Building point cloud three-dimensional roof wireframe reconstruction method based on self-supervised learning

Abstract

The invention provides a building point cloud three-dimensional roof wireframe reconstruction method based on self-supervised learning. The building point cloud three-dimensional roof wireframe reconstruction method comprises the steps of obtaining an input initial building point cloud; performing data preprocessing and feature embedding on the initial building point cloud; performing self-supervised feature extraction and masking area reconstruction on the initial building point cloud after feature embedding by using a mask auto-encoder to obtain a reconstructed building point cloud; performing roof edge identification on the reconstructed building point cloud to obtain candidate roof edge points and an initial roof edge direction; performing edge parameter regression and matching based on the candidate roof edge points and the initial roof edge direction to obtain an optimal edge set; and processing the line segment information by adopting an edge non-maximum suppression strategy to obtain a three-dimensional roof wireframe result. According to the method, the precision and reliability of reconstruction of the three-dimensional wireframe structure are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of 3D modeling, and in particular, to a method for reconstructing a three-dimensional roof wireframe of building point clouds based on self-supervised learning. Background Art

[0002] With the continuous growth of the needs for smart city construction, urban informatization management, and spatial data digitization, high-precision and automated 3D city modeling has become an important foundation for urban management and planning. Among them, the building roof structure, as a key element in urban space modeling, has wide application values in fields such as building recognition, spatial analysis, view reconstruction, and urban navigation. However, due to the complex types, diverse shapes, and fuzzy geometric boundaries of roof structures, as well as problems such as sparsity, noise, occlusion, and corner point missing in the acquisition process of lidar point cloud data, traditional methods based on rule template fitting or supervised learning face many bottlenecks in practical applications, such as low reconstruction accuracy, poor generalization ability, and dependence on a large amount of labeled data.

[0003] Currently, some studies have attempted to use deep learning methods to extract building structure wireframes from point clouds. Especially in the roof modeling task, preliminary structure extraction is achieved through corner point detection and boundary prediction modules. However, these methods often highly rely on the supervised learning framework, and the training process requires a large amount of accurate labeled edge point and corner point information, which severely restricts their adaptability and generalizability on large-scale and highly heterogeneous urban point cloud datasets. At the same time, existing methods are prone to problems such as inaccurate structure recognition, excessive redundant edges, or boundary fitting deviation when facing scenes with complex structures, multi-angle occlusion, and incomplete edge information, affecting the final modeling quality and robustness.

[0004] In recent years, self-supervised learning (SSL), as a paradigm for feature representation learning without explicit label information, has shown strong potential in point cloud understanding tasks. By constructing appropriate pre-training tasks, such as strategies like masked point reconstruction, structure perturbation repair, or edge prediction, the model can learn geometric semantic feature representations on large-scale unlabeled data, thereby achieving higher accuracy and stronger generalization ability in downstream reconstruction tasks. However, the current application of SSL in the field of building wireframe reconstruction is still in its initial stage, and there are still many challenges in designing efficient feature extraction frameworks and structure analysis modules, especially in scenarios lacking corner point information or with sparse edges. Summary of the Invention

[0005] In view of this, the present invention provides a method for reconstructing a three-dimensional roof wireframe of building point clouds based on self-supervised learning to solve the above problems.

[0006] The present invention provides a method for three-dimensional roof wireframe reconstruction of building point clouds based on self-supervised learning, including: obtaining the input initial building point clouds; performing data preprocessing and feature embedding on the initial building point clouds; using a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the initial building point clouds after feature embedding to obtain reconstructed building point clouds; performing roof edge recognition on the reconstructed building point clouds to obtain candidate roof edge points and initial roof edge directions; based on the candidate roof edge points and initial roof edge directions, performing edge parameter regression and matching to obtain an optimal edge set; and using an edge non-maximum suppression strategy to process the line segment information to obtain a three-dimensional roof wireframe result.

[0007] In another implementation manner of the present invention, the performing data preprocessing and feature embedding on the initial building point clouds includes: performing data preprocessing on the initial building point clouds, where the data preprocessing includes coordinate normalization, filtering out abnormal noise points, and height threshold clipping processing; and using a multi-layer perceptron to perform feature mapping on the coordinates of each preprocessed point cloud to obtain an embedding vector.

[0008] In another implementation manner of the present invention, the using a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the initial building point clouds after feature embedding to obtain reconstructed building point clouds includes: using farthest point sampling and spherical query to divide the initial building point clouds after feature embedding into local point cloud segments of different scales; randomly applying a mask to the local point cloud segments and dividing the local point cloud segments into visible point clouds and masked point clouds; inputting the visible point clouds into the masked autoencoder for self-supervised feature extraction to obtain latent features; and using a decoder to perform masked area reconstruction on the masked point clouds based on the latent features to obtain reconstructed building point clouds.

[0009] In another implementation manner of the present invention, the performing roof edge recognition on the reconstructed building point clouds to obtain candidate roof edge points and initial roof edge directions includes: projecting the labeled line segment information onto the reconstructed building point clouds to obtain candidate roof edge points; collecting all candidate roof edge points and performing line segment fitting on the local edge distribution to obtain the initial roof edge direction of each edge.

[0010] In another implementation manner of the present invention, performing edge parameter regression and matching based on the candidate roof edge points and the initial roof edge direction to obtain an optimal edge set, including: performing edge parameter regression on the candidate roof edge point parameters based on the candidate roof edge points and the initial roof edge direction to obtain an edge confidence score, a direction offset, an edge direction, and an edge length; calculating a similarity based on the distance between two lines, the edge length, and the edge direction to obtain an edge similarity; performing bipartite graph matching between the candidate roof edge points and the true edges based on the edge similarity to obtain an optimal edge set.

[0011] In another implementation manner of the present invention, processing the optimal edge set by adopting an edge non-maximum suppression strategy to obtain a three-dimensional roof wireframe result, including: sorting the edges in the optimal edge set according to the edge confidence score; sequentially selecting the edge with the highest confidence and comparing its similarity with the remaining edges, and removing it if the similarity is greater than a preset threshold; until all the edges in the optimal edge set are traversed to obtain a three-dimensional roof wireframe result.

[0012] On the other hand, the present invention provides a three-dimensional roof wireframe reconstruction system for building point clouds based on self-supervised learning, including: a data acquisition module: acquiring the input initial building point cloud; a data processing module: performing data preprocessing and feature embedding on the initial building point cloud; a point cloud reconstruction module: using a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the initial building point cloud after feature embedding to obtain a reconstructed building point cloud; an edge recognition module: performing roof edge recognition on the reconstructed building point cloud to obtain candidate roof edge points and an initial roof edge direction; an edge matching module: performing edge parameter regression and matching based on the candidate roof edge points and the initial roof edge direction to obtain an optimal edge set; a result output module: processing the line segment information by adopting an edge non-maximum suppression strategy to obtain a three-dimensional roof wireframe result.

[0013] On the other hand, the present invention provides an electronic device, including: a memory, a processor, and a computer program stored on the memory and executable on the processor, and when the processor executes the computer program, the steps of a method for reconstructing a three-dimensional roof wireframe of building point clouds based on self-supervised learning as described in any one of the above are implemented.

[0014] On the other hand, the present invention provides a computer storage medium, characterized in that a computer program is stored on the computer storage medium, and when the computer program is executed by a processor, the steps in a method for reconstructing a three-dimensional roof wireframe of building point clouds based on self-supervised learning as described in any one of the above are implemented.

[0015] The 3D roof wireframe reconstruction method for building point clouds based on self-supervised learning of the present invention effectively utilizes unannotated point cloud data to learn edge semantic features, performs high-quality modeling on the 3D roof wireframe through a fine edge point regression mechanism; improves the model expression accuracy and reconstruction reliability through a redundant edge elimination strategy; overall enhances the automation level and engineering practicability of 3D building modeling, and provides efficient and stable technical support for urban digital twins, high-precision map generation, and 3D GIS systems. Brief Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for use in the description of the embodiments or the prior art. By reading the detailed description of the following embodiments, the advantages and benefits in the solutions will become clear to those skilled in the art. The drawings are only for the purpose of showing the preferred embodiments and are not considered to be a limitation of the present invention. In the drawings:

[0017] Figure 1 It is a schematic flow chart of the 3D roof wireframe reconstruction method for building point clouds based on self-supervised learning according to an embodiment of the present invention.

[0018] Figure 2 It is a schematic flow chart of the building wireframe structure reconstruction method of the pre-training - fine-tuning framework according to an embodiment of the present invention.

[0019] Figure 3 It is a graph of the edge point regression loss according to an embodiment of the present invention.

[0020] Figure 4 It is a schematic diagram of the wireframe reconstruction result on a large-scale dense dataset according to an embodiment of the present invention.

[0021] Figure 5 It is a schematic diagram of the result comparison of three conventional methods, two deep learning-based methods, and the method proposed in the present invention according to an embodiment of the present invention. Detailed Embodiments

[0022] In order to enable those skilled in the art to better understand the technical solutions in the embodiments of the present invention, the following will clearly and detailedly describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments in the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art should fall within the protection scope of the embodiments of the present invention.

[0023] Figure 1 It is a schematic flow chart of a 3D roof wireframe reconstruction method for building point clouds based on self-supervised learning provided by an embodiment of the present invention, as Figure 1As shown, this embodiment mainly includes:

[0024] S101. Obtain the input initial building point cloud.

[0025] S102. Perform data preprocessing and feature embedding on the initial building point cloud.

[0026] S103. Use a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the feature-embedded initial building point cloud to obtain a reconstructed building point cloud.

[0027] S104. Perform roof edge recognition on the reconstructed building point cloud to obtain candidate roof edge points and an initial roof edge direction.

[0028] S105. Based on the candidate roof edge points and the initial roof edge direction, perform edge parameter regression and matching to obtain an optimal edge set.

[0029] S106. Use an edge non-maximum suppression strategy to process the line segment information to obtain a three-dimensional roof wireframe result.

[0030] The three-dimensional roof wireframe reconstruction method for building point clouds based on self-supervised learning of the present invention effectively utilizes unlabeled point cloud data to learn edge semantic features, performs high-quality modeling on the three-dimensional roof wireframe through a fine edge point regression mechanism; improves the model expression accuracy and reconstruction reliability through a redundant edge elimination strategy; overall improves the automation level and engineering practicability of three-dimensional building modeling, and provides efficient and stable technical support for urban digital twins, high-precision map generation, and three-dimensional GIS systems.

[0031] In another implementation manner of the present invention, the performing data preprocessing and feature embedding on the initial building point cloud includes: performing data preprocessing on the initial building point cloud, where the data preprocessing includes coordinate normalization, filtering out abnormal noise points, and height threshold clipping processing; using a multi-layer perceptron to perform feature mapping on the coordinates of each preprocessed point cloud to obtain an embedding vector.

[0032] Exemplarily, as Figure 2 shown, perform coordinate normalization, filter out abnormal noise points, and optional height threshold clipping processing on the input initial building point cloud to obtain relatively clean three-dimensional coordinate data. If the point cloud resolution is uneven, downsampling can be performed based on farthest point sampling (FPS) to unify the point cloud scale and improve the calculation efficiency.

[0033] Use a multi-layer perceptron (MLP) to perform feature mapping on the coordinates of each point (and optional attributes such as intensity and normal vector), and output a preliminary point-level embedding vector, which will be used as the input feature for subsequent self-supervised extraction and edge point regression.

[0034] In another implementation of the present invention, the use of a Masked Autoencoder (MAE) to perform self-supervised feature extraction and masked area reconstruction on the initial building point cloud after feature embedding to obtain a reconstructed building point cloud includes: dividing the initial building point cloud after feature embedding into local point cloud segments of different scales by using farthest point sampling and spherical query; randomly applying a mask to the local point cloud segments and dividing the local point cloud segments into visible point cloud and masked point cloud; inputting the visible point cloud into the masked autoencoder for self-supervised feature extraction to obtain latent features; and using a decoder to perform masked area reconstruction on the masked point cloud based on the latent features to obtain a reconstructed building point cloud.

[0035] Exemplarily, farthest point sampling (FPS) and spherical query (BallQuery) are used to divide the point cloud into local point cloud segments (patches) of different scales to capture multi-granularity geometric information from near to far. A random masking operation is performed, randomly applying a mask to the divided point cloud segments and dividing them into "visible point cloud" and "masked point cloud". The masking ratio can be set according to the data situation, such as 50% or other appropriate values.

[0036] Perform masked autoencoder learning, only input the visible point cloud into the encoder to extract latent features, and use the decoder to reconstruct the masked part. The Chamfer Distance is used as the reconstruction loss to encourage the network to learn a general feature representation of the local structure and global distribution. Perform MAE encoding and decoding on the visible subset to learn a feature representation that can reconstruct the masked points at multiple scales.

[0037] In another implementation of the present invention, the identification of the roof edge of the reconstructed building point cloud to obtain candidate roof edge points and the initial roof edge direction includes: projecting the labeled line segment information onto the reconstructed building point cloud to obtain candidate roof edge points; collecting all candidate roof edge points and performing line segment fitting on the local edge distribution to obtain the initial roof edge direction of each edge.

[0038] Exemplarily, for edge point identification, project the true labeled line segment information onto the point cloud. If the distance from a point to the nearest line segment is less than a threshold, it is marked as a "hard edge label", and if it falls into another range, it is marked as a "soft edge label".

[0039] Collect all candidate roof edge points and use pyRANSAC-3D or a similar algorithm to perform line segment fitting on the local edge distribution to obtain the direction vector or center point coordinates of each edge for subsequent fine regression.

[0040] As Figure 3As shown, (a) is the segmentation of the input point cloud; (b) is the detailed description of the edge point loss formula and design. For the regression loss of edge points, factors such as the distance from the point to the true line segment and the distance from the candidate edge point to the projection point are comprehensively considered:

[0041] loss(edgepoint)=dist(p,Lines)+dist(C p3 ,prop)

[0042] Among them, the first term represents the distance from the input point to the nearest true edge, and the second term represents the geometric consistency between the candidate edge point and the predicted edge.

[0043] In another implementation manner of the present invention, based on the candidate roof edge points and the initial roof edge direction, edge parameter regression and matching are performed to obtain an optimal edge set, including: based on the candidate roof edge points and the initial roof edge direction, edge parameter regression is performed on the candidate roof edge point parameters to obtain an edge confidence score, a direction offset, an edge direction, and an edge length; similarity calculation is performed based on the distance between two lines, the edge length, and the edge direction to obtain an edge similarity; based on the edge similarity, bipartite graph matching is performed between the candidate roof edge points and the true edges to obtain an optimal edge set.

[0044] Exemplarily, after obtaining the initial direction, it is necessary to perform edge parameter regression on the specific parameters of each candidate edge, using four dedicated MLP models, each of which is designed to process different tensors: edge confidence score, direction offset, edge direction, and edge length. Losses such as L1, cross-entropy, and cosine similarity are used to ensure the prediction accuracy.

[0045] Edge similarity is quantified based on three key factors: the distance between two lines, the edge length, and the edge direction.

[0046] For the distance between two lines, the Hausdorff Edit Distance (HED) is used for effective pairwise node matching:

[0047]

[0048] In the formula, A and B are two edge sets, and the points associated with these edges are represented as points A i and B j .

[0049] The edge length uses the ratio of the minimum distance (dmin) between the endpoints of two edges to the length of the shorter edge segment to quantify the proximity between two edges:

[0050]

[0051] The edge direction uses cosine similarity to quantify the directional similarity between two edges:

[0052]

[0053] Therefore, the overall edge similarity is determined by an integral equation as follows:

[0054] Edges sim = α·H(A,B) + β·sim(A i ,B j ) + γ·pro(A i ,B j )

[0055] In the formula, α, β, and γ are balance coefficients that control the contributions of different similarity factors.

[0056] Based on the edge similarity, a bipartite graph matching is performed between the predicted and true edges to maximize the "correct match" score, so as to determine the optimal edge set and calculate the loss.

[0057] In another implementation of the present invention, the optimal edge set is processed by adopting an edge non-maximum suppression (NMS) strategy to obtain a three-dimensional roof wireframe result, including: sorting the edges in the optimal edge set according to the edge confidence score; sequentially selecting the edge with the highest confidence, comparing its similarity with the remaining edges, and removing it if the similarity is greater than a preset threshold; until all the edges in the optimal edge set are traversed, a three-dimensional roof wireframe result is obtained.

[0058] Exemplarily, sorting in descending order according to the predicted edge confidence, sequentially selecting the edge with the highest confidence, and comparing it with the remaining candidate edges: removing it if the similarity is greater than a preset threshold, and keeping the unique representative until all candidate edges are traversed, and outputting the final refined wireframe, that is, the three-dimensional roof wireframe result, as Figure 4 shown.

[0059] Edge confidence score and edge direction loss: The cross-entropy loss function and the cosine similarity loss function are respectively used to optimize the predicted edge confidence score and edge direction.

[0060] The formula for the confidence score loss is as follows:

[0061] loss (edge-confidence) = l ce (EP conf -EG conf )

[0062]

[0063] In the formula, here, EP conf and EGconf respectively represent the confidence scores of the predicted edge and the GT edge, while EP ed and EG ed respectively represent the corresponding edge direction components within the predicted edge set E predict and the GT edge set E GT respectively.

[0064] The final optimization objective of the present invention consists of the following loss terms

[0065] Loss final = λ1·d CD + λ2·loss (edge-point)

[0066] + λ3·loss (edge-length) + λ4·loss (edge-confidence)

[0067] + λ5·loss (edge-direction)

[0068] Here, λ represents a set of coefficients used to balance the contributions of the respective loss terms in the overall loss formula.

[0069] As Figure 5 shown, the results of three conventional methods, two DL-based methods, and the proposed method are compared. The experimental results show that on multiple real and synthetic datasets, the method proposed in the present invention can achieve or exceed the performance of some mainstream fully supervised models when only using 50% of the labeled data, significantly improving the automation level and engineering practicability of 3D building modeling, and providing a stable, efficient, and low-cost technical path for large-scale urban 3D modeling.

[0070] The present invention can complete the 3D wireframe structure reconstruction task of urban building roofs with high precision and low supervision. Without a large amount of manually labeled data, it can still effectively restore the geometric shape, edge topology, and spatial position relationship of the roof structure. Compared with the existing technical solutions based on supervised learning or template matching, the present invention introduces self-supervised pre-training and edge point regression mechanisms, and still maintains high reconstruction accuracy and robustness under difficult conditions such as complex building shapes, sparse point cloud distributions, and blurred edges. The edge point regression loss function and edge set similarity measurement strategy designed by the present invention effectively improve the model's spatial cognition ability of building structures, significantly superior to the existing line fitting or pure point feature-driven solutions.

[0071] On the other hand, the present invention provides a 3D roof wireframe reconstruction system for building point clouds based on self-supervised learning, including:

[0072] Data acquisition module: Acquire the input initial building point cloud.

[0073] Data processing module: performs data preprocessing and feature embedding on the initial building point cloud.

[0074] Point cloud reconstruction module: uses a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the initial building point cloud after feature embedding to obtain a reconstructed building point cloud.

[0075] Edge recognition module: performs roof edge recognition on the reconstructed building point cloud to obtain candidate roof edge points and an initial roof edge direction.

[0076] Edge matching module: based on the candidate roof edge points and the initial roof edge direction, performs edge parameter regression and matching to obtain an optimal edge set.

[0077] Result output module: processes the line segment information using an edge non-maximum suppression strategy to obtain a 3D roof wireframe result.

[0078] The 3D roof wireframe reconstruction system for building point clouds based on self-supervised learning of the present invention effectively utilizes unlabeled point cloud data to learn edge semantic features, performs high-quality modeling of the 3D roof wireframe through a fine edge point regression mechanism; improves the model expression accuracy and reconstruction reliability through a redundant edge elimination strategy; overall improves the automation level and engineering practicability of 3D building modeling, and provides efficient and stable technical support for urban digital twins, high-precision map generation, and 3D GIS systems.

[0079] On the other hand, an electronic device of the present invention includes: a processor, a memory, and a communication bus and a communication interface.

[0080] Wherein:

[0081] The processor, the memory, and the communication interface complete communication with each other through the communication bus.

[0082] The communication interface is used to communicate with other electronic devices or servers.

[0083] The processor is used to execute a program, and specifically can execute the steps of any one of the 3D roof wireframe reconstruction methods for building point clouds based on self-supervised learning in the above embodiments.

[0084] Specifically, the program may include program code, and the program code includes computer operation instructions.

[0085] The processor may be a central processing unit (CPU), or an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application. One or more processors included in the intelligent device may be of the same type of processor, such as one or more CPUs; or may be of different types of processors, such as one or more CPUs and one or more ASICs.

[0086] A memory for storing programs. The memory may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk memory.

[0087] The program can specifically be used to cause the processor to execute the steps for implementing any one of the methods for three-dimensional roof wireframe reconstruction of building point clouds based on self-supervised learning described in the embodiments. For the specific implementation of each step in the program, reference can be made to the steps and corresponding descriptions in the units executed by any one of the methods for three-dimensional roof wireframe reconstruction of building point clouds based on self-supervised learning in the above steps, which will not be elaborated here. Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described devices and modules can refer to the corresponding process descriptions in the foregoing method embodiments.

[0088] An exemplary embodiment of the present application also provides a non-transitory computer-readable storage medium storing computer instructions, wherein the computer instructions are used to cause a computer to execute the methods of the embodiments of the present application.

[0089] The methods according to the embodiments of the present invention described above can be implemented in hardware, firmware, or be implemented as software or computer code that can be stored in a recording medium (such as a CD ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or be implemented as computer code originally stored in a remote recording medium or a non-transitory machine-readable medium and downloaded through a network and to be stored in a local recording medium, so that the methods described herein can be processed by such software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or an FPGA). It can be understood that a computer, a processor, a microprocessor controller, or programmable hardware includes a storage component (such as RAM, ROM, flash memory, etc.) that can store or receive software or computer code, and when the software or computer code is accessed and executed by the computer, the processor, or the hardware, the methods described herein are implemented. In addition, when a general-purpose computer accesses the code for implementing the methods shown herein, the execution of the code converts the general-purpose computer into a dedicated computer for executing the methods shown herein.

[0090] Thus far, specific embodiments of the present invention have been described. Other embodiments are within the scope of the appended claims. In some cases, the acts recited in the claims may be performed in a different order and still achieve the desired result. Additionally, the processes depicted in the drawings do not necessarily require the particular order or sequential order shown to achieve the desired result.

[0091] It should be noted that all directional indications (such as up, down, left, right, rear, etc.) in the embodiments of the present invention are only used to explain the relative positional relationship between components in a specific order (as shown in the drawings). If this specific order changes, then the directional indication will change accordingly.

[0092] In the description of the present invention, the terms "first" and "second" are only used for convenience in describing different components or names, and should not be construed as indicating or implying an order relationship, relative importance, or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include at least one of such features.

[0093] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention pertains. The terms used in the description of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0094] It should be noted that although the specific embodiments of the present invention have been described in detail in conjunction with the accompanying drawings, it should not be construed as a limitation on the protection scope of the present invention. Within the scope described in the claims, various modifications and variations that can be made by those of ordinary skill in the art without creative efforts still fall within the protection scope of the present invention.

[0095] The examples of the embodiments of the present invention are intended to briefly illustrate the technical features of the embodiments of the present invention, enabling those of ordinary skill in the art to intuitively understand the technical features of the embodiments of the present invention, and are not intended as an improper limitation on the embodiments of the present invention.

[0096] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional roof wireframe reconstruction method for building point clouds based on self-supervised learning, characterized in that It includes: Obtain the input initial building point cloud; Perform data preprocessing and feature embedding on the initial building point cloud; Use a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the feature-embedded initial building point cloud to obtain a reconstructed building point cloud; Perform roof edge recognition on the reconstructed building point cloud to obtain candidate roof edge points and an initial roof edge direction; Based on the candidate roof edge points and the initial roof edge direction, perform edge parameter regression and matching to obtain an optimal edge set; Adopt an edge non-maximum suppression strategy to process the line segment information to obtain a three-dimensional roof wireframe result.

2. The method according to claim 1, characterized in that, The performing data preprocessing and feature embedding on the initial building point cloud includes: Perform data preprocessing on the initial building point cloud, and the data preprocessing includes coordinate normalization, filtering out abnormal noise points, and height threshold cropping processing; Use a multi-layer perceptron to perform feature mapping on the coordinates of each preprocessed point cloud to obtain an embedding vector.

3. The method according to claim 1, characterized in that The using a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the feature-embedded initial building point cloud to obtain a reconstructed building point cloud includes: Adopt farthest point sampling and spherical query to divide the feature-embedded initial building point cloud into local point cloud segments of different scales; Randomly apply a masking process to the local point cloud segments, and divide the local point cloud segments into visible point clouds and masked point clouds; Input the visible point cloud into the masked autoencoder for self-supervised feature extraction to obtain latent features; Use a decoder to perform masked area reconstruction on the masked point cloud based on the latent features to obtain a reconstructed building point cloud.

4. The method according to claim 1, wherein The performing roof edge recognition on the reconstructed building point cloud to obtain candidate roof edge points and an initial roof edge direction includes: Project the labeled line segment information onto the reconstructed building point cloud to obtain candidate roof edge points; Collect all candidate roof edge points, perform line segment fitting on the local edge distribution, and obtain the initial roof edge direction of each edge.

5. The method according to claim 4, characterized in that, The performing edge parameter regression and matching based on the candidate roof edge points and the initial roof edge direction to obtain an optimal edge set includes: Based on the candidate roof edge points and the initial roof edge direction, perform edge parameter regression on the candidate roof edge point parameters to obtain an edge confidence score, a direction offset, an edge direction, and an edge length; Calculate the similarity based on the distance between two lines, the edge length, and the edge direction to obtain an edge similarity; Perform bipartite graph matching between the candidate roof edge points and the real edges based on the edge similarity to obtain an optimal edge set.

6. The method according to claim 5, characterized in that The adopting an edge non-maximum suppression strategy to process the optimal edge set to obtain a three-dimensional roof wireframe result includes: Sort the edges in the optimal edge set according to the edge confidence score; Sequentially select the edge with the highest confidence, compare the similarity with the remaining edges, and eliminate it if the similarity is greater than a preset threshold; Until the edges in the optimal edge set are traversed, obtain a three-dimensional roof wireframe result.

7. A three-dimensional roof wireframe reconstruction system for building point clouds based on self-supervised learning, characterized in that, It includes: Data acquisition module: acquire the input initial building point cloud; Data processing module: perform data preprocessing and feature embedding on the initial building point cloud; Point cloud reconstruction module: use a masked autoencoder to perform self-supervised feature extraction and masked area reconstruction on the initial building point cloud after feature embedding to obtain a reconstructed building point cloud; Edge recognition module: perform roof edge recognition on the reconstructed building point cloud to obtain candidate roof edge points and an initial roof edge direction; Edge matching module: perform edge parameter regression and matching based on the candidate roof edge points and the initial roof edge direction to obtain an optimal edge set; Result output module: process the line segment information using an edge non-maximum suppression strategy to obtain a three-dimensional roof wireframe result.

8. An electronic device, characterized in that, Comprising: A memory, a processor, and a computer program stored on the memory and executable on the processor, wherein when the processor executes the computer program, the steps of a method for reconstructing a three-dimensional roof wireframe of a building point cloud based on self-supervised learning as described in any one of claims 1 to 6 are implemented.

9. A computer storage medium, characterized in that, A computer program is stored on the computer storage medium, and when the computer program is executed by a processor, the steps in a method for reconstructing a three-dimensional roof wireframe of a building point cloud based on self-supervised learning as described in any one of claims 1 to 6 are implemented.

Images