A direct meshing method based on Gaussian surface element for incremental 3D scene reconstruction

Abstract

The application discloses a kind of based on Gaussian cell direct gridding incremental three-dimensional scene reconstruction method.Gaussian cell set is constructed to the initial first, including the following steps: RGB-D image data and corresponding camera pose of three-dimensional scene are obtained;Then, after optimization and screening to current Gaussian cell set, obtain geometric Gaussian cell set, and then generate local triangular mesh;Local triangular mesh and the latest global mesh are fused to obtain updated global mesh, and the online incremental reconstruction of three-dimensional scene is realized.The application can directly generate and update high-precision watertight mesh during observation while ensuring rendering quality, thereby effectively solving the problem of offline construction of mesh caused by dependence on implicit field or voxel grid as intermediate representation in existing three-dimensional reconstruction method.

Description

Technical Field

[0001] This invention belongs to the field of computer vision, and specifically relates to an incremental 3D scene reconstruction method based on Gaussian surface direct meshing. Background Technology

[0002] With the rapid development of technologies such as intelligent robots, autonomous driving, and augmented reality, the ability to acquire consistent local and global 3D geometric structures online during scene observation has become an urgent need. Existing 3D reconstruction methods mainly include neural reconstruction methods based on implicit fields and voxel fusion methods based on TSDF.

[0003] While implicit representation methods based on Neural Radiation Fields (NeRF) have advantages in handling complex topologies and generating high-fidelity reconstructions, they typically require preprocessing of the entire scene, resulting in enormous computational costs and making real-time incremental reconstruction difficult. On the other hand, traditional online reconstruction methods (such as KinectFusion) rely on voxel fusion mechanisms based on Truncated Symbolic Distance Fields (TSDF), which are limited by fixed voxel resolution and significant memory consumption, making it difficult to maintain sufficient geometric detail in large-scale scenes.

[0004] In recent years, 3D Gaussian Splatting (3DGS), as an explicit and differentiable 3D representation method, has demonstrated excellent performance in real-time novel view synthesis tasks. However, existing 3DGS-based reconstruction methods typically require offline optimization of the Gaussian surface, followed by conversion of the optimized Gaussian surface into an intermediate implicit field (such as SDF or density field), and finally extraction of the mesh from the implicit field using the Marching Cubes algorithm. Such methods rely on the intermediate implicit field for mesh extraction, making it impossible to directly obtain and update the scene mesh during observation, thus failing to meet the requirements for immediacy and online feedback in tasks such as robotics.

[0005] Therefore, there is an urgent need for a reconstruction method that does not require the construction of implicit fields or rely on voxel meshes. Summary of the Invention

[0006] To address the problems existing in the background technology, this invention provides an incremental 3D scene reconstruction method based on direct Gaussian grid generation. This invention can incrementally recover high-quality meshes online from continuous RGB-D image data, while simultaneously achieving high-fidelity rendering. This invention can directly generate explicit meshes based on 3D Gaussian representation and supports incremental updates during observation, thereby meeting the needs of online incremental 3D reconstruction.

[0007] The technical solution adopted in this invention is: I. An incremental 3D scene reconstruction method based on Gaussian surface direct meshing S1: Acquire RGB-D image data of the 3D scene and the corresponding camera pose; S2: Construct an initial Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose; then optimize and filter the current Gaussian surface set to obtain a geometric Gaussian surface set, and then generate a local triangular mesh. S3: Merge the current local triangular mesh with the latest global mesh to obtain an updated global mesh; S4: Repeat S1-S3 to continuously acquire RGB-D image data of the 3D scene and the corresponding camera pose, generate local triangular meshes, and then update the global mesh to achieve online incremental reconstruction of the 3D scene.

[0008] In step S2, constructing the initial Gaussian element set based on the currently acquired RGB-D image data and the corresponding camera pose includes: A directed point cloud is generated by backprojecting the depth map from the currently acquired RGB-D image and the camera pose, and then combined with the RGB image from the RGB-D image to construct an initial Gaussian surface set.

[0009] In step S2, after optimizing and filtering the current Gaussian surface element set, a geometric Gaussian surface element set is obtained, including: Gaussian rendering is performed based on the current Gaussian polygon set to obtain a rendered image. By minimizing the loss function between the rendered image and the current RGB-D image data, the parameters of each Gaussian polygon in the current Gaussian polygon set are jointly optimized so that the optimized Gaussian polygons can closely fit the local geometric surface of the scene represented by the Gaussian polygon set and maintain geometric consistency, thereby obtaining the optimized Gaussian polygon set.

[0010] The loss function includes color loss, depth loss, plane-based traction loss, normal vector consistency loss, and sparsity loss.

[0011] In step S2, after optimizing and filtering the current Gaussian surface element set, a geometric Gaussian surface element set is obtained, including: Based on the currently acquired RGB-D image data and the corresponding camera pose, a geometric Gaussian surface set is selected from the optimized Gaussian surface set.

[0012] The step of selecting a geometric Gaussian surface set from the optimized Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose specifically includes: First, the optimized 3D Gaussian surface set is projected onto the camera view using the camera pose of each frame, and then the corresponding predicted depth map is obtained through Gaussian surface rendering. For Gaussian surfaces that satisfy the opacity constraint, the difference between the predicted depth value along the viewing direction and the actual depth value of the corresponding pixel position in the current frame depth map is calculated. Only Gaussian surfaces with a depth difference less than a preset depth threshold are retained, thereby obtaining the geometric Gaussian surface set.

[0013] In step S2, generating a local triangular mesh includes: After performing a direct triangulation mesh generation operation on the geometric Gaussian surface set, a local triangular mesh is obtained. The direct triangulation mesh generation operation includes projection on the tangent space and topological connectivity.

[0014] II. An incremental 3D scene reconstruction system based on Gaussian surface element direct meshing The data acquisition unit acquires RGB-D image data of the 3D scene and the corresponding camera pose; The Gaussian surface set construction unit is used to construct an initial Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose. The optimization unit is used to optimize the initial Gaussian surface set to obtain the optimized Gaussian surface set. The filtering unit is used to filter out the geometric Gaussian surface set from the optimized Gaussian surface set; Mesh generation unit, used to obtain local triangular meshes after performing direct triangulation mesh generation operation on geometric Gaussian surface set; The mesh update unit is used to merge the current local triangular mesh with the latest global mesh to obtain an updated global mesh.

[0015] III. A computer device The computer device includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the incremental 3D scene reconstruction method based on Gaussian surface direct meshing.

[0016] IV. A computer-readable storage medium The storage medium stores a computer program, which, when executed by a processor, implements the steps of the incremental 3D scene reconstruction method based on Gaussian surface direct meshing.

[0017] The beneficial effects of this invention are: This invention eliminates the cumbersome process of converting optimized Gaussian fields into intermediate implicit fields before mesh extraction, which is common in existing technologies. Instead, it directly triangulates the dense geometric Gaussian representation. This invention can extract and update the mesh from the Gaussian set in real time, without waiting for the entire scene scan to complete, achieving efficient online incremental reconstruction and overcoming the bottleneck of offline processing. This end-to-end direct meshing strategy significantly reduces computational and memory overhead.

[0018] In the 3D Gaussian optimization part of this invention, in order to dynamically align the Gaussian center to the local geometric plane, a plane-based traction constraint is introduced, which enables the generated 3D mesh to closely fit the surface of the real object.

[0019] This invention proposes a dense geometric Gaussian representation, which uses Gaussian elements as unified geometric primitives for rendering and mesh generation. Under the same representation framework, it simultaneously takes into account appearance expression and geometric modeling, and achieves high-fidelity rendering and high-quality geometric reconstruction. Attached Figure Description

[0020] Figure 1 This is a flowchart of the method proposed in this invention.

[0021] Figure 2 This is a framework diagram of the method proposed in this invention.

[0022] Figure 3 The images are visualization results of the model after reconstruction using the method proposed in this invention; where (a) is an RGB image of the three-dimensional Gaussian surface model rendered from a new perspective, and (b) is a schematic diagram of the mesh from the same perspective as (a). Detailed Implementation

[0023] The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0024] like Figure 1 and Figure 2 As shown, the incremental 3D scene reconstruction method based on Gaussian surface direct meshing proposed in this invention includes the following steps: S1: Acquire RGB-D image data of the 3D scene and the corresponding camera pose through camera acquisition; In one feasible implementation, two general indoor datasets, Replica and ScanNet++, are selected. Both datasets contain color image datasets acquired by cameras, depth image datasets acquired by depth cameras, and known camera poses.

[0025] RGB-D image data can be a single frame or multiple consecutive frames, and the camera pose can also be a single frame or multiple consecutive frames.

[0026] S2: Construct an initial dense Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose; then optimize and filter the current Gaussian surface set to obtain a geometric Gaussian surface set, and then generate a local triangular mesh. In one feasible implementation, constructing an initial dense Gaussian element set based on the currently acquired RGB-D image data and the corresponding camera pose includes: A directed point cloud is generated by back-projecting the depth map and camera pose from the currently acquired RGB-D image, and then combined with the RGB image in the RGB-D image to construct an initial dense Gaussian element set.

[0027] Specifically: The initial point cloud position of the current frame is obtained by back-projecting each frame of depth image captured by the depth camera into space according to the camera's intrinsic parameters. The normal vector corresponding to each point in the point cloud The point cloud position is estimated by the cross product of adjacent points. If the currently acquired RGB-D image data contains multiple frames, each frame is processed to obtain the initial point cloud position corresponding to each frame. Then, a weighted fusion strategy is used to fuse the point clouds from each frame, thereby constructing a globally consistent directed point cloud, thus obtaining the final directed point cloud. Next, combining the RGB images from the RGB-D image, the directed point cloud is initialized as a 3D Gaussian distribution, with each Gaussian surface containing the mean. Rotation matrix, RGB color values Scale vector Opacity Among them, Gaussian mean Taken from point cloud location The rotation matrix is ​​derived from the point cloud normal vectors. Parameterization; RGB color values Originating from the color information of corresponding points in a directed point cloud; the color information originates from the acquired RGB image stream; scale vector The first two components and The third component is determined based on the geometric distance between the current point cloud and its nearest neighbor point clouds. Setting it to approach 0, specifically by setting the scale vector... The third component Set to a very small value (e.g., 0.000001); Opacity The initial value is set to 0.99.

[0028] In one feasible implementation, after optimizing and filtering the current Gaussian surface element set, a geometric Gaussian surface element set is obtained, including: Gaussian rendering is performed based on the current Gaussian polygon set to obtain a rendered image. By minimizing the loss function between the rendered image and the current RGB-D image data, the parameters of each Gaussian polygon in the current Gaussian polygon set are jointly optimized so that the optimized Gaussian polygons can closely fit the local geometric surface of the scene represented by the Gaussian polygon set and maintain geometric consistency, thereby obtaining the optimized Gaussian polygon set.

[0029] The loss function L includes color loss, depth loss, plane-based traction loss, normal vector consistency loss, and sparsity loss. The specific formula is as follows:

[0030]

[0031]

[0032] =

[0033]

[0034]

[0035]

[0036] in, For color loss; This is a deep loss; For planar traction loss; For normal vector consistency loss; This is a sparsity loss; , , , , These are the first through fifth loss weights; N represents the total number of pixels in the rendered image. This indicates the absolute value operation; For the first The actual observed color value at each pixel location; The predicted color value is obtained by rendering from the current set of Gaussian elements; For the first The actual observed depth value at each pixel location; The predicted depth value is obtained by rendering the current set of Gaussian pixels. The pixel depth is obtained by calculating the intersection of the view and the opaque Gaussian ellipsoid in front. r is the unit view direction vector in the world coordinate system. and These are the position vector and normal vector of the Gaussian surface element that intersects with the unit line-of-sight vector r, respectively; This is the transformation matrix from the world coordinate system to the current camera coordinate system. This indicates retrieving the depth value of the 3D point; T indicates the transpose operation. Indicates the center position of the Gaussian element; This indicates the position of a point in the corresponding directed point cloud; Represents the normal vector of a point; This represents the set of correspondences between Gaussian surface elements and directed point clouds; Indicates the first The normal vector of a Gaussian surface element; This represents the opacity value of the Gaussian surface element. express Norm; This indicates the number of Gaussian elements.

[0037] In the loss function of this invention, Ensure that the appearance properties of Gaussian elements accurately reproduce the visual features of the scene. Used to constrain the depth map rendered by Gaussian polygons and the depth map acquired by the sensor. This involves pulling Gaussian elements onto a local geometric plane, providing smooth, continuous Gaussian elements for subsequent direct meshing extraction. This is used to guide the Gaussian ellipsoid to remain consistent with the tangent vector of the object surface throughout the optimization process, providing an accurate normal field for subsequent direct meshing operations, and ensuring that the generated triangular mesh can accurately fit complex geometric boundaries. By penalizing Gaussian elements with intermediate opacity, the system automatically eliminates redundant elements that contribute very little to the rendering (i.e., reduces their opacity to 0), or strengthens key elements into solid surfaces (i.e., their opacity approaches 1). During incremental reconstruction, this constraint effectively suppresses floating noise in space, resulting in a clearer and more compact surface representation, greatly reducing the topological complexity during meshing.

[0038] This invention optimizes the parameters of Gaussian facets by minimizing the loss function between the rendered image and the current RGB-D image in terms of color and depth differences, and by combining this with a traction constraint based on local geometry (used to pull the center of the Gaussian facet to the local geometric plane). In other words, this invention updates the geometric and appearance properties of Gaussian facets under multi-view observation constraints, jointly optimizing attributes such as the center position, normal vector, scale, and opacity of the Gaussian facets.

[0039] In one feasible implementation, after optimizing and filtering the current Gaussian surface element set, a geometric Gaussian surface element set is obtained, including: Based on the currently acquired RGB-D image data and the corresponding camera pose, a geometric Gaussian surface set is selected from the optimized Gaussian surface set.

[0040] Based on the currently acquired RGB-D image data and the corresponding camera pose, a geometric Gaussian surface set is selected from the optimized Gaussian surface set, specifically including: First, the optimized 3D Gaussian facet set is projected onto the camera view using the camera pose of each frame, and then the corresponding predicted depth map is obtained through Gaussian facet rendering. For Gaussian facets that satisfy the opacity constraint (whose opacity is higher than the preset opacity threshold), the difference between the predicted depth value along the viewing direction and the actual depth value of the corresponding pixel position in the current frame depth map is calculated. Only Gaussian facets with a depth difference less than the preset depth threshold are retained. These Gaussian facets satisfy both the opacity constraint and the depth consistency constraint. Facets that contribute little or no to geometry are eliminated, thus obtaining the geometric Gaussian facet set.

[0041] If the currently acquired RGB-D image data contains several frames, then the RGB-D image of each frame and the corresponding camera pose are traversed frame by frame to obtain the geometric Gaussian surface set corresponding to each frame. Finally, the geometric Gaussian surface sets corresponding to each frame are combined and processed to obtain the final geometric Gaussian surface set.

[0042] In one feasible implementation, generating a local triangular mesh includes: After performing a direct triangulation mesh generation operation on the geometric Gaussian surface set, a local triangular mesh is obtained. The direct triangulation mesh generation operation includes projection on the tangent space and topological connectivity.

[0043] Specifically: For each geometric Gaussian surface element in the geometric Gaussian surface element set, with its center position... and normal vector Based on this, a local tangent plane is constructed using the normal vector of the central Gaussian, providing a unified reference space for subsequent geometric analysis and topology construction.

[0044] Adaptive neighborhood search radius of the geometric Gaussian surface element The process involves retrieving adjacent Gaussian surfaces, projecting them (i.e., Gaussian surfaces within the neighborhood) onto a local tangent plane, and then combining their relative spatial position and normal relationship within the tangent space to determine the consistency of the normals of adjacent Gaussian surfaces. This process filters out a set of candidate Gaussian surfaces that are locally geometrically consistent with the central Gaussian surface.

[0045] Adaptive Domain Search Radius The calculation is based on the following formula:

[0046] in, Let the three-dimensional spatial coordinates of the central Gaussian be denoted as and the three-dimensional spatial coordinates of its k nearest neighbors be denoted as . , This is a scaling factor.

[0047] In one feasible implementation, normal consistency specifically means: The adjacent Gaussian surfaces are projected onto the local tangent plane, and the cosine of the angle between the normals of the adjacent Gaussian surfaces and the central Gaussian surface is calculated. Gaussian surfaces with a cosine of the angle between the normals greater than a certain threshold (such as 0.95) are selected, thus obtaining a set of candidate Gaussian surfaces that are locally geometrically consistent with the central Gaussian surface.

[0048] Next, based on the candidate Gaussian surface set, the three-dimensional mean position of the central Gaussian surface is updated by weighted fusion to enhance the stability and continuity of the local surface estimation.

[0049] Then, in the updated central Gaussian tangent plane, the neighboring Gaussian elements are reprojected, and local topological connections are constructed based on their angular order and relative distance relationship in the tangent space. Under the conditions of satisfying angular constraints, distance constraints and visibility constraints, several local triangular pieces are generated, thereby completing the direct triangulation operation of the current geometric Gaussian elements based on Gaussian elements.

[0050] The remaining geometric Gaussian elements in the geometric Gaussian element set are traversed and processed to obtain several corresponding local triangular patches. Finally, the local triangular patches generated for each central Gaussian element are collected to form a complete local triangular mesh, which is the output result of the direct triangulation mesh generation step.

[0051] S3: Merge the current local triangular mesh with the latest global mesh to obtain an updated global mesh; In one feasible implementation, fusing the current local triangular mesh with the latest global mesh specifically involves updating the global mesh topology using a remeshing operation, as follows: For mesh regions where geometric inconsistencies arise due to persistent changes in the properties of some Gaussian elements caused by repeated observations, a remeshing operation is triggered on the affected global mesh region. Remeshing refers to adaptively splitting and shrinking excessively long or short edges based on the local average edge length, for example, according to the local average edge length. Split excessively long edges in the global mesh to ensure their length does not exceed 1.5. And for those shorter than 0.5 The edges are contracted to simplify the topology. While maintaining visibility and geometric consistency, edge flipping is used to optimize vertex degree distribution. Then, Laplacian smoothing is applied to adjust vertex positions, thereby maintaining overall topological stability while achieving online incremental updates and geometric consistency maintenance of the global mesh. For example, edge flipping is used to increase the degree of adjacent vertices to near the target degree of 6; and Laplacian smoothing is used to migrate vertex positions to the centroids of their neighboring vertices to improve mesh uniformity and consistency.

[0052] S4: Repeat S1-S3 to continuously acquire RGB-D image data of the 3D scene and the corresponding camera pose, generate local triangular meshes, and then update the global mesh to achieve online incremental reconstruction of the 3D scene.

[0053] The visualization result of the model reconstructed by the incremental 3D reconstruction method based on Gaussian surface element direct meshing after collecting data in an indoor scene according to the embodiments of the present invention is shown in the figure. Figure 3 As shown. Figure 3 (a) is an RGB image rendered from a new perspective of a 3D Gaussian surface model, showing a high-fidelity representation of the scene appearance. Figure 3 (b) is with Figure 3 (a) A schematic diagram of the mesh from the same perspective, showing a topologically complete and geometrically accurate watertight mesh. Figure 3 The invention verifies that the method of the present invention, while ensuring rendering quality during incremental reconstruction, is free from the limitations of traditional implicit intermediate representation and can directly extract high-quality geometric structures from explicit Gaussian polygon sets.

[0054] To further verify the reconstruction effect of this invention, experiments were conducted on the public dataset Replica, and the performance of the method of this invention was compared with that of existing mainstream reconstruction methods. As shown in Table 1, the method of this invention significantly outperforms existing mainstream methods in all geometric evaluation metrics on the Replica dataset. Specifically, the accuracy error of the method of this invention is only 1.34 cm, which is a 51.6% improvement compared to the similar MonoGS method based on Gaussian models, and achieves several times the accuracy improvement compared to the classic KinectFusion and NICE-SLAM methods. Furthermore, the accuracy of the method of this invention reaches 99.7%, proving that the direct meshing algorithm proposed in this invention can restore the true geometric structure of the scene with extremely high accuracy.

[0055] Table 1. Performance comparison of the method of this invention with existing reconstruction methods on the Replica dataset.

[0056] The present invention proposes an incremental 3D scene reconstruction method based on Gaussian surface direct meshing, which directly utilizes dense geometric Gaussian representation for explicit mesh reconstruction, eliminating the drawback of traditional methods that require the construction of implicit fields as an intermediate medium, and achieving the unity of high-fidelity rendering and accurate geometric reconstruction.

[0057] The incremental 3D scene reconstruction system based on Gaussian surface direct meshing proposed in this invention includes: The data acquisition unit acquires RGB-D image data of the 3D scene and the corresponding camera pose; The Gaussian surface set construction unit is used to construct an initial Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose. The optimization unit is used to optimize the initial Gaussian surface set to obtain the optimized Gaussian surface set. The filtering unit is used to filter out the geometric Gaussian surface set from the optimized Gaussian surface set; Mesh generation unit, used to obtain local triangular meshes after performing direct triangulation mesh generation operation on geometric Gaussian surface set; The mesh update unit is used to merge the current local triangular mesh with the latest global mesh to obtain an updated global mesh.

[0058] The present invention proposes a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of an incremental three-dimensional scene reconstruction method based on Gaussian surface direct meshing.

[0059] The present invention proposes a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of an incremental 3D scene reconstruction method based on Gaussian surface direct meshing.

Claims

1. An incremental 3D scene reconstruction method based on Gaussian surface direct meshing, characterized in that, Includes the following steps: S1: Acquire RGB-D image data of the 3D scene and the corresponding camera pose; S2: Construct an initial Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose; then optimize and filter the current Gaussian surface set to obtain a geometric Gaussian surface set, and then generate a local triangular mesh. S3: Merge the current local triangular mesh with the latest global mesh to obtain an updated global mesh; S4: Repeat S1-S3 to continuously acquire RGB-D image data of the 3D scene and the corresponding camera pose, generate local triangular meshes, and then update the global mesh to achieve online incremental reconstruction of the 3D scene.

2. The incremental 3D scene reconstruction method based on Gaussian surface direct meshing according to claim 1, characterized in that, In step S2, constructing the initial Gaussian element set based on the currently acquired RGB-D image data and the corresponding camera pose includes: A directed point cloud is generated by backprojecting the depth map from the currently acquired RGB-D image and the camera pose, and then combined with the RGB image from the RGB-D image to construct an initial Gaussian surface set.

3. The incremental 3D scene reconstruction method based on Gaussian surface direct meshing according to claim 1, characterized in that, In step S2, after optimizing and filtering the current Gaussian surface element set, a geometric Gaussian surface element set is obtained, including: Gaussian rendering is performed based on the current Gaussian polygon set to obtain a rendered image. By minimizing the loss function between the rendered image and the current RGB-D image data, the parameters of each Gaussian polygon in the current Gaussian polygon set are jointly optimized so that the optimized Gaussian polygons can closely fit the local geometric surface of the scene represented by the Gaussian polygon set and maintain geometric consistency, thereby obtaining the optimized Gaussian polygon set.

4. The incremental 3D scene reconstruction method based on Gaussian surface direct meshing according to claim 3, characterized in that, The loss function includes color loss, depth loss, plane-based traction loss, normal vector consistency loss, and sparsity loss.

5. The incremental 3D scene reconstruction method based on Gaussian surface direct meshing according to claim 1, characterized in that, In step S2, after optimizing and filtering the current Gaussian surface element set, a geometric Gaussian surface element set is obtained, including: Based on the currently acquired RGB-D image data and the corresponding camera pose, a geometric Gaussian surface set is selected from the optimized Gaussian surface set.

6. The incremental 3D scene reconstruction method based on Gaussian surface direct meshing according to claim 1, characterized in that, The step of selecting a geometric Gaussian surface set from the optimized Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose specifically includes: First, the optimized 3D Gaussian surface set is projected onto the camera view using the camera pose of each frame, and then the corresponding predicted depth map is obtained through Gaussian surface rendering. For Gaussian surfaces that satisfy the opacity constraint, the difference between the predicted depth value along the viewing direction and the actual depth value of the corresponding pixel position in the current frame depth map is calculated. Only Gaussian surfaces with a depth difference less than a preset depth threshold are retained, thereby obtaining the geometric Gaussian surface set.

7. The incremental 3D scene reconstruction method based on Gaussian surface direct meshing according to claim 1, characterized in that, In step S2, generating a local triangular mesh includes: After performing a direct triangulation mesh generation operation on the geometric Gaussian surface set, a local triangular mesh is obtained. The direct triangulation mesh generation operation includes projection on the tangent space and topological connectivity.

8. An incremental 3D scene reconstruction system based on Gaussian surface direct meshing, characterized in that, include: The data acquisition unit acquires RGB-D image data of the 3D scene and the corresponding camera pose; The Gaussian surface set construction unit is used to construct an initial Gaussian surface set based on the currently acquired RGB-D image data and the corresponding camera pose. The optimization unit is used to optimize the initial Gaussian surface set to obtain the optimized Gaussian surface set. The filtering unit is used to filter out the geometric Gaussian surface set from the optimized Gaussian surface set; Mesh generation unit, used to obtain local triangular meshes after performing direct triangulation mesh generation operation on geometric Gaussian surface set; The mesh update unit is used to merge the current local triangular mesh with the latest global mesh to obtain an updated global mesh.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the incremental three-dimensional scene reconstruction method based on Gaussian surface direct meshing as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the incremental three-dimensional scene reconstruction method based on Gaussian surface direct meshing as described in any one of claims 1 to 7.

Images