Underground engineering three-dimensional reconstruction and grid extraction method based on 3D Gaussian splashing

Abstract

The invention relates to an underground engineering three-dimensional reconstruction and grid extraction method based on 3D Gaussian splashing, and the method comprises the following steps: S1, obtaining a static scene image: photographing a plurality of static scene images from different angles and heights of an underground engineering through photographing equipment or an unmanned plane; s2, sparse point cloud generation: processing the image set collected in the step S1; s3, initializing to generate 3D Gaussian: taking the sparse point clouds obtained in the step S2 as input, dividing the point clouds into a plurality of regions, and generating one or more 3D Gaussian models in each region based on the world space position mean value and distribution characteristics of the point clouds; s4, optimizing a 3D Gaussian parameter; s5, extracting a grid; and S6, model lightweight: in order to reduce the storage demand of the model and accelerate the rendering speed, performing lightweight processing on the model. According to the method, efficient generation and real-time rendering of the three-dimensional model of the underground engineering can be realized while the reconstruction quality is maintained, and digital support is provided for design, construction and management of the underground engineering.

Description

Technical Field

[0001] The present invention relates to the technical field of computer graphics and three-dimensional reconstruction, and more specifically, to a three-dimensional reconstruction and mesh extraction method for underground engineering based on 3D Gaussian splashing. Background Art

[0002] Underground engineering visualization refers to the creation of a three-dimensional model of an underground project based on geological survey data, engineering design data, etc., and the display of graphic images on a computer platform. This process is designed to allow users to browse the structure and internal environment of underground projects in an intuitive way, or interact with external data to achieve status monitoring and response. The application range of underground engineering visualization technology is wide, including but not limited to the design, construction monitoring, maintenance and management of complex underground facilities such as tunnels, mines, and pipeline networks. With the development of computer technologies such as the Internet of Things, Web3D, and three-dimensional reconstruction, web-based underground engineering visualization systems have become possible. Internet of Things technology allows sensors and other devices to collect real-time data in underground environments, which can be integrated into three-dimensional models to provide important information about the status of underground projects. Web3D technology allows three-dimensional models to be presented in web browsers without the need to install additional software or plug-ins, thereby improving the ease of use and accessibility of the system. Three-dimensional reconstruction technology refers to the process of generating three-dimensional models using data collected by various sensors (such as lidar, cameras, etc.). The combination of these technologies provides a comprehensive digital solution for underground engineering, enabling engineers, designers, and managers to remotely monitor and manage engineering projects.

[0003] Although 3D reconstruction technology has made significant progress in many fields, it still faces many challenges when dealing with underground projects. Although traditional 3D reconstruction methods (such as laser scanning, photogrammetry, motion structure recovery (SfM), multi-view stereo vision (MVS), etc.) can provide a certain reconstruction quality, it is difficult to achieve ideal reconstruction effects in the specific environment of underground projects due to the following reasons: Insufficient lighting: Underground projects usually have weak light, which leads to low image quality obtained by methods such as photogrammetry, affecting the reconstruction accuracy. Lack of texture: The surfaces of materials such as rocks and soils commonly found in underground environments often lack obvious texture features, which makes it difficult for reconstruction algorithms based on texture matching to accurately correspond to point clouds or images. Many obstacles: The internal structure of underground projects is complex, and there are a large number of obstacles such as supporting structures, cables, and pipelines. These obstacles block the field of view of the sensor and cause occlusion problems during reconstruction. Environmental noise: Due to the closed nature of the underground environment, data collection is easily affected by various noises, such as echo interference and electromagnetic interference, which further reduces the data quality. These factors work together to cause a large number of holes and artifacts in traditional 3D reconstruction methods when dealing with underground projects, which cannot meet the needs of high-quality reconstruction. On the other hand, although deep learning-based methods such as NeRF can achieve high-quality reconstruction under certain conditions, they suffer from deficiencies in training time and rendering performance. Methods such as NeRF usually require a lot of computing resources and a long training process, which is an obvious bottleneck for real-time applications. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a method for three-dimensional reconstruction and mesh extraction of underground engineering based on 3D Gaussian splattering, which can quickly achieve high-quality three-dimensional reconstruction, effectively reduce holes and artifacts, and improve rendering performance. While maintaining the reconstruction quality, it can achieve efficient generation and real-time rendering of the three-dimensional model of the underground engineering, thereby providing more reliable digital support for the design, construction and management of underground engineering.

[0005] The technical solution adopted by the present invention to solve the technical problem is: constructing a 3D reconstruction and mesh extraction method for underground engineering based on 3D Gaussian splashing, comprising the following steps: S1. Obtain static scene images: Use photographic equipment or drones to take multiple static scene images from different angles and heights of the underground project; S2, generating sparse point cloud: processing the image set collected in step S1; S3, initialization generation of 3D Gaussian: taking the sparse point cloud obtained in step S2 as input, dividing the point cloud into multiple regions, and generating one or more 3D Gaussian models in each region based on the world space position mean and distribution characteristics of the point cloud; S4, optimize 3D Gaussian parameters: involves iteratively adjusting the center coordinates, covariance matrix and opacity parameters of the Gaussian model to make the model better fit the original point cloud data; S5, grid extraction; S6. Model lightweighting: To reduce the storage requirements of the model and speed up the rendering, the model is lightweighted.

[0006] According to the above scheme, in step S1, the static scene image covers the key areas of the project including entrances, passages, intersections, and important equipment to ensure that comprehensive information can be obtained in subsequent processing.

[0007] According to the above scheme, in step S2, the collected image set is processed using the SfM algorithm; The SfM algorithm analyzes feature points in an image and tracks the position changes of the feature points in different images, thereby inferring the motion trajectory of the camera and the three-dimensional structure of the scene. The SfM algorithm generates a sparse point cloud model, which contains the three-dimensional coordinate information of key points in the scene. According to the above scheme, in step S3, the point cloud is divided into multiple regions by K-means clustering or a grid-based partitioning method; The 3D Gaussian model has an ellipsoidal volume appearance, and the center coordinates, shape and opacity parameters controlled by the 3D covariance matrix are all calculated based on the point cloud data. According to the above scheme, step S4 involves iteratively adjusting the center coordinates, covariance matrix and opacity parameters of the Gaussian model so that the model better fits the original point cloud data; During the optimization process, in order to ensure the convergence speed and optimization effect, the optimization algorithm of gradient descent, Newton method or conjugate gradient is used.

[0008] According to the above scheme, in step S5, the method for mesh extraction includes the following steps: S501, constructing a Gaussian layer: constructing a Gaussian distribution around each vertex of the initial mesh according to a Gaussian function; For each vertex , define a Gaussian function: , Where x is any point in space, Vertex The position vector of is a 3x3 covariance matrix, the Gaussian function defines the width and direction of the Gaussian distribution, and the covariance matrix is ​​dynamically adjusted according to the position of the vertex and the required level of detail; S502, adaptive thickness adjustment: using a feature detection algorithm to identify the features of edges and corners in the scene, and automatically adjusting the thickness of the Gaussian layer according to the detection results of the features; When a region is detected as an edge or corner, the thickness of the Gaussian layer in the region is reduced to better capture these features; when a region is relatively flat, the thickness of the Gaussian layer in the region is increased to reduce the computational cost; S503, final mesh extraction: After Gaussian layer refinement and optimization, the final triangular mesh is extracted from the Gaussian distribution using the Poisson reconstruction method.

[0009] According to the above scheme, in step S502, a feature detection algorithm is used, including a Canny edge detector and a Shi-Tomasi corner detector. The mathematical basis of the feature detection algorithm is briefly described as follows: Canny edge detector: Based on the smoothing process of Gaussian filtering, it determines the edge by calculating the gradient amplitude and direction, using non-maximum suppression and double threshold detection. The core lies in gradient calculation, using the Sobel operator, and the gradient calculation formula is:

[0010] in, are the gradients in the horizontal and vertical directions respectively; Shi-Tomasi corner detector: By calculating the autocorrelation matrix of each pixel in the image and finding points with two larger eigenvalues ​​as corners, the autocorrelation matrix M can be expressed as:

[0011] in, are the gradients of the image in the x and y directions respectively; After feature detection, the thickness of the Gaussian layer is adjusted according to the detection results; assuming that the detected feature intensity is , the adjustment of Gaussian layer thickness T(x,y) is expressed as:

[0012] in, is the minimum value of Gaussian layer thickness, is the base thickness, To adjust the coefficient, adjust it according to the specific application scenario; After constructing the Gaussian layer, the Poisson reconstruction algorithm is used for refinement; the core of the Poisson reconstruction algorithm is to solve a linear system, and its general form is expressed as:

[0013] Among them, L is the system matrix, which is related to the Laplacian operator of the grid; is the position of the mesh vertex to be solved; is the vector associated with the input point cloud data; In order to further improve the mesh quality, a regularization term is added and the objective function is expressed as:

[0014] Among them, the first term is the data fitting term, which ensures that the grid is aligned with the Gaussian layer data; the second term is the regularization term, is a regularization function used to smooth the mesh surface and reduce noise; is the regularization coefficient.

[0015] According to the above scheme, in step S6, in order to reduce the storage requirement of the model and speed up the rendering speed, the model is lightweighted by pruning redundant Gaussian points and performing knowledge distillation.

[0016] The present invention also provides a 3D reconstruction and mesh extraction system for underground engineering based on 3D Gaussian splashing, comprising a mesh extraction module, a model lightweight processing module and a web page real-time rasterization rendering module; The mesh extraction module is used to extract a high-quality mesh model from the 3D Gaussian representation; The model lightweight processing module performs lightweight processing on the Gaussian model, thereby reducing the user's model loading time and storage overhead and improving the user experience; The web-side real-time rasterization rendering module is used to implement a 3D Gaussian splash rasterization renderer on the web-side, supporting pure grid representation and grid + 3D Gaussian mixed representation.

[0017] The implementation of the underground engineering three-dimensional reconstruction and mesh extraction method based on 3D Gaussian splashing of the present invention has the following beneficial effects: 1. The present invention can significantly improve the level of detail on the model surface while maintaining the accuracy of the overall shape. By adding a Gaussian distribution layer on the basis of the existing grid and adaptively adjusting the width of the Gaussian distribution according to the surface features, it ensures that sufficient details are provided where needed, so that the reconstructed model can better capture the true geometric characteristics of the object surface, especially the edges, sharp features and areas with rich details. Traditional 3D reconstruction methods often cause a large number of holes and artifacts in the reconstruction results due to problems such as insufficient lighting, lack of texture, many obstacles, and environmental noise. The present invention uses feature-aware refinement technology to strengthen the Gaussian distribution density around key features, effectively enhancing the expression of details; 2. The present invention has excellent performance in computational efficiency. Compared with other complex detail enhancement technologies, it maintains high computational efficiency while ensuring details. Through efficient fusion and reconstruction algorithms, high-quality triangular meshes are extracted from the fused Gaussian distribution, making the entire reconstruction process not only fast but also able to process large-scale data sets, which is crucial for real-time applications, because the three-dimensional reconstruction of underground engineering often requires the generation of a large amount of data in a short period of time, and this requirement can be met.

[0018] 3. The present invention has extremely high flexibility and can be flexibly applied to different types of 3D reconstruction tasks. It is particularly suitable for application scenarios that require a high degree of realism and detailed reproduction. Whether it is a tunnel, a mine or a complex underground pipeline network, it can provide a meticulous reconstruction effect. It is not limited to underground engineering, but can also be extended to other fields that require high-quality 3D reconstruction, such as urban planning, cultural relics protection, virtual reality, etc. 4. The present invention also has significant value at the commercial and social levels. For underground engineering companies, high-quality 3D reconstruction models not only help improve the accuracy of design and construction, but also significantly reduce the cost increase and construction delays caused by information asymmetry. Through real-time rendering and online display, project management becomes more efficient, customer satisfaction will also increase, and it also promotes collaboration within the industry. Through the real-time rendering function of the web page, different project participants can easily share information and work together, improving the overall execution efficiency of the project. For scientific research institutions and educational institutions, this method also provides a powerful tool for teaching and research, allowing students and researchers to more intuitively understand and analyze various complex situations of underground engineering.

[0019] 5. The present invention not only achieves high-quality 3D reconstruction and efficient mesh extraction in terms of technology, but also promotes the digital transformation of the underground engineering industry at the commercial and social levels, improves work efficiency and coordination, and lays a solid foundation for the sustainable development of the industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which: Figure 1 It is a system framework diagram of the underground engineering three-dimensional reconstruction and mesh extraction method based on 3D Gaussian splashing of the present invention; Figure 2 It is a flow chart of mesh extraction steps of the underground engineering three-dimensional reconstruction and mesh extraction method based on 3D Gaussian splashing of the present invention; Figure 3 It is the overall workflow diagram of the underground engineering three-dimensional reconstruction and grid extraction method based on 3D Gaussian splashing of the present invention. DETAILED DESCRIPTION

[0021] In order to have a clearer understanding of the technical features, purposes and effects of the present invention, specific embodiments of the present invention are now described in detail with reference to the accompanying drawings.

[0022] like Figure 1-3 As shown, the underground engineering 3D reconstruction and mesh extraction method based on 3D Gaussian splashing of the present invention comprises the following steps: S1. Acquire a static scene image.

[0023] Use professional photography equipment or drones to take multiple static scene images from different angles and heights of the underground project. The static scene images cover all key areas of the project, including entrances, passages, intersections, important equipment, etc., to ensure that subsequent processing can obtain comprehensive information. At the same time, pay attention to lighting conditions when shooting to ensure that the image is clear, without blur and shadow interference.

[0024] S2. Generate sparse point cloud The image set collected in step S1 is processed using the SfM algorithm. The SfM algorithm infers the camera's motion trajectory and the 3D structure of the scene by analyzing the feature points in the image and tracking the position changes of these feature points in different images. Ultimately, the algorithm generates a sparse point cloud model that contains the 3D coordinate information of the key points in the scene.

[0025] S3, initialize and generate 3D Gaussian After the sparse point cloud is obtained in step S2, it is used as input and the point cloud is divided into multiple regions by a specific algorithm. In this embodiment, K-means clustering or a grid-based division method is used. Then, in each region, one or more 3D Gaussian models are generated based on the world space position mean and distribution characteristics of the point cloud. Each 3D Gaussian model has an ellipsoidal volume appearance, and its center coordinates, shape (controlled by the 3D covariance matrix) and opacity and other parameters are calculated based on the point cloud data.

[0026] S4. Optimize 3D Gaussian parameters In order to improve the quality of scene representation, the parameters of the 3D Gaussian model need to be optimized. This involves iteratively adjusting parameters such as the center coordinates, covariance matrix, and opacity of the Gaussian model to make the model better fit the original point cloud data. During the optimization process, optimization algorithms such as gradient descent, Newton method, or conjugate gradient can be used to ensure convergence speed and optimization effect.

[0027] S5. Grid extraction: The network extraction method comprises the following steps: S501, construct Gaussian layer: A Gaussian distribution is constructed around each vertex of the initial mesh according to the Gaussian function. , define a Gaussian function , where x is any point in space, Vertex The position vector of It is a 3x3 covariance matrix. The Gaussian function defines the width and direction of the Gaussian distribution. The covariance matrix can be dynamically adjusted based on the location of the vertex and the required level of detail.

[0028] Using feature point detection algorithms (such as SIFT, SURF, etc.) to identify significant feature points in the scene, the eigenvalues ​​of the covariance matrix can be reduced near these feature points, thereby reducing the width of the Gaussian distribution to retain more details. Conversely, in flat areas, the eigenvalues ​​of the covariance matrix can be increased to increase the width of the Gaussian distribution, thereby reducing the amount of calculation.

[0029] S502, Adaptive thickness adjustment: In order to further improve the accuracy and efficiency of the grid, a feature detection algorithm is used to identify features such as edges and corners in the scene. This embodiment uses a Canny edge detector, a Shi-Tomasi corner detector, etc. Then, the thickness of the Gaussian layer is automatically adjusted according to the detection results of these features. If a certain area is detected as an edge or corner, the thickness of the Gaussian layer in the area is reduced to better capture these features; if a certain area is relatively flat, the thickness of the Gaussian layer in the area is increased to reduce the computational cost.

[0030] Using feature detection algorithms, such as Canny edge detector and Shi-Tomasi corner detector, the mathematical basis of feature detection algorithms are briefly described as follows: Canny edge detector: Based on the smoothing process of Gaussian filtering, it determines the edge by calculating the gradient amplitude and direction, using non-maximum suppression and double threshold detection. Its core lies in gradient calculation, using the Sobel operator, and the gradient calculation formula is:

[0031] in, are the gradients in the horizontal and vertical directions respectively.

[0032] Shi-Tomasi corner detector: By calculating the autocorrelation matrix of each pixel in the image, and finding points with two larger eigenvalues ​​as corners. The autocorrelation matrix M can be expressed as:

[0033] in, are the gradients of the image in the x and y directions, respectively.

[0034] After feature detection, the thickness of the Gaussian layer is adjusted according to the detection results. Assume that the detected feature strength is , the adjustment of Gaussian layer thickness T(x,y) can be expressed as:

[0035] in, is the minimum value of Gaussian layer thickness, is the base thickness, To adjust the coefficient, adjust it according to the specific application scenario.

[0036] After constructing the Gaussian layer, the Poisson reconstruction algorithm is used for refinement. The core of the Poisson reconstruction algorithm is to solve a linear system, which is generally expressed as:

[0037] Where L is the system matrix, usually related to the Laplacian of the grid; is the position of the mesh vertex to be solved; is the vector associated with the input point cloud data.

[0038] In order to further improve the mesh quality, a regularization term is added and the objective function is expressed as:

[0039] Among them, the first term is the data fitting term, which ensures that the grid is aligned with the Gaussian layer data; the second term is the regularization term, is a regularization function used to smooth the mesh surface and reduce noise. is the regularization coefficient. The specific form of the regularization function may include the sum of squares of the second-order derivatives of the mesh vertices.

[0040] By solving the above optimization problem, a mesh surface that is closely aligned with the Gaussian layer data and has good smoothness can be obtained.

[0041] S502, final mesh extraction: After the Gaussian layer is refined and optimized, the final triangular mesh is extracted from the Gaussian distribution using methods such as Poisson reconstruction. This mesh not only contains the information of the initial mesh, but also incorporates the additional details and feature information provided by the Gaussian layer. Each triangle in the mesh corresponds to a local area in the Gaussian layer, and the shape and size of the triangle are affected by the Gaussian distribution of the area. In this way, a mesh that is both accurate and efficient can be generated to represent the true geometric features of the scene.

[0042] S6. Model Lightweight In order to reduce the storage requirements of the model and speed up the rendering speed, the model is lightweighted by pruning redundant Gaussian points and knowledge distillation. Pruning redundant Gaussian points can remove points or areas that contribute less to the scene representation; knowledge distillation can compress complex models into simpler models while maintaining high accuracy and efficiency.

[0043] The present invention also provides a 3D reconstruction and mesh extraction system for underground engineering based on 3D Gaussian splashing, including a mesh extraction module, a model lightweight processing module and a web-side real-time rasterization rendering module. The mesh extraction module is used to extract a high-quality mesh model from a 3D Gaussian representation; the model lightweight processing module reduces the user's model loading time and storage overhead by performing lightweight processing on the Gaussian model, thereby improving the user experience; the web-side real-time rasterization rendering module is used to implement a 3D Gaussian splashing rasterization renderer on the web-side, supporting pure mesh representation and mesh + 3D Gaussian mixed representation.

[0044] Adaptive Gaussian layer: Enhance the detail capture capability by adding a Gaussian distribution layer to the existing mesh, so that the reconstructed model can better capture the true geometric characteristics of the object surface. Feature-aware refinement: Strengthen the Gaussian distribution density around key features to effectively enhance the expression of details. Efficient fusion and reconstruction: Provides an efficient method to fuse multiple Gaussian distributions and extract high-quality triangular meshes from them. Adaptive Gaussian layer construction method: Adaptively adjust the width of the Gaussian distribution according to surface features to ensure that sufficient details are provided in required areas. Feature-aware refinement technology: Strengthen the Gaussian distribution density around key features to effectively enhance the expression of details. Efficient mesh extraction algorithm: A method for extracting high-quality triangular meshes from the fused Gaussian distribution.

[0045] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation modes, which are merely illustrative rather than restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which are within the protection of the present invention.

Claims

1. A method for 3D reconstruction and mesh extraction of underground engineering based on 3D Gaussian splashing, characterized in that: The following steps are involved: S1. Obtain static scene images: Use photographic equipment or drones to take multiple static scene images from different angles and heights of the underground project; S2, generating sparse point cloud: processing the image set collected in step S1; S3, initialization generation of 3D Gaussian: taking the sparse point cloud obtained in step S2 as input, dividing the point cloud into multiple regions, and generating one or more 3D Gaussian models in each region based on the world space position mean and distribution characteristics of the point cloud; S4, optimize 3D Gaussian parameters: involves iteratively adjusting the center coordinates, covariance matrix and opacity parameters of the Gaussian model to make the model better fit the original point cloud data; S5, grid extraction; S6. Model lightweighting: To reduce the storage requirements of the model and speed up the rendering, the model is lightweighted.

2. The method for underground engineering 3D reconstruction and mesh extraction based on 3D Gaussian splashing according to claim 1 is characterized in that: In step S1, the static scene image covers key areas of the project including entrances, passages, intersections, and important equipment to ensure that comprehensive information can be obtained in subsequent processing.

3. The underground engineering 3D reconstruction and mesh extraction method based on 3D Gaussian splashing according to claim 1 is characterized in that: In the step S2, the collected image set is processed using the SfM algorithm; The SfM algorithm analyzes feature points in an image and tracks the position changes of the feature points in different images, thereby inferring the motion trajectory of the camera and the three-dimensional structure of the scene. The SfM algorithm generates a sparse point cloud model, which contains the three-dimensional coordinate information of key points in the scene.

4. The underground engineering 3D reconstruction and mesh extraction method based on 3D Gaussian splashing according to claim 1 is characterized in that: In the step S3, the point cloud is divided into a plurality of regions by K-means clustering or a grid-based partitioning method; The 3D Gaussian model has an ellipsoidal volume appearance, and the center coordinates, shape and opacity parameters controlled by the 3D covariance matrix are all calculated based on the point cloud data.

5. The method for underground engineering 3D reconstruction and mesh extraction based on 3D Gaussian splashing according to claim 1, characterized in that: In step S4, it involves iteratively adjusting the center coordinates, covariance matrix and opacity parameters of the Gaussian model so that the model better fits the original point cloud data; During the optimization process, in order to ensure the convergence speed and optimization effect, the optimization algorithm of gradient descent, Newton method or conjugate gradient is used.

6. The method for underground engineering 3D reconstruction and mesh extraction based on 3D Gaussian splashing according to claim 1, characterized in that: In step S5, the method for mesh extraction comprises the following steps: S501, constructing a Gaussian layer: constructing a Gaussian distribution around each vertex of the initial mesh according to a Gaussian function; For each vertex v i , define a Gaussian function: G(x,μ i ,S i ), Where x is any point in space, μ i For vertex v i The position vector, ∑ i is a 3x3 covariance matrix, the Gaussian function defines the width and direction of the Gaussian distribution, and the covariance matrix is ​​dynamically adjusted according to the position of the vertex and the required level of detail; S502, adaptive thickness adjustment: using a feature detection algorithm to identify features of edges and corners in the scene, and automatically adjusting the thickness of the Gaussian layer according to the detection results of the features; When a region is detected as an edge or corner, the thickness of the Gaussian layer in the region is reduced to better capture these features; when a region is relatively flat, the thickness of the Gaussian layer in the region is increased to reduce the computational cost; S503, final mesh extraction: After Gaussian layer refinement and optimization, the final triangular mesh is extracted from the Gaussian distribution using the Poisson reconstruction method.

7. The method for underground engineering 3D reconstruction and mesh extraction based on 3D Gaussian splashing according to claim 6, characterized in that: In step S502, a feature detection algorithm is used, including a Canny edge detector and a Shi-Tomasi corner detector. The mathematical foundations of the feature detection algorithms are briefly described as follows: Canny edge detector: Based on the smoothing process of Gaussian filtering, it determines the edge by calculating the gradient amplitude and direction, using non-maximum suppression and double threshold detection. The core lies in gradient calculation, using the Sobel operator, and the gradient calculation formula is: Among them, G x and G y are the gradients in the horizontal and vertical directions respectively; Shi-Tomasi corner detector: By calculating the autocorrelation matrix of each pixel in the image and finding points with two larger eigenvalues ​​as corners, the autocorrelation matrix M can be expressed as: Among them, I x and I y are the gradients of the image in the x and y directions respectively; After feature detection, the thickness of the Gaussian layer is adjusted according to the detection results; assuming that the detected feature intensity is F(x, y), the adjustment of the Gaussian layer thickness T(x, y) is expressed as: T(x,y)=max(T min ,T base -α·F(x,y)) Among them, T min is the minimum value of Gaussian layer thickness, T base is the base thickness, α is the adjustment coefficient, which is adjusted according to the specific application scenario; After constructing the Gaussian layer, the Poisson reconstruction algorithm is used for refinement; the core of the Poisson reconstruction algorithm is to solve a linear system, and its general form is expressed as: Lu=f Among them, L is the system matrix, which is related to the Laplacian operator of the grid; u is the grid vertex position to be solved; f is the vector related to the input point cloud data; In order to further improve the mesh quality, a regularization term is added and the objective function is expressed as: Among them, the first item is the data fitting item, which ensures that the grid is aligned with the Gaussian layer data; the second item is the regularization item, R(u) is the regularization function, which is used to smooth the grid surface and reduce noise; λ is the regularization coefficient.

8. The method for underground engineering 3D reconstruction and mesh extraction based on 3D Gaussian splashing according to claim 1, characterized in that: In step S6, in order to reduce the storage requirement of the model and speed up the rendering speed, the model is lightweighted by pruning redundant Gaussian points and performing knowledge distillation.

9. A 3D reconstruction and mesh extraction system for underground engineering based on 3D Gaussian splashing, characterized in that: It includes mesh extraction module, model lightweight processing module and web-side real-time rasterization rendering module; The mesh extraction module is used to extract a high-quality mesh model from the 3D Gaussian representation; The model lightweight processing module performs lightweight processing on the Gaussian model, thereby reducing the user's model loading time and storage overhead and improving the user experience; The web-side real-time rasterization rendering module is used to implement a 3D Gaussian splash rasterization renderer on the web-side, supporting pure grid representation and grid + 3D Gaussian mixed representation.

Images