A multispectral LiDAR point cloud building three-dimensional gridding reconstruction method

Abstract

This invention discloses a method for 3D mesh reconstruction of buildings from multispectral LiDAR point clouds, belonging to the field of multispectral LiDAR point cloud processing technology. The method includes: extracting building instances from the input point cloud using a multispectral LiDAR point cloud classification method; extracting geometric features from the height map obtained from the building point cloud, and then extracting spectral features from the spectral map using a spectral adaptive convolution module; inputting the data into a spectral-geometric fusion module to fuse the features of the spectral map and the geometric map to achieve accurate prediction of 2D corner points; projecting the 2D corner points onto the height index of the height map to obtain 3D corner points, and using an attention mechanism jointly guided by multispectral and geometric features to predict the boundary structure of the 3D corner points, generating a wireframe model; finally, achieving seamless and watertight 3D mesh reconstruction of the building under the guidance of the topological relationships contained in the wireframe model. This invention can effectively reduce artifacts and planar misalignment problems during the reconstruction process.

Description

Technical Field

[0001] This invention relates to the field of multispectral LiDAR point cloud processing technology, specifically to a method for three-dimensional mesh reconstruction of buildings using multispectral LiDAR point clouds. Background Technology

[0002] In the field of remote sensing building reconstruction, 3D point cloud data, with its direct spatial structure representation capability, distinguishes itself from traditional 2D images and has become a core data source for modern digital city modeling. With the iteration of airborne LiDAR and dense matching technology, high-precision point cloud data has been widely applied in urban planning, disaster monitoring, and geographic information updates. However, limited by the sensor's overhead scanning perspective and complex physical environmental interference, building point clouds in remote sensing scenarios often suffer from sparse roof points and missing facade points, posing a significant challenge to geometric reconstruction of buildings within remote sensing point cloud scenarios.

[0003] Traditional building reconstruction methods typically rely solely on point cloud geometric features for structural segmentation, a strategy that demands high-quality point cloud datasets. For sparse buildings with missing points, the reconstruction effect based on the geometry of key regions significantly decreases. In remote sensing scenarios, point cloud-based building reconstruction is prone to artifacts and structural misalignment on the model surface. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method for 3D mesh reconstruction of buildings using multispectral LiDAR point clouds. By combining spectral features, this invention reduces the dependence of the reconstruction method on geometric features, thereby obtaining a more accurate representation of key geometric structures and improving the reconstruction quality of complex building structures in remote sensing scenarios.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for three-dimensional mesh reconstruction of buildings using multispectral LiDAR point clouds, specifically including the following steps: S1. Collect raw multispectral LiDAR point cloud data. After constructing superpoints from the raw multispectral LiDAR point cloud data, construct spectral maps and geometric maps respectively. Based on the spectral maps and geometric maps, extract spectral features and spatial features respectively. Then, fuse the spectral features and spatial features to extract building category instances and output building point clouds. S2. After normalizing the building point cloud, the farthest point FPS sampling algorithm is used to extract a fixed number of points from the normalized point cloud and output a sparse building point cloud. S3. Based on sparse building point clouds, after obtaining a multi-scale point cloud set through multi-scale sampling, the multi-scale point cloud set is subjected to high-dimensional position encoding and multi-layer perceptron mapping to extract features, generating a shape encoding containing global position encoding; the shape encoding is mapped to obtain coarse full points, which are merged with the sparse building point cloud to form a coarse point cloud; the coarse point cloud is passed through a cascaded Transformer layer to obtain enhanced features, and finally reconstructed and decoded and projected onto three-dimensional space to output a refined and complete structural point cloud. S4. Based on the refined complete structural point cloud, a height map and a multi-channel spectral map are generated through projection. The geometric features of the height map and the spectral features of the spectral map are extracted respectively. After feature fusion and non-maximum suppression constraints, two-dimensional corner points are obtained. After mapping the two-dimensional corner points to the height map and then using the spectral-guided boundary attention mechanism to predict the boundary, a wireframe model is output. S5. Based on the wireframe model, the roof topology is extracted through global vertical clustering. The minimum cyclic basis algorithm is used to identify closed loops and fit planes to complete local triangulation. The roof outline boundary is extracted and the vertical projection is used to generate the wall surface. Combined with the normal inversion of the bottom surface reconstruction, a watertight three-dimensional building mesh model is constructed to complete the three-dimensional mesh reconstruction method of multispectral LiDAR point cloud building.

[0006] Preferably, step S2 specifically includes the following steps: S2.1 Based on the building point cloud, calculate the geometric center of the point cloud and then translate it to the origin. Then, scale and normalize it according to the maximum modulus to obtain the normalized point cloud. The expression for calculating the maximum modulus is as follows: in, For the corresponding spatial coordinates, To input the number of point clouds, for index, The geometric center of the point cloud; S2.2 Based on the normalized point cloud, a fixed number of point clouds are extracted from the normalized point cloud using the farthest point sampling algorithm with a preset threshold, and the sparse building point cloud is output.

[0007] Preferably, step S3 specifically includes the following steps: S3.1 Based on sparse building point clouds, multi-scale sampling is performed by setting the threshold of the FPS algorithm and the aggregation operation to obtain a multi-scale point cloud set; S3.2 Input a multi-scale point cloud set, and obtain a high-dimensional vector by performing high-dimensional position encoding; input the high-dimensional feature vector into a multilayer perceptron to map it into high-dimensional features; The expression for mapping to high-dimensional features is as follows: in, This indicates the preset feature dimension parameters. Represents a multi-scale point cloud set. This is the first layer of the multilayer perceptron mapping. This indicates a normalization operation; S3.3. The high-dimensional features are fed into the feature extractor to generate a shape code containing global position encoding. The shape code is input into a multilayer perceptron for mapping to obtain coarse complete points for missing structures. The coarse complete points are merged with the sparse building point cloud to form a coarse point cloud. The coarse point cloud and the shape code are respectively input into the multilayer perceptron to extract features, and then input into three cascaded Transformer layers to output enhanced features. The enhanced features are then reshaped, decoded, and projected into three-dimensional space through a preset expansion ratio to obtain a refined complete structure point cloud.

[0008] Preferably, step S4 specifically includes the following steps: S4.1, Based on a refined and complete point cloud structure, through Value projection is used to output an average height map; after normalizing the average height map, an output height map is also generated. The expression for the output heightmap is as follows: in, Point cloud coordinates Minimum height of the shaft, For maximum height, It is a very small constant. This is an average height map; S4.2 Based on the height map, the spectral map is output after spectral intensity projection, spectral intensity averaging and channel normalization; The expression for the output spectrum is as follows: in, The minimum spectral intensity of the channel, The maximum spectral intensity of the channel. The original spectral intensity; S4.3 Based on the height map and spectral map, after freezing the preset number of ResNetbackone encoding layers and the preset number of Transformer encoding layers in the pre-trained model, the geometric features are extracted, the spectral adaptive convolution module extracts spectral features, the geometric features and spectral features are fused, and the two-dimensional corner points are output. S4.4 Based on two-dimensional corner points and height maps, three-dimensional corner points are obtained through three-dimensional mapping. Then, spectral-guided boundary attention is used to predict the boundary, and the prediction results are constrained by a non-maximum suppression algorithm to output a wireframe model.

[0009] Preferably, step S5 specifically includes the following steps: S5.1 Based on the wireframe model, after performing global vertical clustering and height sorting in sequence, the roof topology map is constructed by maximizing the preset vertical gap separation distance; S5.2 Based on the roof topology map, closed loops are identified through the minimum cyclic basis algorithm, and then the normal vector is calculated through singular value decomposition to obtain the roof triangular mesh model; S5.3 Based on the roof triangular mesh model, the roof outline boundary is extracted by surface value, the boundary vertices are vertically projected onto the base plane to generate the wall surface, and combined with the reconstruction of the bottom surface with reversed normal, a watertight three-dimensional mesh model is constructed.

[0010] Compared with existing technologies, this invention provides a method for three-dimensional mesh reconstruction of buildings using multispectral LiDAR point clouds, which has the following beneficial effects: This invention integrates the spectral and geometric features of multispectral LiDAR point clouds and introduces a multi-scale attention mechanism for point cloud completion, spectral adaptive convolution, and geometric fusion modules to achieve accurate corner prediction. This solves the problem that traditional reconstruction methods are prone to artifacts and structural misalignment in sparse and missing point clouds due to over-reliance on geometric information. Ultimately, it achieves high-precision, seamless, watertight 3D mesh reconstruction of complex structures in remote sensing scenes. Attached Figure Description

[0011] Figure 1 This is a flowchart of the steps of the present invention; Figure 2 This is a diagram showing the global reconstruction results of buildings based on the University of Houston dataset. Figure 3 This is a comparison diagram of the results of the present invention and existing methods. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] like Figure 1 As shown, a method for 3D mesh reconstruction of buildings using multispectral LiDAR point clouds includes the following steps: S1. Collect raw multispectral LiDAR point cloud data. After constructing superpoints from the raw multispectral LiDAR point cloud data, construct spectral maps and geometric maps respectively. Based on the spectral maps and geometric maps, extract spectral features and spatial features respectively. Then, fuse the spectral features and spatial features to extract building category instances and output building point clouds. S2. After normalizing the building point cloud, the farthest point FPS sampling algorithm is used to extract a fixed number of points from the normalized point cloud and output a sparse building point cloud. S2 specifically includes the following steps: S2.1 Based on the building point cloud, calculate the geometric center of the point cloud and then translate it to the origin. Then, scale and normalize it according to the maximum modulus to obtain the normalized point cloud. First, based on the building point cloud, calculate the geometric center of the point cloud. The calculation formula is as follows: in, For the corresponding spatial coordinates, To input the number of point clouds, for The index.

[0014] After calculating the geometric center of the point cloud, translate the point cloud to the origin, based on the maximum modulus. Scaling normalization is performed to decentralize and scale the input point cloud. The maximum modulus is calculated using the following formula: The final normalized coordinates are: , and the normalized point cloud composed of normalized coordinates.

[0015] S2.2 Based on the normalized point cloud, the Farthest Point Sampling (FPS) algorithm is used to extract a fixed number of points from the normalized point cloud and output a sparse building point cloud. In this embodiment, the sampling threshold for the FPS algorithm is set to 2048.

[0016] S3. Based on sparse building point clouds, after obtaining a multi-scale point cloud set through multi-scale sampling, the multi-scale point cloud set is subjected to high-dimensional position encoding and multi-layer perceptron mapping to extract features, generating a shape encoding containing global position encoding; the shape encoding is mapped to obtain coarse full points, which are merged with the sparse building point cloud to form a coarse point cloud; the coarse point cloud is passed through a cascaded Transformer layer to obtain enhanced features, and finally reconstructed and decoded and projected onto three-dimensional space to output a refined and complete structural point cloud. S3 specifically includes the following steps: S3.1 Based on sparse building point clouds, multi-scale sampling is performed by setting the threshold of the FPS algorithm and the aggregation operation to obtain a multi-scale point cloud set; In this embodiment, the input sparse building point cloud is processed by setting the threshold values ​​of the FPS algorithm and aggregation operation to obtain point clouds with three representations of quantity: 512, 256, and 128. These point clouds are then fed into three multi-scale feature extraction networks.

[0017] S3.2 Input a multi-scale point cloud set, and obtain a high-dimensional vector by performing high-dimensional position encoding; input the high-dimensional feature vector into a multilayer perceptron to map it into high-dimensional features; For the input multi-scale point cloud set, high-dimensional location encoding is first performed, and the encoding method is as follows: in, This represents the preset feature dimension parameter, whose value is related to the number of FPS sampling points, and is 512, 256, and 128 respectively. Represents a series of cosine and sine function operations. Represents a multi-scale point cloud set. This represents a multilayer perceptron mapping.

[0018] After obtaining the high-dimensional positional encoding, a multilayer perceptron is used to map the high-dimensional vector to the corresponding high-dimensional feature. The calculation expression is as follows: in, The values ​​are the same as those for the encoding operations mentioned above, which are 512, 256, and 128 respectively. This indicates a normalization operation. This is the first layer of the multilayer perceptron mapping. This is the mapping for the second layer of the multilayer perceptron.

[0019] S3.3. Feed the high-dimensional features into the feature extractor to generate a shape code containing global position encoding; input the shape code into a multilayer perceptron for mapping to obtain coarse-complete points for missing structures; merge the coarse-complete points with the sparse building point cloud to form a coarse point cloud; input the coarse point cloud and the shape code into the multilayer perceptron respectively to extract features, and then input them into three cascaded Transformer layers to output enhanced features. Reshape, decode and project them into three-dimensional space through a preset expansion ratio to obtain a refined complete structure point cloud. The high-dimensional features are fed into a feature extractor, which outputs a shape code that includes a global positional encoding. The shape code is a dimensional... The location encoding vector is 3D, where each column represents the location encoding of key structural features of the building. The shape encoding is then mapped through a multilayer perceptron to obtain coarsely complete points for missing structures, which are then merged with the original sparse building point cloud to obtain a coarsely complete structure point cloud. The coarsely complete structure point cloud and the shape encoding are then processed by a multilayer perceptron for feature extraction and fed into three cascaded Transformer layers. The output enhanced features are projected back into 3D coordinate space through a reshaping operation. Here, this invention sets the point cloud augmentation factor required for the reshaping operation to 8, resulting in a refined complete structure point cloud containing 1024*8=8192 points.

[0020] S4. Based on the refined complete structural point cloud, a height map and a multi-channel spectral map are generated through projection. The geometric features of the height map and the spectral features of the spectral map are extracted respectively. After feature fusion and non-maximum suppression constraints, two-dimensional corner points are obtained. After mapping the two-dimensional corner points to the height map and then using the spectral-guided boundary attention mechanism to predict the boundary, a wireframe model is output. S4 specifically includes the following steps: S4.1, Based on a refined and complete point cloud structure, through Value projection is used to output an average height map; after normalizing the average height map, an output height map is also generated. This invention is based on the refined and complete structural point cloud obtained in S3, and firstly requires following... The coordinate axis values ​​are projected onto a two-dimensional plane to generate a corresponding height map. Specifically, for each pixel in the image to be projected... ), calculate the average z-value of all points falling within the pixel grid. Let the set be... For falling pixels ( The expression for outputting the average height map from a point set is: in, This indicates the number of points within that pixel. If there are no points within the pixel, then... . For the first The height of each point.

[0021] Finally, the calculated average heightmap is normalized, and the expression for the output heightmap is as follows: in, Point cloud coordinates Minimum height of the shaft, For maximum height, It is a very small constant. This is used to prevent division by zero errors.

[0022] S4.2 Based on the height map, the spectral map is output after spectral intensity projection, spectral intensity averaging and channel normalization; This invention, after obtaining the height map, generates a spectral map through spectral intensity projection and channel normalization. The difference lies in calculating the falling pixel for each selected spectral channel. The original spectral intensity is obtained by averaging the spectral intensities of all points. The specific calculation method is as follows: in, This indicates the use of 3 spectral channel features. Indicates spectral intensity.

[0023] Similarly, each channel is independently normalized and mapped to... Within the interval, the expression for the output spectrum is as follows: in, The minimum spectral intensity of the channel, This represents the maximum spectral intensity of the channel.

[0024] S4.3 Based on the height map and spectral map, after freezing the 4 ResNetbackone encoding layer and 1 Transformer encoding layer in the pre-trained model, the geometric features are extracted, the spectral adaptive convolution module extracts spectral features, the features are fused and non-maximum suppression is performed in sequence, and then the two-dimensional corner points are output. This invention first freezes the four ResNet backone encoding layers and one Transformer encoding layer in the pre-trained model. Then, geometric features are extracted from the resulting heightmap, and spectral features are extracted from the spectral map using a spectral adaptive convolution module, which includes two Conv2d convolutional layers, one normalization layer, and a ReLU activation function. After extracting the geometric and spectral features, a spectral-geometric fusion module fuses the features from the spectral and geometric maps, and a non-maximum suppression algorithm is used to constrain the corner prediction process. After extracting the required features, the ResNet backone encoding layers and Transformer layers are unfrozen layer by layer to adapt to the feature changes, resulting in accurate two-dimensional corner points.

[0025] S4.4 Based on two-dimensional corner points and height maps, three-dimensional corner points are obtained through three-dimensional mapping. Then, spectral-guided boundary attention is used to predict the boundary, and the prediction results are constrained by a non-maximum suppression algorithm to output a wireframe model. The two-dimensional corner points predicted by S4.3 are combined with the height map obtained by S4.1 to map the two-dimensional corner points into three-dimensional corner points. For the obtained three-dimensional corner points, a spectral-guided boundary attention mechanism is used to predict the boundary between any two points in the three-dimensional corner point set, and the non-maximum suppression algorithm is also used to constrain the prediction process. Finally, a wireframe model that can accurately represent the global structure of the building is generated.

[0026] S5. Based on the wireframe model, the roof topology is extracted through global vertical clustering. The minimum cyclic basis algorithm is used to identify closed loops and fit planes to complete local triangulation. The roof outline boundary vertical projection is extracted to generate the wall surface. Combined with the normal inversion bottom surface reconstruction, a watertight three-dimensional building mesh model is constructed to complete the three-dimensional mesh reconstruction method of multispectral LiDAR point cloud building. S5 specifically includes the following steps: S5.1 Based on the wireframe model, after performing global vertical clustering and height sorting in sequence, the roof topology map is constructed by maximizing the preset vertical gap separation distance; Based on the wireframe model obtained from S4, this invention first uses a global vertical clustering algorithm to sort the vertex heights of the roof topology. By setting the maximum vertical gap to 3 meters, the optimal separation surface is determined, thereby isolating and constructing a topology map that only contains the roof portion. .

[0027] S5.2 Based on the roof topology map, closed loops are identified through the minimum cyclic basis algorithm, and then the normal vector is calculated through singular value decomposition to obtain the roof triangular mesh model; Specifically, this invention uses the Minimum Cyclic Basis (MCB) algorithm to identify closed loops as candidate planes through plane fitting and local triangulation, and calculates robust normal vectors through singular value decomposition (SVD) to eliminate jitter, thereby completing the optimal plane fitting and local triangulation to obtain the roof triangular mesh model.

[0028] S5.3 Based on the roof triangular mesh model, the roof outline boundary is extracted by surface value, the boundary vertices are vertically projected onto the base plane to generate the wall surface, and combined with the reconstruction of the bottom surface with reversed normal, a watertight three-dimensional mesh model is constructed. The present invention generates a watertight three-dimensional mesh model based on extracting the roof outline boundary from the surface value, projecting the boundary vertices vertically onto the base plane to generate the wall surface, and combining the reconstruction of the bottom surface with the normal inversion, finally forming a building mesh model that strictly satisfies the Euler characteristic number.

[0029] To verify the effectiveness of the method of the present invention, the following experiments were conducted and explained: 1. Experimental setup: This invention is implemented based on the PyTorch 2.1.1 framework, using the AdamW optimizer with an initial learning rate of 0.0004. The learning rate is dynamically adjusted using cosine annealing in OneCycleLR, and the total number of epochs is set to 800. A total of 1320 pairs of input data from HU (including original point clouds and ground truth wireframe models) are used as training data. 132 pairs of point cloud data from HU are used as test data.

[0030] 2. Experimental Results: As shown in Table 1, the method of the present invention achieves optimal point distance RMSE (root mean square error) and normal vector RMSE. This indicates that the method of the present invention can better reconstruct selected multispectral LiDAR point cloud buildings compared to existing technologies. Figure 2 , Figure 3 The visualization results show the global reconstruction results of the method of the present invention and the comparison with the reconstruction results of other methods, which more intuitively demonstrates the advantages of the invention.

[0031] Table 1: Comparison of Reconstruction Accuracy of the University of Houston Dataset The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A multi-spectral LiDAR point cloud building 3D meshing reconstruction method, characterized in that, The method comprises the following steps: S1, collecting original multi-spectral LiDAR point cloud data, constructing a super point based on the original multi-spectral LiDAR point cloud data, and respectively constructing a spectrum graph and a geometry graph; based on the spectrum graph and the geometry graph, respectively extracting spectral features and spatial features, and then fusing the spectral features and the spatial features to extract building class instances, and outputting building point clouds; S2, after normalizing the building point clouds, extracting a preset number of point clouds from the normalized point clouds using the farthest point FPS sampling algorithm, and outputting sparse building point clouds; S3, based on the sparse building point clouds, obtaining a multi-scale point cloud set through multi-scale sampling, and then performing high-dimensional position coding and multi-layer perception machine mapping on the multi-scale point cloud set to extract features and generate shape coding containing global position coding; the shape coding is mapped to obtain coarse completion points, which are combined with the sparse building point clouds to form rough point clouds; The rough point clouds are output through a cascade Transformer layer to output enhanced features, and finally are reshaped and decoded to project to a three-dimensional space to output fine complete structure point clouds; S4, based on the fine complete structure point clouds, generating a height map and a spectrum graph through projection; respectively extracting geometric features of the height map and spectral features of the spectrum graph, and then performing feature fusion and non-maximum suppression constraint to obtain two-dimensional corner points; after mapping the two-dimensional corner points to three-dimensional corner points in combination with the height map, performing boundary prediction using a spectrum-guided boundary attention mechanism, and outputting a wireframe model; S5, based on the wireframe model, extracting a roof topology through global vertical clustering, identifying a closed loop using a minimum cycle basis algorithm and fitting a plane, and completing local triangulation; extracting a roof contour boundary vertical projection to generate a wall surface, combining a normal inversion bottom surface reconstruction, and constructing a watertight three-dimensional building mesh model, thereby completing a multi-spectral LiDAR point cloud building three-dimensional mesh reconstruction method.

2. The method of claim 1, wherein, In the S2, the following steps are specifically included: S2.1, based on the building point clouds, calculating a point cloud geometric center and then translating to the origin, and performing scaling normalization according to a maximum module length to obtain normalized point clouds; The calculation expression of the maximum module length is as follows: wherein, is the corresponding spatial coordinate, is the number of input point clouds, is the index of is the point cloud geometry center;​ S2.2, based on the normalized point clouds, extracting a fixed number of point clouds from the normalized point clouds using a farthest point sampling algorithm with a preset threshold, and outputting sparse building point clouds.

3. The method of claim 1, wherein, In the S3, the following steps are specifically included: S3.1, based on the sparse building point clouds, performing multi-scale sampling by setting the threshold values of the FPS algorithm and the aggregation operation to obtain a multi-scale point cloud set; S3.2, inputting the multi-scale point cloud set, performing high-dimensional position coding to obtain a high-dimensional vector; and inputting the high-dimensional feature vector into a multi-layer perception machine to map it into a high-dimensional feature; The expression of mapping into a high-dimensional feature is as follows: wherein, denotes a preset feature dimension parameter, denotes a multi-scale point cloud set, is a first layer multi-layer perception mapping, denotes a normalization operation; S3.

3. The high-dimensional features are fed into the feature extractor to generate a shape code containing global position encoding. The shape code is input into a multilayer perceptron for mapping to obtain coarse complete points for missing structures. The coarse complete points are merged with the sparse building point cloud to form a coarse point cloud. The coarse point cloud and the shape code are respectively input into the multilayer perceptron to extract features, and then input into three cascaded Transformer layers to output enhanced features. The enhanced features are then reshaped, decoded, and projected into three-dimensional space through a preset expansion ratio to obtain a refined complete structure point cloud.

4. The method of claim 1, wherein, S4 specifically includes the following steps: S4.1, based on the refined complete structure point cloud, by value projection, output the average height map; after normalizing the average height map, output the height map; The expression for the output heightmap is as follows: wherein, is the minimum height of the axis, is the minimum height of the axis, is the maximum height, is the minimum height of the axis, is the average height map; S4.2 Based on the height map, the spectral map is output after spectral intensity projection, spectral intensity averaging and channel normalization; The expression for the output spectrum is as follows: wherein, is the minimum spectral intensity of the channel, is the maximum spectral intensity of the channel, is the original spectral intensity; S4.3 Input height map and spectral map. After freezing the preset number of ResNetbackone encoding layers and the preset number of Transformer encoding layers in the pre-trained model, extract geometric features in sequence, extract spectral features through the spectral adaptive convolution module, fuse geometric features and spectral features, and output two-dimensional corner points. S4.4 Based on two-dimensional corner points and height maps, three-dimensional corner points are obtained through three-dimensional mapping. Then, spectral-guided boundary attention is used to predict the boundary, and the prediction results are constrained by a non-maximum suppression algorithm to output a wireframe model.

5. The method of claim 1, wherein, S5 specifically includes the following steps: S5.1 Based on the wireframe model, after performing global vertical clustering and height sorting in sequence, the roof topology map is constructed by maximizing the preset vertical gap separation distance; S5.2 Based on the roof topology map, closed loops are identified through the minimum cyclic basis algorithm, and then the normal vector is calculated through singular value decomposition to obtain the roof triangular mesh model; S5.3 Based on the roof triangular mesh model, the roof outline boundary is extracted by surface value, the boundary vertices are vertically projected onto the base plane to generate the wall surface, and combined with the reconstruction of the bottom surface with reversed normal, a watertight three-dimensional mesh model is constructed.

6. The method of claim 4, wherein, In S4.3, the spectral adaptive convolution module sequentially includes two Conv2d convolutional layers, one normalization layer, and a ReLU activation function.

7. The method of claim 3, wherein, In S3.1, the multi-scale point cloud includes three scales: 512, 256, and 128.

Images