Three-dimensional reconstruction method, storage medium and program product

Abstract

The invention provides a three-dimensional reconstruction method, a storage medium and a program product, and relates to the technical field of three-dimensional reconstruction. The method comprises the following steps: acquiring grid data, and extracting geometric features in the grid data; according to feature keeping importance corresponding to different regions in the geometric features, determining constraint strength corresponding to the different regions in the geometric features; on the basis of a progressive Laplacian smoothing algorithm, optimizing the geometric features according to constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features; and obtaining three-dimensional reconstruction data corresponding to the grid data according to the optimized geometric features. The accuracy of a three-dimensional reconstruction result can be improved.

Description

Technical Field

[0001] This application relates to the field of three-dimensional reconstruction technology, and in particular to a three-dimensional reconstruction method, storage medium, and program product. Background Technology

[0002] With the rapid development of 3D scanning and computer vision technologies, 3D reconstruction methods based on multi-view stereo vision, LiDAR, or structured light are widely used in fields such as industrial design, autonomous driving, and digital twins. In the process of digital modeling vehicle exteriors, these technologies can efficiently acquire geometric information about the vehicle body surface, providing an important data foundation for virtual display and quality inspection.

[0003] In the 3D reconstruction process, related technologies employ the Laplacian smoothing operator to iteratively optimize the mesh vertex positions, achieving surface smoothness by minimizing curvature changes; anisotropic diffusion filtering is used to remove high-frequency noise while preserving edge features; and geometric noise is eliminated by simulating the surface evolution process based on the curvature flow method. However, this approach suffers from poor 3D reconstruction results. Summary of the Invention

[0004] This application provides a three-dimensional reconstruction method, storage medium, and program product to improve the accuracy of three-dimensional reconstruction results.

[0005] In a first aspect, embodiments of this application provide a three-dimensional reconstruction method, including:

[0006] Acquire grid data and extract geometric features from the grid data;

[0007] Based on the feature retention importance corresponding to different regions in the geometric feature, determine the constraint strength corresponding to different regions in the geometric feature;

[0008] Based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features;

[0009] Based on optimized geometric features, obtain the 3D reconstruction data corresponding to the mesh data.

[0010] In one possible implementation, the geometric feature includes feature line vertex regions, feature line neighborhood regions, transition regions, and flat regions with decreasing feature preservation importance; the constraint strength corresponding to different regions in the geometric feature is determined based on the feature preservation importance corresponding to different regions, including:

[0011] The constraint strength corresponding to the vertex region of the feature line is determined as the first constraint strength;

[0012] The constraint strength corresponding to the neighborhood region of the feature line is determined as the second constraint strength;

[0013] The constraint strength corresponding to the transition region is determined to be the third constraint strength;

[0014] The constraint strength corresponding to the flat region is determined to be the fourth constraint strength;

[0015] Among them, the corresponding constraint strengths of the first constraint strength, the second constraint strength, the third constraint strength and the fourth constraint strength decrease in sequence.

[0016] In one possible implementation, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features, including:

[0017] In the first resolution dimension, based on the Laplacian smoothing algorithm, according to the constraint strength corresponding to different regions in the geometric features, large-scale noise and global distortion in the geometric features are removed to obtain the first optimized feature. The first resolution dimension is the dimension that can display the vertex of the feature line.

[0018] In the second resolution dimension, based on the Laplacian smoothing algorithm, the first optimized feature is restored in detail according to the constraint strength corresponding to different regions in the geometric features to obtain the second optimized feature. The second resolution dimension is the dimension that can display all geometric features.

[0019] In the third resolution dimension, based on the Laplacian smoothing algorithm, the details of the second optimized feature are restored according to the constraint strength of different regions in the geometric features, and the position of the mesh vertices in the second optimized feature is optimized to obtain the optimized geometric features. The resolution corresponding to the third resolution dimension is smaller than the resolution corresponding to the second resolution dimension.

[0020] In one possible implementation, the progressive Laplace smoothing algorithm includes the following mechanism:

[0021] The dihedral angle restoration mechanism forces the faces on both sides of the feature edge to retain the original dihedral angle;

[0022] A curvature continuity optimization mechanism maintains curvature continuity along the direction of the feature line;

[0023] An adaptive iterative stopping mechanism stops smoothing when the following stopping conditions are detected: the change in position of each vertex after smoothing is less than the position change threshold, the change in the average dihedral angle of the feature edges tends to stabilize, and the system residual norm is less than the preset tolerance.

[0024] In one possible implementation, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features, including:

[0025] The grid data is divided into multiple sub-regions, and the Schwarz alternation method is used to coordinate the boundary consistency of multiple sub-regions.

[0026] Within multiple sub-regions, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features, resulting in optimized geometric features.

[0027] In one possible implementation, it also includes:

[0028] During the optimization of geometric features, when the positional change of the geometric feature is detected to be greater than the preset positional change threshold, the Laplacian matrix terms in the progressive Laplacian smoothing algorithm are updated, and a hot-start strategy based on the conjugate gradient method is used to improve the computational efficiency of the progressive Laplacian smoothing algorithm.

[0029] In one possible implementation, the geometric features include corner points or endpoints. The constraint strength corresponding to different regions within the geometric feature is determined based on the importance of features corresponding to different regions, including:

[0030] The protection range is determined by using the corner point or endpoint as the center and the preset protection radius length;

[0031] Within the protection zone, the constraint strength increases exponentially as the distance to the corner or endpoint decreases.

[0032] In one possible implementation, the 3D reconstruction data corresponding to the mesh data is obtained based on optimized geometric features, including:

[0033] Based on the feature lines contained in the optimized geometric features, retopology sampling points are generated using the feature distance function;

[0034] The feature lines contained in the optimized geometric features are used as constraint edges. Based on the Delaunay triangulation algorithm, the first topological mesh is generated according to the retopological sampling points.

[0035] An improved Ruppert algorithm is used to optimize the first-level topological mesh to obtain the second-level topological mesh;

[0036] Based on the preset mesh quality, the second topological mesh is iteratively optimized to obtain the 3D reconstruction data corresponding to the acquired mesh data.

[0037] In one possible implementation, extracting geometric features from the grid data includes:

[0038] Based on an improved quadratic error metric algorithm, the curvature corresponding to the grid vertices in the grid data is calculated;

[0039] Identify potential feature regions based on the curvature corresponding to the grid vertices;

[0040] A dihedral angle calculation algorithm based on a half-side data structure is adopted to filter feature edges from potential feature regions according to a preset angle threshold.

[0041] Gaussian curvature distribution map is used to identify concave and convex regions. Combined with prior knowledge of the shape of the 3D reconstructed object, false features in the feature edges are eliminated to obtain geometric feature data.

[0042] Using prior engineering knowledge, semantic features and prior confidence maps corresponding to geometric feature data are obtained based on geometric feature data.

[0043] Geometric features are obtained by multimodal fusion of geometric feature data, semantic features, and prior confidence maps.

[0044] Secondly, embodiments of this application provide a three-dimensional reconstruction apparatus, comprising:

[0045] The extraction module is used to acquire grid data and extract geometric features from the grid data;

[0046] The determination module is used to determine the constraint strength corresponding to different regions of a geometric feature based on the feature retention importance corresponding to different regions in the geometric feature;

[0047] The feature optimization module is used to optimize geometric features based on the progressive Laplacian smoothing algorithm, according to the constraint strength corresponding to different regions in the geometric features, to obtain optimized geometric features;

[0048] The 3D reconstruction module is used to obtain 3D reconstruction data corresponding to the mesh data based on optimized geometric features.

[0049] Thirdly, embodiments of this application provide a three-dimensional reconstruction device, including: a memory and a processor;

[0050] The memory stores instructions that the computer executes;

[0051] The processor executes computer execution instructions stored in memory, causing the processor to perform the methods described in the various possible implementations of the first aspect above.

[0052] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the methods described in the various possible implementations of the first aspect above.

[0053] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed, implements the methods described in the various possible implementations of the first aspect above.

[0054] The 3D reconstruction method, storage medium, and program product provided in this application acquire mesh data and extract geometric features from the mesh data; determine the constraint strength corresponding to different regions of the geometric features based on the feature preservation importance; optimize the geometric features based on the constraint strength corresponding to different regions of the geometric features to obtain optimized geometric features; and acquire the 3D reconstruction data corresponding to the mesh data based on the optimized geometric features. Based on the hierarchical constraint mechanism, excessive smoothing of feature lines is avoided. The progressive Laplacian smoothing algorithm optimizes the geometric features at multiple scales based on feature strength, balancing global noise removal and local detail preservation, thus improving the accuracy of the 3D reconstruction results. Attached Figure Description

[0055] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0056] Figure 1 A schematic flowchart illustrating the three-dimensional reconstruction method provided in this application embodiment;

[0057] Figure 2 This is a schematic diagram of the three-dimensional reconstruction device provided in the embodiments of this application;

[0058] Figure 3 This is a schematic diagram of the structure of the three-dimensional reconstruction device provided in the embodiments of this application.

[0059] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0060] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0061] In practical engineering applications of 3D reconstruction technology, the following technical problems are encountered:

[0062] Geometric noise issues: Due to the influence of sensor accuracy, ambient lighting conditions, and surface material characteristics (such as highly reflective paint), the reconstructed triangular mesh models generally exhibit geometric noise, manifested as irregular surface depressions, bulges, and distortions. The amplitude of this type of noise can typically reach 0.1%–1% of the diagonal length of the mesh bounding box, severely affecting the visual quality and engineering usability of the model.

[0063] • Difficulty in preserving features: Key engineering feature lines in vehicle design (such as skirt lines, waistlines, shoulder lines, and rooflines) are prone to losing their geometric sharpness or even breaking during reconstruction due to noise interference and uneven sampling density. These feature lines play an important role in vehicle aesthetic design and aerodynamics, and their distortion can seriously affect subsequent engineering analysis, computer-aided design (CAD) model reconstruction, and manufacturing process evaluation.

[0064] Topological quality defects: Automatically reconstructed meshes often contain a large number of elongated triangles, non-manifold edges, isolated vertices, and other topological errors. Triangle quality is usually measured by aspect ratio, and the average aspect ratio of the reconstructed mesh often exceeds 5:1, far below the 2:1 standard required for engineering applications, making the mesh difficult to use directly for engineering applications such as CAD modeling, finite element analysis, and computational fluid dynamics.

[0065] In the 3D modeling process, related technologies employ Laplacian smoothing operators to iteratively optimize the positions of mesh vertices, achieving surface smoothness by minimizing curvature changes; anisotropic diffusion filtering is used to remove high-frequency noise while preserving edge features; and curvature flow methods are used to eliminate geometric noise by simulating the surface evolution process. However, this approach suffers from the following technical drawbacks: traditional Laplacian smoothing is isotropic, homogenizing all geometric features and causing important engineering feature lines to become over-smoothed and lose their sharpness; anisotropic diffusion requires manual adjustment of multiple parameters, resulting in poor adaptability to the surface characteristics of different vehicle models; the curvature flow method requires solving complex partial differential equations, leading to high computational costs and poor real-time performance when processing large-scale meshes; these methods are prone to getting trapped in local optima and cannot balance noise removal and feature preservation globally. In summary, these technologies suffer from poor 3D reconstruction results.

[0066] To address the aforementioned technical problems, embodiments of this application provide a three-dimensional reconstruction method, including the following technical concepts:

[0067] First, a hierarchical constraint weight allocation mechanism is designed to dynamically adjust the smoothing intensity based on feature importance. A multi-scale progressive smoothing strategy is proposed, removing large-scale noise at the coarse-grained level and optimizing local geometry at the fine-grained level. A feature-preserving energy function is established, transforming feature preservation into a constraint optimization problem. Hierarchical constraints prevent feature lines from being over-smoothed; multi-scale processing balances global noise removal and local detail preservation; the energy function integrates feature preservation into the mathematical optimization framework to achieve adaptive balance.

[0068] Second, a feature-guided density field generation algorithm maintains high sampling density in the feature region; it implements a constrained triangulation process to ensure accurate recovery of feature edges after retopology; and it constructs a mesh quality evaluation and optimization loop to continuously improve topology quality through local operations. The density field ensures sufficient geometric sampling in the feature region; constrained triangulation preserves the feature edge structure; and the quality optimization loop generates a high-quality mesh suitable for engineering applications.

[0069] Third, a joint optimization model combining geometric features and engineering priors is established, integrating geometric features such as discrete curvature and dihedral angles with prior knowledge of feature line positions within the vehicle bounding box; a feature path optimization method based on the graph cut algorithm is designed to find the optimal continuous path within the feature candidate region; and a feature confidence evaluation system is constructed to improve the reliability of feature extraction through multi-index fusion. By fusing geometric data and engineering priors, noise and true features are accurately distinguished; the graph cut algorithm ensures the continuity and integrity of feature lines; and confidence evaluation reduces misidentification and improves robustness.

[0070] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0071] Figure 1 This is a schematic flowchart illustrating the three-dimensional reconstruction method provided in an embodiment of this application. Figure 1 As shown in the figure, this application provides a three-dimensional reconstruction method, including:

[0072] S101. Obtain grid data and extract geometric features from the grid data.

[0073] Specifically, geometric features include vertex regions, feature line neighborhoods, transition regions, and flat regions in the grid data. In one implementation, a quadratic error metric algorithm (QEM) is used to extract geometric features from the grid data.

[0074] Optionally, the QEM algorithm can be improved by calculating the sum of the basic quadratic form matrices of all triangular faces within a ring of its surrounding neighborhood for each vertex in the grid data. This matrix encapsulates the geometric information of the local region of that vertex. Using the basic quadratic form matrix of the vertex, a best-fit quadratic surface can be defined for its local neighborhood. From this quadratic surface function, the principal curvature, Gaussian curvature, and mean curvature can be analytically estimated by analyzing its Hessian matrix (second derivative matrix) at the vertex. Based on the principal curvature, Gaussian curvature, and mean curvature, geometric features among geometric features can be identified.

[0075] In one implementation, the improved QEM algorithm is integrated with prior engineering knowledge to enhance the accuracy of geometric feature extraction.

[0076] S102. Determine the constraint strength corresponding to different regions in the geometric feature based on the feature retention importance of different regions in the geometric feature.

[0077] Specifically, for core regions where features like mesh feature lines, sharp corners, and high-curvature abrupt changes are of high importance, higher constraint strength is applied. This strictly limits vertex merging, positional shifts, and mesh deformation within these regions, maximizing the preservation of core geometric details and feature recognizability. For non-core regions where features like flat transition areas and low-curvature smooth areas are of lower importance, the constraint strength is appropriately reduced, allowing for greater mesh simplification operations in these regions, thus improving mesh reduction efficiency.

[0078] S103. Based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features.

[0079] Specifically, leveraging the multi-round iteration and step-by-step convergence characteristics of the progressive Laplacian smoothing algorithm, the differentiated constraint strength corresponding to different regions of the geometric features is used as the optimization basis. Matching constraints are applied to each region during the algorithm's successive smoothing iterations. For core regions where feature preservation is of high importance, high constraint strength strictly limits the vertex position offset, precisely locking the key geometric shape, sharp edges, and feature line orientation of the mesh, preventing distortion or loss of core features during smoothing. For smoothing transition regions where feature preservation is of low importance, low constraint strength appropriately relaxes the vertex position adjustment range, efficiently eliminating mesh noise and local burrs in these regions, achieving uniform smoothing optimization of the mesh surface.

[0080] S104. Based on the optimized geometric features, obtain the 3D reconstruction data corresponding to the mesh data.

[0081] The 3D reconstruction method provided in this application proposes a hierarchical constraint weight allocation mechanism to dynamically adjust the smoothing intensity based on the importance of feature retention, thus avoiding excessive smoothing of feature lines. It also proposes a multi-scale progressive smoothing strategy to remove large-scale noise at the coarse-grained layer, optimize local geometry at the fine-grained layer, and continuously adapt to the constraint intensity differences in each region during progressive iteration. This dynamically balances the feature retention accuracy and smoothing optimization effect in different regions, ultimately achieving refined polishing and quality improvement across the entire geometric feature domain. The result is optimized geometric features that combine core feature integrity, surface smoothness, and topological regularity, thereby improving the accuracy of the 3D reconstruction results.

[0082] In one possible implementation, the geometric feature includes feature line vertex regions, feature line neighborhood regions, transition regions, and flat regions with decreasing feature preservation importance; the constraint strength corresponding to different regions in the geometric feature is determined based on the feature preservation importance corresponding to different regions, including:

[0083] The constraint strength corresponding to the vertex region of the feature line is determined as the first constraint strength;

[0084] The constraint strength corresponding to the neighborhood region of the feature line is determined as the second constraint strength;

[0085] The constraint strength corresponding to the transition region is determined to be the third constraint strength;

[0086] The constraint strength corresponding to the flat region is determined to be the fourth constraint strength;

[0087] Among them, the corresponding constraint strengths of the first constraint strength, the second constraint strength, the third constraint strength and the fourth constraint strength decrease in sequence.

[0088] For example, a four-level constraint system is established, and constraint strength is allocated according to the differences in importance based on the characteristics of different regions:

[0089] The first level is a hard constraint layer targeting the vertex region of the feature line. The vertices on the feature line are completely fixed and cannot be moved at any position. The constraint strength coefficient of the hard constraint layer is... =∞ (mathematically treated as an equality constraint). In the implementation of the Laplace smoothing algorithm, the corresponding row in the Laplace matrix is ​​set as a unit vector, and the right side is set as the target position.

[0090] The second level is a strong constraint layer targeting the direct neighborhood region, with 2-3 rings of neighboring vertices on each side of the feature line, allowing for a very small range of movement. The constraint strength coefficient of the strong constraint layer is: =10.0~100.0 (adaptively adjusted according to curvature), and a buffer is established for strongly constrained layers to prevent feature diffusion effects.

[0091] The third level is a moderately constrained layer for the transition region, located 4-6 ring vertices from the feature line, balancing smoothness and preservation. The constraint strength coefficient of the moderately constrained layer is: =1.0~10.0, for medium constraint layers, use secondary constraints to allow moderate deformation but control the magnitude of change.

[0092] The fourth level is a weak constraint layer for flat regions; regions far from features require sufficiently smooth optimization. The constraint strength coefficient of the weak constraint layer is as follows: =0.1~1.0, for weakly constrained layers, use standard Laplacian smoothing to maximize the noise reduction effect.

[0093] Then, the constraint weights for the i-th vertex are dynamically calculated based on the local geometric features:

[0094]

[0095] in, The baseline constraint strength is directly assigned by the aforementioned hierarchical constraint system based on the constraint layer to which the vertex belongs, and it determines the approximate protection level for that region. Fine-tuning is then performed based on the local geometry of the vertex, building upon this baseline constraint strength. For curvature, vertices with large curvature are usually located in characteristic regions and require stronger protection; This is the geodesic distance from the current vertex to the nearest feature line. The closer to the feature line, the stronger the protection should be to ensure a smooth transition of the feature's influence range. For local mesh quality, regions with poor quality (such as narrow triangles) are prone to distortion during smoothing, requiring appropriate increases in constraints to stabilize the optimization process. α, β, and γ are coefficients that adjust the weights of the three factors, and can be obtained through fitting, calibration, and other methods.

[0096] This application provides an embodiment of establishing a hierarchical constraint system, which maintains the importance of features corresponding to different regions and dynamically determines the constraint strength of different levels. This provides a basis for subsequent smoothing of different regions with different strengths, ensuring that important features are not over-smoothed while fully optimizing non-important features, thereby improving the accuracy of 3D modeling.

[0097] In one possible implementation, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features, including:

[0098] In the first resolution dimension, based on the Laplacian smoothing algorithm, according to the constraint strength corresponding to different regions in the geometric features, large-scale noise and global distortion in the geometric features are removed to obtain the first optimized feature. The first resolution dimension is the dimension that can display the vertex of the feature line.

[0099] In the second resolution dimension, based on the Laplacian smoothing algorithm, the first optimized feature is restored in detail according to the constraint strength corresponding to different regions in the geometric features to obtain the second optimized feature. The second resolution dimension is the dimension that can display all geometric features.

[0100] In the third resolution dimension, based on the Laplacian smoothing algorithm, the details of the second optimized feature are restored according to the constraint strength of different regions in the geometric features, and the position of the mesh vertices in the second optimized feature is optimized to obtain the optimized geometric features. The resolution corresponding to the third resolution dimension is smaller than the resolution corresponding to the second resolution dimension.

[0101] Specifically, constraint weights It is about building the entire optimization model. The key part, in the equation middle, It is the Laplace matrix, representing the smoothing term. It is a diagonal constraint weight matrix, and the first element on the diagonal is... The element is precisely the vertex. constraint weights , The term is the constraint term. The larger the value, the more likely it is to be the corresponding vertex during the solution process. The more a feature is positioned, the more likely it is to remain unchanged (or move only slightly), thus protecting the feature.

[0102] After receiving the data with assigned constraint weights, a three-tiered processing pyramid is built, processing the data sequentially from coarse to fine resolution. The essential difference between the three levels lies in the mesh resolution (scale) and the level of the constraint objective. This is a progressive refinement process from "global" to "local" and then to "detail." When constructing the corresponding linear system at each resolution dimension, the constraint weights are recalculated based on the mesh at the current resolution. In the first resolution dimension, only the identified core feature lines (hard constraints) and their immediate neighboring regions (strong constraints) are assigned high resolution. Other areas The resolution is very low, allowing for significant smoothing to remove large-scale noise. In the second resolution dimension, coarse-grained constraint information is transmitted and subdivided, with all four levels of constraint systems participating. Recalculated based on the subdivided vertex positions, with the intensity slightly lower than the final level to allow for some mesoscale adjustment. Full, precise constraint weights are used in the second resolution dimension. Then, make final fine-tuning adjustments.

[0103] Specifically, at the first resolution dimension, large-scale noise and global distortion are removed, and constraint processing retains only hard and strong constraints. The goal at the first resolution dimension is to eliminate global distortion and large-scale noise. Operating on an extremely simplified model results in low computational cost and allows for rapid correction of the overall shape. Retaining only the most critical hard and strong constraints while ignoring detailed features is equivalent to correcting the overall outline on a "sketch".

[0104] In the second resolution dimension, the results are subdivided based on the coarse-grained outcome to restore medium-scale details. All level-four constraints are involved, but their intensity is appropriately reduced. The goal in the second resolution dimension is to restore and optimize medium-scale geometric features based on the corrected overall outline. The level-four constraint system begins to work in full force, but its intensity is appropriately reduced to leave room for final fine adjustments, which is equivalent to depicting the main light and shadow relationships on a "sketch".

[0105] In the third resolution dimension, the medium-granularity results are refined to the original resolution. The complete constraint system precisely controls the movement of each vertex, and additional detail enhancement steps are added. Precise optimization is performed to perfectly preserve all features while removing small-scale noise and enhancing details. The complete and precise constraint system works at full capacity, which is equivalent to performing final polishing and finishing on a "high-definition canvas".

[0106] The output of each resolution dimension serves as the input for the next dimension, ultimately generating optimized grid data.

[0107] A sparse linear system Ax=b is constructed using a mathematical optimization model and solved using the preprocessed conjugate gradient method, where A=LᵀL+λW_c, b is the right-hand vector composed of constraint terms, and x is the vertex position vector to be solved.

[0108] The 3D reconstruction method provided in this application employs a multi-scale progressive smoothing strategy, optimizing geometric features to varying degrees at different resolution dimensions. It removes large-scale noise at the coarse-grained layer and optimizes local geometric features at the fine-grained layer. While preserving the complex feature structure, it effectively removes inherent noise and irregularities in the 3D scan data, laying a solid foundation for subsequent retopology processing. Through hierarchical constraints and progressive optimization strategies, an optimal balance is achieved between computational efficiency and result quality.

[0109] In one possible implementation, the progressive Laplace smoothing algorithm includes the following mechanism:

[0110] The dihedral angle restoration mechanism forces the faces on both sides of the feature edge to retain the original dihedral angle;

[0111] A curvature continuity optimization mechanism maintains curvature continuity along the direction of the feature line;

[0112] An adaptive iterative stopping mechanism stops smoothing when the following stopping conditions are detected: the change in position of each vertex after smoothing is less than the position change threshold, the change in the average dihedral angle of the feature edges tends to stabilize, and the system residual norm is less than the preset tolerance.

[0113] Specifically, the following mechanisms are applied sequentially during the smoothing process to maintain the sharpness of geometric features:

[0114] Dihedral angle recovery mechanism: The original dihedral angle is forcibly maintained on both sides of the feature edge and integrated into the optimization system through the angle constraint term E_angle; Curvature continuity mechanism: Curvature continuity is maintained along the feature line direction and a curvature smoothing term E_curvature is added to avoid curvature abrupt changes; Adaptive iteration stopping mechanism: Intelligent stopping conditions are designed to avoid over-smoothing or under-smoothing.

[0115] Intelligent stopping condition combining multiple criteria: Calculate the maximum change in position of all vertices between two iterations, when... ( Stop when the average dihedral angle of the feature edge changes; stop when the change tends to stabilize to prevent over-smoothing from causing the feature to become rounded; in solving linear systems When, when residual The iteration stops when the norm is less than the set tolerance.

[0116] The three-dimensional reconstruction method provided in this application embodiment sets up a dihedral angle recovery mechanism, a curvature continuous optimization mechanism, and an adaptive iteration stopping mechanism during the smoothing process. The three mechanisms process the data in sequence to ensure that the sharpness of geometric features is effectively maintained.

[0117] In one possible implementation, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features, including:

[0118] The grid data is divided into multiple sub-regions, and the Schwarz alternation method is used to coordinate the boundary consistency of multiple sub-regions.

[0119] Within multiple sub-regions, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features, resulting in optimized geometric features.

[0120] Specifically, based on the topological structure, geometric feature distribution patterns, and feature preservation priorities of the grid data, the overall grid data is adaptively divided into multiple sub-regions with independent topological integrity and interconnected boundaries. Then, the Schwarz alternation method is introduced to carry out cross-domain boundary coordination and iterative optimization for each sub-region. Using the overlapping boundaries of each sub-region as the information transmission carrier, a closed-loop process of alternating solution between sub-regions, bidirectional feedback of boundary information, and constraint correction is used to continuously calibrate the boundary vertex positions, feature line directions, and topological connection relationships of each sub-region. This effectively resolves problems such as boundary misalignment and feature discontinuity that occur after sub-region segmentation, and strictly ensures the geometric consistency and topological continuity of the boundaries between multiple sub-regions.

[0121] Simultaneously, within each independent sub-region after boundary coordination, the advantages of the progressive Laplacian smoothing algorithm's multi-round iterations and step-by-step convergence are fully utilized. Targeted optimization is carried out by combining the pre-set differentiated constraint strengths for different regions in the geometric features. High constraint strength is applied to core regions where feature preservation is of high importance to lock key geometric shapes, sharp edges, and feature details. Constraint strength is appropriately reduced in smooth transition regions where feature preservation is of low importance to efficiently filter out mesh noise and local burrs, achieving a dynamic balance between feature preservation accuracy and smoothing effect within the sub-region. Finally, through boundary consistency coordination of the entire sub-region and refined smoothing optimization of the partitions, optimized geometric features are obtained with smooth boundary connections between each sub-region, complete preservation of core geometric features, and a smooth and regular overall surface.

[0122] The 3D reconstruction method provided in this application divides a large-scale mesh into multiple sub-regions. Each sub-region independently constructs and solves a linear system. The Schwarz alternation method is used to coordinate boundary consistency, enabling parallel processing of mesh data and improving the processing efficiency of the 3D reconstruction process.

[0123] In one possible implementation, it also includes:

[0124] During the optimization of geometric features, when the positional change of the geometric feature is detected to be greater than the preset positional change threshold, the Laplacian matrix terms in the progressive Laplacian smoothing algorithm are updated, and a hot-start strategy based on the conjugate gradient method is used to improve the computational efficiency of the progressive Laplacian smoothing algorithm.

[0125] Specifically, in the later stages of smooth iteration, most vertices move very little, and only the Laplacian matrix terms of vertices whose position changes exceed the threshold are updated. The hot-start strategy using the conjugate gradient method can improve computational efficiency.

[0126] In the Laplace smoothing algorithm, each element L_ij in the Laplace matrix L defines the geometric relationship weight between vertices i and j, which is related to the length of the edge, the area of ​​the adjacent triangle, etc. Updating its Laplace matrix elements means that when the position of a vertex changes significantly during iteration, the geometric relationships of all edges within the one or two-ring neighborhood centered on that vertex change. Therefore, it is necessary to recalculate all non-zero elements in the rows and columns of matrix L related to these vertices, i.e., L_ij.

[0127] The 3D reconstruction method provided in this application adopts an incremental update mechanism, updating only the relevant matrix items for "active vertices" that have undergone significant positional changes. This avoids reconstructing the entire huge sparse matrix L in each iteration, thereby greatly saving computation time and memory bandwidth. Combined with warm start, that is, using the solution of the previous iteration as the initial value for the current iteration, it can significantly accelerate the solution of large linear systems.

[0128] In one possible implementation, the geometric features include corner points or endpoints. The constraint strength corresponding to different regions within the geometric feature is determined based on the importance of features corresponding to different regions, including:

[0129] The protection range is determined by using the corner point or endpoint as the center and the preset protection radius length;

[0130] Within the protection zone, the constraint strength increases exponentially as the distance to the corner or endpoint decreases.

[0131] Specifically, enhanced protection is adopted for corners and endpoints: corners are identified by the rate of change of normal and curvature extrema, the radius-weighted constraint strength increases exponentially, and edge folding or vertex merging operations are prohibited near corners.

[0132] Optionally, corner points can be located by the extreme points of abrupt changes in the rate of change of normal and the extreme points of Gaussian curvature, such as the intersection of the roofline and the A-pillar.

[0133] A protection radius R is defined with the corner point or endpoint as the center. Within this radius, the constraint strength λ increases exponentially as the distance r from the corner point decreases, i.e. This means that the closer to a corner point, the stronger the constraint becomes, forming a "hardened" protective shell that strictly prohibits vertex movement. Within this protected area, operations in the optimization algorithm that might change the topology or drastically alter the geometry, such as edge folding and vertex merging, are explicitly prohibited.

[0134] In one implementation, a combination of piecewise linearization and spline fitting is used for curve features: the continuous curve is discretized into piecewise linear segments, each segment is processed separately to ensure continuity within the segment, and finally, B-spline fitting is used to recover the smooth curve features. Special feature processing strategies and performance optimization techniques work together to ensure comprehensive feature protection.

[0135] The 3D reconstruction method provided in this application defines the protection boundary by delineating the protection radius, avoiding the interference of constraints in irrelevant areas that affect the overall smoothing efficiency. It also relies on the exponential growth distribution law of constraint intensity to ensure that the core positions closer to the corners and endpoints receive stronger constraint control, firmly locking the geometric shape of key features to prevent distortion and displacement. At the same time, the constraints are gradually weakened in the area extending outward from the protection range, achieving a precise balance between the rigid preservation of core features and the flexible smoothing of the surrounding areas. This effectively avoids problems such as corner blunting and endpoint loss, maximizing the integrity and recognizability of the mesh geometric features.

[0136] In one possible implementation, the 3D reconstruction data corresponding to the mesh data is obtained based on optimized geometric features, including:

[0137] Based on the feature lines contained in the optimized geometric features, retopology sampling points are generated using the feature distance function;

[0138] The feature lines contained in the optimized geometric features are used as constraint edges. Based on the Delaunay triangulation algorithm, the first topological mesh is generated according to the retopological sampling points.

[0139] An improved Ruppert algorithm is used to optimize the first-level topological mesh to obtain the second-level topological mesh;

[0140] Based on the preset mesh quality, the second topological mesh is iteratively optimized to obtain the 3D reconstruction data corresponding to the acquired mesh data.

[0141] Specifically, the retopology process employs a feature-guided strategy, with each step executed sequentially:

[0142] Density field generation:

[0143] The density field is generated based on the feature distance function: ρ(v) = 1 + α·exp(-d(v,F) / σ) + β·κ(v), where d(v,F) is the distance from the vertex to the feature line, and κ(v) is the curvature value. α and β are weighting coefficients, obtained through fitting or calibration. High-density sampling points are generated around the feature line, and flat regions are appropriately simplified. The result is then passed to the constraint triangulation step.

[0144] Constrained Triangulation: Receive the density field data and embed the extracted feature lines as constraint edges into the Delaunay triangulation process. An improved Ruppert algorithm is used to ensure the integrity of the feature edges, and a minimum angle constraint (≥25°) is set to avoid the generation of elongated triangles. The triangulated mesh data is output for local topology optimization.

[0145] Local topology optimization:

[0146] The mesh receives the constrained triangulation output, performs edge flipping, vertex relocation, and local subdivision operations, uses quality evaluation metrics (aspect ratio, minimum angle, Jacobian condition number) to guide the optimization process, and establishes an iterative optimization loop until all mesh quality metrics meet the standards.

[0147] The 3D reconstruction method provided in this application embodiment is based on a feature-guided density field generation algorithm, which maintains a high sampling density in the feature region; implements a constraint triangulation process to ensure that feature edges are accurately recovered after retopology; and constructs a mesh quality evaluation and optimization loop to continuously improve the topology quality through local operations.

[0148] In one possible implementation, extracting geometric features from the grid data includes:

[0149] Based on an improved quadratic error metric algorithm, the curvature corresponding to the grid vertices in the grid data is calculated;

[0150] Identify potential feature regions based on the curvature corresponding to the grid vertices;

[0151] A dihedral angle calculation algorithm based on a half-side data structure is adopted to filter feature edges from potential feature regions according to a preset angle threshold.

[0152] Gaussian curvature distribution map is used to identify concave and convex regions. Combined with prior knowledge of the shape of the 3D reconstructed object, false features in the feature edges are eliminated to obtain geometric feature data.

[0153] Using prior engineering knowledge, semantic features and prior confidence maps corresponding to geometric feature data are obtained based on geometric feature data.

[0154] Geometric features are obtained by multimodal fusion of geometric feature data, semantic features, and prior confidence maps.

[0155] Specifically, this stage employs a multi-level feature recognition strategy, with each level processing the data sequentially, as detailed below:

[0156] First, geometric feature analysis is performed on the input mesh. The improved QEM algorithm is used to calculate vertex curvature, and potential feature regions are identified by the rate of change of normal. The standard QEM algorithm is mainly used for mesh simplification. It calculates the sum of squared distances from a vertex to its associated triangular face, i.e., quadratic error, as a measure of the geometric distortion caused by deleting that vertex.

[0157] The standard QEM matrix (or Q matrix for short) is calculated for each vertex by summing the basic quadratic form matrices corresponding to all triangular faces within a ring neighborhood. This matrix encapsulates the geometric information of the local region of that vertex. The improved QEM algorithm's process and key improvement lie in using the Q matrix of vertex v to define a best-fit quadratic surface f(p) = p^TQp (where p is a 3D point) for its local neighborhood. From this quadratic surface function, the principal curvature, Gaussian curvature, and mean curvature can be analytically estimated by analyzing its Hessian matrix (second derivative matrix) at vertex v. Traditional discrete curvature estimation (such as based on normal variation or ring length ratio) is very sensitive to noise. This method performs a least-squares best-fit surface locally using the QEM matrix, which is equivalent to smoothing and regularizing the noisy data. Then, the curvature is extracted from the fitted smooth model, thus exhibiting stronger noise resistance and greater stability.

[0158] Secondly, dihedral angle calculation based on a half-edge data structure is employed, with a dynamic threshold (15°-25° adaptive) set to filter feature edges. Concave and convex regions are identified using a Gaussian curvature distribution map, and false features are eliminated by combining prior knowledge of vehicle shape. After processing, the geometric feature data is passed to the next level for further fusion. The geometric feature data is a structured set of feature information, rather than the original mesh, and mainly includes:

[0159] Feature edge / feature point set: Edges and their endpoints that identify significant geometric transitions (such as body edges) and are selected by dynamic dihedral angle thresholding;

[0160] Vertex curvature: The curvature value of each vertex (Gaussian curvature, mean curvature, etc.), used to identify concave and convex regions;

[0161] Region Labels: Based on Gaussian curvature and other factors, the mesh surface is divided into semantic parts such as "ridges", "valleys", and "flat regions";

[0162] Additional attributes: such as the calculated dihedral angle, rate of change of normal, etc.

[0163] Furthermore, taking the 3D reconstruction process of vehicles as an example, a probability distribution model of vehicle feature lines is established based on geometric feature data. The vehicle body bounding box is divided into 8 semantic regions to obtain semantic features. Corresponding feature line template libraries are loaded for different vehicle models (sedans, SUVs, trucks), and a Gaussian mixture model is used to probabilistically model the feature line positions, generating a prior confidence map. Engineering semantic information and geometric feature data are initially fused, and the results are used for multimodal feature fusion.

[0164] Finally, a weighted fusion function is designed: F_final = α·F_geom + β·F_prior + γ·F_context, where F_geom represents the geometric feature data, F_prior is the prior confidence map, F_context represents the semantic features, and α, β, and γ are adaptive weight coefficients that are dynamically adjusted based on mesh quality and noise level. A graph cut algorithm is used to find the optimal continuous path within the feature candidate region to ensure feature line continuity. The final output is the fused geometric features, which are used for subsequent smoothing and 3D retopology stages.

[0165] The three-dimensional reconstruction method provided in this application establishes a joint optimization model of geometric features and engineering priors, combining geometric features such as discrete curvature and dihedral angles with prior knowledge of feature line positions in the vehicle bounding box; designs a feature path optimization method based on graph cut algorithm to find the optimal continuous path in the feature candidate region; and constructs a feature confidence evaluation system to improve the reliability of feature extraction through multi-index fusion.

[0166] Figure 2 This is a schematic diagram of the three-dimensional reconstruction device provided in an embodiment of this application. Figure 2 As shown, this application embodiment provides a three-dimensional reconstruction device 20, including:

[0167] Extraction module 201 is used to acquire grid data and extract geometric features from the grid data;

[0168] The determination module 202 is used to determine the constraint strength corresponding to different regions of the geometric feature based on the feature retention importance corresponding to different regions in the geometric feature;

[0169] The feature optimization module 203 is used to optimize the geometric features based on the progressive Laplace smoothing algorithm and according to the constraint strength corresponding to different regions in the geometric features, so as to obtain optimized geometric features.

[0170] The 3D reconstruction module 204 is used to obtain the 3D reconstruction data corresponding to the mesh data based on the optimized geometric features.

[0171] In one possible implementation, the geometric features include feature line vertex regions, feature line neighborhood regions, transition regions, and flat regions where feature importance decreases; the determining module 202 is specifically used for:

[0172] The constraint strength corresponding to the vertex region of the feature line is determined as the first constraint strength;

[0173] The constraint strength corresponding to the neighborhood region of the feature line is determined as the second constraint strength;

[0174] The constraint strength corresponding to the transition region is determined to be the third constraint strength;

[0175] The constraint strength corresponding to the flat region is determined to be the fourth constraint strength;

[0176] Among them, the corresponding constraint strengths of the first constraint strength, the second constraint strength, the third constraint strength and the fourth constraint strength decrease in sequence.

[0177] In one possible implementation, the feature optimization module 203 is specifically used for:

[0178] In the first resolution dimension, based on the Laplacian smoothing algorithm, according to the constraint strength corresponding to different regions in the geometric features, large-scale noise and global distortion in the geometric features are removed to obtain the first optimized feature. The first resolution dimension is the dimension that can display the vertex of the feature line.

[0179] In the second resolution dimension, based on the Laplacian smoothing algorithm, the first optimized feature is restored in detail according to the constraint strength corresponding to different regions in the geometric features to obtain the second optimized feature. The second resolution dimension is the dimension that can display all geometric features.

[0180] In the third resolution dimension, based on the Laplacian smoothing algorithm, the details of the second optimized feature are restored according to the constraint strength of different regions in the geometric features, and the position of the mesh vertices in the second optimized feature is optimized to obtain the optimized geometric features. The resolution corresponding to the third resolution dimension is smaller than the resolution corresponding to the second resolution dimension.

[0181] In one possible implementation, the feature optimization module 203 includes the following mechanism:

[0182] The dihedral angle restoration mechanism forces the faces on both sides of the feature edge to retain the original dihedral angle;

[0183] A curvature continuity optimization mechanism maintains curvature continuity along the direction of the feature line;

[0184] An adaptive iterative stopping mechanism stops smoothing when the following stopping conditions are detected: the change in position of each vertex after smoothing is less than the position change threshold, the change in the average dihedral angle of the feature edges tends to stabilize, and the system residual norm is less than the preset tolerance.

[0185] In one possible implementation, the feature optimization module 203 is further configured to:

[0186] The grid data is divided into multiple sub-regions, and the Schwarz alternation method is used to coordinate the boundary consistency of multiple sub-regions.

[0187] Within multiple sub-regions, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features, resulting in optimized geometric features.

[0188] In one possible implementation, the feature optimization module 203 is further configured to:

[0189] During the optimization of geometric features, when the positional change of the geometric feature is detected to be greater than the preset positional change threshold, the Laplacian matrix terms in the progressive Laplacian smoothing algorithm are updated, and a hot-start strategy based on the conjugate gradient method is used to improve the computational efficiency of the progressive Laplacian smoothing algorithm.

[0190] In one possible implementation, the geometric features include corner points or endpoints, and the determining module 202 is further configured to:

[0191] The protection range is determined by using the corner point or endpoint as the center and the preset protection radius length;

[0192] Within the protection zone, the constraint strength increases exponentially as the distance to the corner or endpoint decreases.

[0193] In one possible implementation, the three-dimensional reconstruction module 204 is specifically used for:

[0194] Based on the feature lines contained in the optimized geometric features, retopology sampling points are generated using the feature distance function;

[0195] The feature lines contained in the optimized geometric features are used as constraint edges. Based on the Delaunay triangulation algorithm, the first topological mesh is generated according to the retopological sampling points.

[0196] An improved Ruppert algorithm is used to optimize the first-level topological mesh to obtain the second-level topological mesh;

[0197] Based on the preset mesh quality, the second topological mesh is iteratively optimized to obtain the 3D reconstruction data corresponding to the acquired mesh data.

[0198] In one possible implementation, the extraction module 201 is specifically used for:

[0199] Based on an improved quadratic error metric algorithm, the curvature corresponding to the grid vertices in the grid data is calculated;

[0200] Identify potential feature regions based on the curvature corresponding to the grid vertices;

[0201] A dihedral angle calculation algorithm based on a half-side data structure is adopted to filter feature edges from potential feature regions according to a preset angle threshold.

[0202] Gaussian curvature distribution map is used to identify concave and convex regions. Combined with prior knowledge of the shape of the 3D reconstructed object, false features in the feature edges are eliminated to obtain geometric feature data.

[0203] Using prior engineering knowledge, semantic features and prior confidence maps corresponding to geometric feature data are obtained based on geometric feature data.

[0204] Geometric features are obtained by multimodal fusion of geometric feature data, semantic features, and prior confidence maps.

[0205] The three-dimensional reconstruction device provided in this embodiment performs the method provided in the above method embodiment, and its implementation principle and technical effect are similar, so it will not be described in detail here.

[0206] Figure 3 This is a schematic diagram of the structure of the three-dimensional reconstruction device provided in an embodiment of this application. Figure 3 As shown, the 3D reconstruction device 30 provided in this embodiment includes at least one processor 301 and a memory 302. Optionally, the 3D reconstruction device 30 further includes a communication interface 303. The processor 301, memory 302, and communication interface 303 are connected via a communication bus 304.

[0207] In a specific implementation, at least one processor 301 executes computer execution instructions stored in memory 302, causing at least one processor 301 to perform the above-described method.

[0208] The specific implementation process of processor 301 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0209] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0210] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0211] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0212] This application also provides a computer program product, including a computer program that, when executed, implements the above-described method.

[0213] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed, implement the above-described method.

[0214] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0215] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an application-specific integrated circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0216] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0217] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0218] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0219] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0220] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0221] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A three-dimensional reconstruction method, characterized in that, include: Acquire grid data and extract geometric features from the grid data; The constraint strength corresponding to different regions of the geometric feature is determined based on the feature retention importance of different regions in the geometric feature. Based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features; Based on the optimized geometric features, obtain the 3D reconstruction data corresponding to the mesh data.

2. The three-dimensional reconstruction method according to claim 1, characterized in that, The geometric feature includes feature line vertex regions, feature line neighborhood regions, transition regions, and flat regions, with feature preservation importance decreasing. Determining the constraint strength corresponding to different regions in the geometric feature based on their feature preservation importance includes: The constraint strength corresponding to the vertex region of the feature line is determined as the first constraint strength; The constraint strength corresponding to the neighborhood region of the feature line is determined as the second constraint strength; The constraint strength corresponding to the transition region is determined to be the third constraint strength; The constraint strength corresponding to the flat region is determined to be the fourth constraint strength; The constraint strengths of the first constraint strength, the second constraint strength, the third constraint strength, and the fourth constraint strength decrease sequentially.

3. The three-dimensional reconstruction method according to claim 2, characterized in that, The progressive Laplacian smoothing algorithm optimizes the geometric features based on the constraint strength corresponding to different regions within the geometric features, resulting in optimized geometric features, including: In the first resolution dimension, based on the Laplacian smoothing algorithm, according to the constraint strength corresponding to different regions in the geometric features, large-scale noise and global distortion in the geometric features are removed to obtain the first optimized feature. The first resolution dimension is the dimension that can display the vertex of the feature line. In the second resolution dimension, based on the Laplacian smoothing algorithm, according to the constraint strength corresponding to different regions in the geometric features, the first optimized feature is restored in detail to obtain the second optimized feature. The second resolution dimension is the dimension that can display all geometric features. In the third resolution dimension, based on the Laplacian smoothing algorithm, according to the constraint strength corresponding to different regions in the geometric feature, the second optimized feature is restored in detail, and the position of the mesh vertex in the second optimized feature is optimized to obtain the optimized geometric feature. The resolution corresponding to the third resolution dimension is less than the resolution corresponding to the second resolution dimension.

4. The three-dimensional reconstruction method according to any one of claims 1 to 3, characterized in that, The progressive Laplace smoothing algorithm includes the following mechanisms: The dihedral angle restoration mechanism forces the faces on both sides of the feature edge to retain the original dihedral angle; A curvature continuity optimization mechanism maintains curvature continuity along the direction of the feature line; An adaptive iterative stopping mechanism stops smoothing when the following stopping conditions are detected: the change in position of each vertex after smoothing is less than the position change threshold, the change in the average dihedral angle of the feature edges tends to stabilize, and the system residual norm is less than the preset tolerance.

5. The three-dimensional reconstruction method according to any one of claims 1 to 3, characterized in that, The progressive Laplacian smoothing algorithm optimizes the geometric features based on the constraint strength corresponding to different regions within the geometric features, resulting in optimized geometric features, including: The grid data is divided into multiple sub-regions, and the Schwarz alternation method is used to coordinate the boundary consistency of the multiple sub-regions. Within the multiple sub-regions, based on the progressive Laplace smoothing algorithm, the geometric features are optimized according to the constraint strength corresponding to different regions in the geometric features to obtain optimized geometric features.

6. The three-dimensional reconstruction method according to any one of claims 1 to 3, characterized in that, Also includes: During the optimization of the geometric features, when the positional change of the geometric features is detected to be greater than the preset positional change threshold, the Laplacian matrix terms in the progressive Laplacian smoothing algorithm are updated, and a hot-start strategy based on the conjugate gradient method is used to improve the computational efficiency of the progressive Laplacian smoothing algorithm.

7. The three-dimensional reconstruction method according to any one of claims 1 to 3, characterized in that, The geometric features include corner points or endpoints. Determining the constraint strength corresponding to different regions of the geometric feature based on the importance of features corresponding to different regions includes: The protection range is determined with the corner point or endpoint as the center and a preset protection radius length. Within the protection range, the constraint strength increases exponentially as the distance to the corner or endpoint decreases.

8. The three-dimensional reconstruction method according to any one of claims 1 to 3, characterized in that, The step of obtaining the 3D reconstruction data corresponding to the mesh data based on the optimized geometric features includes: Based on the feature lines contained in the optimized geometric features, retopology sampling points are generated using a feature distance function; Using the feature lines contained in the optimized geometric features as constraint edges, and based on the Delaunay triangulation algorithm, a first topological mesh is generated according to the retopological sampling points. The first-level topological mesh is optimized using an improved Ruppert algorithm to obtain the second-level topological mesh; Based on the preset mesh quality, the second topological mesh is iteratively optimized to obtain the 3D reconstruction data corresponding to the obtained mesh data.

9. The three-dimensional reconstruction method according to any one of claims 1 to 3, characterized in that, The extraction of geometric features from the grid data includes: Based on the improved quadratic error metric algorithm, the curvature corresponding to the grid vertices in the grid data is calculated; Identify potential feature regions based on the curvature corresponding to the grid vertices; A dihedral angle calculation algorithm based on a half-side data structure is used to filter feature edges from the potential feature region according to a preset angle threshold. The Gaussian curvature distribution map is used to identify concave and convex regions. By combining prior knowledge of the shape of the 3D reconstructed object, false features in the feature edges are eliminated to obtain geometric feature data. Using prior engineering knowledge, semantic features and prior confidence maps corresponding to the geometric feature data are obtained based on the geometric feature data. The geometric features are obtained by multimodal fusion of the geometric feature data, semantic features, and prior confidence map.

10. A three-dimensional reconstruction device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-9.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed, are used to implement the method as described in any one of claims 1-9.

12. A computer program product, characterized in that, Includes a computer program, which, when executed, implements the method according to any one of claims 1-9.

Images