by setting a route or path.

4. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 1, characterized in that: Step S2 includes the following steps: Point cloud data denoising: S2-1, Bilateral Filter: It removes noise while retaining the edge features of the point cloud. The filter reduces the influence of noise points by considering the spatial neighborhood and the surface normal vector of the point cloud. S2-11, Statistical Outlier Removal: Removes outliers that deviate from the normal distribution based on the neighborhood distribution of points in the point cloud. It is suitable for processing accidental noise in large-scale point cloud data. Remote sensing image denoising: S2-2, Gaussian filter: removes noise by blurring the high-frequency part of the image, and processes remote sensing images with large changes in light intensity; S2-21, TotalVariation regularization: retain the main edge features in the image while reducing random noise; S2-22. Data Retention: After noise removal, key features of buildings are retained in the data, such as wall edges, roof lines, and details of building facades.

5. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 1, characterized in that: Step S3 includes the following steps: Point cloud data normalization: S3-1. Scale normalization: Ensure that the model can handle buildings of different sizes by mapping the coordinates of the point cloud data to the unit cube [0,1] range or scaling according to known building sizes; S3-11, Centering: Align the geometric center of the point cloud to the origin of the coordinate system to ensure that the model can input building geometry data in a standard posture; Image data normalization: S3-2, pixel value normalization: normalize the grayscale or color value of the image to the [0,1] interval to reduce the impact of different image brightness or contrast on the deep learning model; S3-21. Spatial alignment: Align multiple remote sensing images to ensure that the positions of buildings are consistent from different viewing angles and reduce deviations caused by differences in image angles and shooting heights.

6. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 1, characterized in that: Step S4 includes the following steps: Frame field design and optimization details: S4-1, frame field structure: Each pixel in the frame field contains four direction vectors, which point to the main direction of the building outline; S4-11 Corner point enhancement: A special corner point detection mechanism is designed in the model to ensure that the frame field generates high-precision corner points at the corners of the building to avoid overly smooth contours; Frame field training details: S4-2, Direction consistency loss: guide the model by calculating the consistency between the frame field vector and the actual building edge direction to optimize the accuracy of the frame field; S4-21. Corner enhancement loss: For the corner areas of buildings, a loss function is designed to enhance the model’s ability to detect corners and ensure that the frame field accurately reflects the directional changes of these locations.

7. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 1, characterized in that: Step S5 includes the following steps: The active skeleton model optimizes the outline to align with the frame field and the frame field detects corners: S5-1, Active Skeleton Model: gradually align the building outline with the frame field output by moving and adjusting the skeleton points; Optimize the contour steps: S5-2, initial skeleton generation: based on the four-vector field output by the frame field, the activity skeleton of the building is preliminarily generated; S5-21, skeleton adjustment: using the direction vector information in the frame field, the position of the skeleton points is optimized by gradient descent to make it consistent with the sharp angles and main directions of the building edges; S5-22, corner point detection: detect the corner positions of buildings and mark them through the corner point information of the frame field. For these corner points, use the local optimization algorithm to adjust their positions to ensure that the generated polygonal contour has high accuracy; Skeleton and frame field alignment details: S5-3, Dynamically adjust skeleton points: When aligning skeleton points with frame fields, adjust the displacement of skeleton points through iterative optimization to ensure that they are highly consistent with the actual building boundary; S5-31. Error control: The difference between the skeleton and the frame field is calculated through the error function, and a regularization term is introduced to avoid generating contours that are too complex or smooth.

8. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 1, characterized in that: In step S6, when the polygonal outline is generated, the skeleton of the building is converted into a polygonal outline by detecting the turning points and edge points of the skeleton. To avoid too many vertices, a sparseness algorithm is used to simplify non-critical vertices, thereby generating a simplified and accurate polygonal outline.

9. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 8, characterized in that: In step S6, the three-dimensional model generation is divided into three-dimensional model establishment and three-dimensional model deepening processing, wherein the three-dimensional model establishment is divided into the following steps: S6-1. Acquisition of height information: The height information of buildings can be inferred from the shadow length in remote sensing images or directly obtained through GIS data. For complex buildings, the height of multiple layers can be determined in combination with the existing building outline map; S6-11, Contour-based 3D model construction: Generate a 2D plane based on polygonal contours and expand it into a 3D model using building height information; S6-12. Multi-storey building modeling: For high-rise buildings, the model will generate multiple floors according to the number of floors of the building, and adjust the building shape of each floor to fit the actual structure; Facade details: S6-13, facade texture: Enhance the visual effect of the building by fitting the facade texture of the actual building on the generated 3D model; S6-14, Complex geometry processing: For complex geometric forms in buildings, the details of polygon outlines are adjusted according to the directional information of the frame field to ensure the accuracy of the building's geometric details.

10. The method for generating a complex-shaped three-dimensional building model based on frame field deep learning according to claim 9, characterized in that: In step S6, the three-dimensional model deepening process includes the following steps: S6-2, Smoothing: Apply the Laplace smoothing algorithm on the generated 3D model to eliminate sharp edges in the building shape while retaining key geometric features; S6-21, Denoising: Remove artifacts and redundant vertices in the model through wavelet transform to optimize the surface smoothness of the model; S6-22, Texture mapping: Map the actual image of the building or the corresponding material onto the surface of the 3D model to improve its realism and visual effect. Depending on the different parts of the building, multi-layer texture mapping technology is used; Model format conversion: S6-23. Model export: The generated 3D building model can be exported to common 3D formats.

Images