Automobile curved surface fitting method and device based on deep implicit function and physical constraint

Abstract

The invention provides an automobile curved surface fitting method and device based on a deep implicit function and physical constraints, and relates to the technical field of automobile design and manufacturing. The method comprises the steps of obtaining three-dimensional point cloud data of a target automobile, inputting the three-dimensional point cloud data into a pre-trained deep implicit function network, and generating a closed manifold triangular mesh curved surface; inputting the triangular mesh curved surface into a physically constrained NURBS conversion network, performing iterative optimization on NURBS curved surface parameters by minimizing a physical constraint loss function, and generating an NURBS parameterized curved surface conforming to engineering constraints; according to the method, the quality of the curved surface is evaluated, and when the target area with the fitting error exceeding the preset engineering tolerance is identified, the adaptive iterative optimization process is triggered until the fitting errors of all the areas are lower than the corresponding engineering tolerance, so that the automation degree, the fitting precision, the physical rationality and the like of automobile curved surface fitting are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of automotive design and manufacturing technology, and in particular to a method and apparatus for fitting automotive surfaces based on deep implicit functions and physical constraints. Background Technology

[0002] In the field of automotive design and manufacturing, the quality of the car body's curved surfaces directly affects the product's aerodynamic performance, aesthetic value, and structural strength, representing a core aspect of the fusion of engineering and art. With increasing market competition and growing personalized demands, the industry is placing higher demands on design efficiency and quality. Especially in reverse engineering and rapid prototyping, the ability to quickly and automatically convert massive, noisy, and even incomplete point cloud data acquired from 3D scanning equipment into high-precision, editable, and engineering-constrained computer-aided design (CAD) models has become a key technological bottleneck restricting industrial upgrading. Traditional modeling methods relying on human experience are time-consuming, costly, and struggle to guarantee the objective optimality of results, failing to meet the digital demands of modern intelligent manufacturing.

[0003] In related technologies, surface fitting schemes based on traditional implicit functions and those based on deep learning are commonly employed. Traditional implicit function-based surface fitting schemes include those utilizing radial basis functions (RBFs) or multi-level partitions of unity (MPUs). These schemes construct a spatial scalar field (implicit function) and define the zero isosurface of this function as the target surface, thereby achieving the transformation from discrete point clouds to continuous watertight meshes. Deep learning-based surface fitting schemes, such as deep implicit functions (DIFs), typically utilize neural networks to learn complex mappings from spatial coordinates to symbolic distances or occupancy probabilities, enabling end-to-end fitting of triangular mesh models with arbitrary topologies from point clouds. However, both of these schemes cannot simultaneously meet the multiple requirements of automation, fitting accuracy, and physical plausibility.

[0004] Therefore, there is an urgent need to provide a solution for fitting automotive surfaces that can overcome the limitations of existing technologies, so as to meet the modern automotive industry's demand for efficient, accurate, and reliable surface modeling. Summary of the Invention

[0005] This application provides a method and apparatus for fitting automotive surfaces based on deep implicit functions and physical constraints, in order to improve the problem that related technologies cannot simultaneously meet multiple requirements such as automation, fitting accuracy and physical rationality.

[0006] Firstly, this application provides a method for fitting automotive surfaces based on deep implicit functions and physical constraints, including:

[0007] Obtain the 3D point cloud data of the target vehicle;

[0008] The 3D point cloud data is input into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface;

[0009] A closed manifold triangular mesh surface is input into a physically constrained Non-Uniform Rational B-Spline (NURBS) transformation network. The NURBS surface parameters are iteratively optimized by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints. The physical constraint loss function includes data fitting loss, smoothness loss, and continuity loss.

[0010] The quality of the NURBS parametric surface is evaluated. When a target region with a fitting error exceeding the preset engineering tolerance is identified, an adaptive iterative optimization process for the target region is triggered until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0011] In one possible implementation, the deep implicit function network includes a point cloud feature encoder and an implicit function decoder. The process involves inputting 3D point cloud data into the pre-trained deep implicit function network to generate a closed manifold triangular mesh surface. This includes: extracting global shape feature vectors and local geometric features from the 3D point cloud data using the point cloud feature encoder; constructing a 3D query mesh within the spatial bounding box of the 3D point cloud data; inputting the global shape feature vectors and the spatial coordinates of each vertex in the 3D query mesh into the implicit function decoder to obtain the geometric state scalars corresponding to each vertex; wherein the geometric state scalars include signed distance values, unsigned distance values, or occupancy probability values; and using the moving cube algorithm to extract zero isosurfaces based on the geometric state scalars corresponding to each vertex of the 3D query mesh to generate a closed manifold triangular mesh surface.

[0012] In one possible implementation, a closed manifold triangular mesh surface is input into a physically constrained NURBS transformation network. The NURBS surface parameters are iteratively optimized by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints. This includes extracting the geometric features of the closed manifold triangular mesh surface, including boundary curves, feature lines, and curvature information. The boundary curves include the outer contour and internal opening boundaries of the closed manifold triangular mesh surface, and the feature lines are used to identify regions where the closed manifold triangular mesh surface undergoes drastic geometric changes. Curvature information includes principal curvature and Gaussian curvature; the closed manifold triangular mesh surface and extracted geometric features are input into a pre-defined parameter prediction network, which predicts the initial values ​​of NURBS surface parameters, including the positions, weights, and distribution of node vectors of the control point mesh; based on the predicted initial values ​​of the NURBS surface parameters, a differentiable physical constraint loss function is constructed; by minimizing the physical constraint loss function, the control point mesh, weights, and node vectors are iteratively optimized to generate a NURBS parametric surface that conforms to engineering constraints.

[0013] In one possible implementation, the physical constraint loss function satisfies the following formula:

[0014]

[0015] Where L_data is the data fitting loss, used to minimize the geometric distance between the NURBS surface and the closed manifold triangular mesh surface; L_fairness is the smoothness loss; and L_continuity is the continuity loss, used to force adjacent NURBS surface patches to achieve the target geometric continuity at the splicing boundary. , and These are the weighting coefficients for the corresponding losses.

[0016] In one possible implementation, the quality assessment of the NURBS parametric surface includes: determining the geometric error between the NURBS parametric surface and the closed manifold triangular mesh surface, and generating a corresponding error distribution chromatogram; based on the error distribution chromatogram, identifying local regions where the fitting error exceeds a preset engineering tolerance as target optimization regions, and obtaining a quality assessment result containing the identification information of the target optimization regions; wherein, the target optimization regions include at least one of edge regions where the curvature value exceeds a preset curvature threshold and splicing areas where multiple surfaces intersect.

[0017] In one possible implementation, triggering an adaptive iterative optimization process for the target region includes: performing an adaptive sampling operation within the corresponding target optimization region based on the target optimization region identification information in the quality assessment results to obtain encrypted sampling information. The adaptive sampling operation includes encrypting sampling points and / or increasing the control point grid density of the NURBS parametric surface within the parameter domain; feeding the encrypted sampling information back to the NURBS transformation network to perform an iterative optimization of the NURBS surface parameters by minimizing the physical constraint loss function until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0018] Secondly, this application provides a vehicle surface fitting device based on deep implicit functions and physical constraints, comprising:

[0019] The acquisition module is used to acquire the 3D point cloud data of the target vehicle.

[0020] The first processing module is used to input 3D point cloud data into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface.

[0021] The second processing module is used to input the closed manifold triangular mesh surface into the physically constrained NURBS transformation network, and iteratively optimize the NURBS surface parameters by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints; wherein, the physical constraint loss function includes data fitting loss, smoothness loss and continuity loss;

[0022] The quality assessment module is used to assess the quality of NURBS parametric surfaces. When a target region with a fitting error exceeding the preset engineering tolerance is identified, an adaptive iterative optimization process is triggered for the target region until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0023] Thirdly, this application provides an electronic device, including: a processor, and a memory communicatively connected to the processor;

[0024] Memory is used to store instructions executed by the computer;

[0025] A processor for executing computer-executable instructions stored in memory to implement any of the methods of the first aspect.

[0026] Fourthly, this application provides a computer-readable storage medium storing computer-executable instructions, which, when executed, are used to implement the method of any one of the first aspects.

[0027] Fifthly, this application provides a computer program product, including a computer program that, when executed, implements the method of any one of the first aspects.

[0028] This application provides a method and apparatus for fitting automotive surfaces based on deep implicit functions and physical constraints. The method acquires 3D point cloud data of a target vehicle, inputs the 3D point cloud data into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface, and inputs the closed manifold triangular mesh surface into a physically constrained NURBS transformation network. The NURBS surface parameters are iteratively optimized by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints. The physical constraint loss function includes data fitting loss, smoothness loss, and continuity loss. The quality of the NURBS parametric surface is evaluated. When a target region with a fitting error exceeding a preset engineering tolerance is identified, an adaptive iterative optimization process for the target region is triggered until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0029] In this process, through the collaborative work of a deep implicit function network and a physically constrained NURBS transformation network, end-to-end automatic fitting from raw 3D point cloud data to a high-quality parametric CAD model is achieved, effectively solving the problems of traditional solutions relying on human experience and having low automation. Among them, the deep implicit function network can robustly reconstruct a complete watertight mesh from noisy and incomplete point clouds, significantly improving the ability to handle complex topologies and data defects. The physically constrained NURBS transformation network, by fusing data fitting loss, smoothness loss, and continuity loss into a physical constraint loss function, ensures that the generated NURBS parametric surface maintains both geometric accuracy and meets engineering physical constraints, effectively overcoming the defect of physical unreliability in the output results of traditional pure data-driven solutions. In addition, the adaptive iterative optimization mechanism achieves intelligent control of fitting accuracy by automatically identifying and locally refining areas where errors exceed the standard. This results in the output of industry-standard NURBS surfaces that can be directly used in computer-aided engineering (CAE) or computer-aided manufacturing (CAM) processes, thereby significantly improving the efficiency, quality, and automation level of automotive surface design. Attached Figure Description

[0030] 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.

[0031] Figure 1A schematic diagram illustrating an application scenario of the automotive surface fitting method based on deep implicit functions and physical constraints, provided as an exemplary embodiment of this application.

[0032] Figure 2 A flowchart illustrating a method for fitting automotive surfaces based on deep implicit functions and physical constraints, provided as an exemplary embodiment of this application;

[0033] Figure 3 Another flowchart illustrating the automotive surface fitting method based on deep implicit functions and physical constraints provided for an exemplary embodiment of this application;

[0034] Figure 4 A schematic diagram of a vehicle surface fitting device based on deep implicit functions and physical constraints, provided as an exemplary embodiment of this application;

[0035] Figure 5 A schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application.

[0036] The accompanying drawings illustrate 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 particular embodiments. Detailed Implementation

[0037] 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.

[0038] The terms “first,” “second,” etc., used in the specification and claims of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented, for example, in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, products, or apparatus.

[0039] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0040] Considering that traditional automotive surface design heavily relies on CAD systems, its foundation is parametric curve and surface theory, such as... Curves, surfaces, B-splines, and NURBS—these methods define surface shapes using control points and node vectors. While technically mature and highly accurate, they reveal several bottlenecks when addressing the demands of efficient and intelligent modern design: 1) High dependence on human experience and intervention: The placement of initial control points, the determination of node vectors, and the smoothing of complex surfaces (such as Class A surfaces) heavily rely on engineers' experience and judgment, resulting in low automation and long design cycles with low efficiency; 2) Insufficient ability to handle complex topologies and large datasets: When faced with high-precision 3D (3... When acquiring massive, noisy, and potentially incomplete point cloud data (a common situation in reverse engineering) using Dimensions (3D) scanners, traditional parametric solutions require first dividing the point cloud into simple quadrilaterals and then fitting the pieces. This process is not only cumbersome, but also extremely difficult to ensure the smooth continuity between surfaces (such as G1 or G2 continuity) during stitching. 3) It lacks intelligence and adaptability, meaning that traditional methods have difficulty "understanding" the physical laws or design intentions behind the surface. For example, they cannot adaptively adjust the fitting complexity according to the point cloud characteristics, nor can they actively ensure that the generated surface meets specific engineering constraints (such as curvature continuity, aerodynamic performance, etc.).

[0041] In recent years, with the significant progress of deep learning technology in computer vision and geometric processing, some research has begun to explore its application in 3D geometric reconstruction. Deep implicit functions, in particular, have shown great potential. Specifically, by learning a neural network, any point (x, y, z) in space is mapped to a scalar value (such as signed distance or occupancy probability), thus implicitly representing the complex shapes of arbitrary topologies. However, purely data-driven deep learning methods still face significant challenges when applied to demanding industrial CAD applications: 1) The "black box" problem and physical unreliability: The network may output surfaces that violate geometric laws (such as curvature abrupt changes and self-intersections), lacking interpretability and failing to meet the high requirements of CAE and manufacturing CAM; 2) Decoupling from downstream CAD / CAM processes: The generated triangular mesh model lacks parametric information and cannot be intuitively edited and modified like NURBS surfaces, limiting its application in forward design. Therefore, the industry urgently needs a new method that can integrate the powerful adaptive capabilities of deep learning with the high precision and editability of traditional geometric modeling.

[0042] During their research, the inventors discovered that surface fitting schemes based on traditional implicit functions, which are not based on deep learning, have certain limitations, even though they have been applied in reverse engineering: a) Low level of intelligence: their fitting quality heavily depends on parameter settings (such as the kernel function of RBF or the subdivision threshold of MPU), lacking the ability to autonomously learn the optimal fitting method from data; b) Inability to incorporate physical constraints: it is difficult to embed engineering constraints such as the smoothness and continuity of the surface as prior knowledge into the fitting process; c) Limited adaptive capability: although MPU uses octree subdivision, its "adaptiveness" is mainly reflected in the response to spatial complexity, rather than the intelligent recognition and preservation of surface geometric features (such as edges and rounded corners). In addition, while deep learning-based surface fitting schemes demonstrate the speed and potential of deep learning in handling surface optimization problems and provide a new paradigm for data-driven geometric processing, they also have certain limitations: a) shallow constraints: their loss functions (such as loss based on Gaussian curvature) are usually constraints at the geometric and statistical level, rather than strong constraints derived from underlying physical laws such as continuum mechanics and aerodynamics, which may lead to the surface not being physically optimal; b) different objectives: this scheme mainly optimizes triangular meshes, and its final output is still a discrete mesh; c) failure to solve the CAD integration problem: this scheme does not address how to convert the results of deep learning into industry-standard CAD models.

[0043] In summary, the relevant technologies exhibit three interrelated and prominent contradictions: the contradiction between "automation requirements" and "reliance on human experience," the contradiction between "complex data processing" and "model universality," and the contradiction between "data-driven fitting" and "physical constraints." Based on these issues, this application proposes a vehicle surface fitting scheme based on deep implicit functions and physical constraints. Specifically, it proposes a technical concept that integrates the adaptive capabilities of deep learning with the high precision and editability of traditional geometric modeling. Its core lies in constructing an end-to-end intelligent surface fitting system, aiming to systematically solve the three core contradictions of traditional methods: reliance on human experience and insufficient ability to handle defective data; and the disconnect between purely data-driven methods and industrial processes, lacking physical guarantees. Specifically, this includes the following three aspects: 1) Addressing the contradiction between automation requirements and reliance on human experience, a deep implicit function network is designed, enabling the processing device to automatically "understand" the geometric features and topological structure of point cloud data, autonomously reconstructing a complete watertight surface from defective data, effectively eliminating reliance on manual segmentation, point placement, and splicing. 2) To resolve the contradiction between complex data processing and model universality, a physically constrained NURBS transformation network is introduced. This network not only converts the intelligently reconstructed mesh into industry-standard NURBS parametric surfaces, but also ensures that the generated surfaces inherently conform to engineering principles by deeply integrating physical constraints such as data fitting loss, smoothness loss, and continuity loss. This results in strong adaptability to defective data and compatibility with downstream CAD / CAE / CAM processes. 3) To ensure the reliability of the fitting results, an adaptive iterative optimization mechanism is integrated. Through a closed loop of quality assessment and re-optimization, the surface quality is autonomously optimized, ensuring that the final output model meets stringent engineering and aesthetic requirements in terms of accuracy, smoothness, and continuity. This, in turn, connects the entire chain from 3D scanning to digital manufacturing.

[0044] Figure 1 This is a schematic diagram illustrating an application scenario of the automotive surface fitting method based on deep implicit functions and physical constraints, provided as an exemplary embodiment of this application. For example... Figure 1 As shown, this application scenario includes a client 11 and a server 12, wherein the number of clients 11 can be at least one. In practical applications, when the server 12 detects the 3D point cloud data of the target car submitted by the client 11, it executes the car surface fitting method based on depth implicit functions and physical constraints provided in this application, and outputs the NURBS parametric surface corresponding to the target car.

[0045] It should be noted that server 12 can also be replaced by a server cluster or other computing devices with a certain computing power. Both the first and second clients can be mobile phones, computers, laptops, or personal digital assistants (PDAs).

[0046] The following is combined Figure 1 Application scenarios, refer to Figure 2 This paper describes a method for fitting automotive surfaces based on depth implicit functions and physical constraints, according to exemplary embodiments of this application. It should be noted that the above application scenarios are shown only to facilitate understanding the spirit and principles of this application, and the embodiments of this application are not limited to... Figure 1 The limitations of the application scenarios shown are explained. It should be noted that the following 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.

[0047] Figure 2 This is a flowchart illustrating a method for fitting automotive surfaces based on deep implicit functions and physical constraints, provided as an exemplary embodiment of this application. Figure 2 As shown, the method for fitting car surfaces based on deep implicit functions and physical constraints includes the following steps:

[0048] S201. Obtain the 3D point cloud data of the target vehicle.

[0049] For example, high-precision 3D scanning equipment, such as laser scanners or structured light scanners, is used to perform non-contact 3D scanning of the target car's body, clay model, or key components to obtain a set of 3D spatial coordinates of its body surface, i.e., 3D point cloud data. This data serves as the raw input for the entire surface fitting process. Correspondingly, referencing... Figure 1 The client 11 acquires the 3D point cloud data of the target car collected by the 3D scanning device and sends it to the server 12; correspondingly, the server 12 acquires the 3D point cloud data of the target car.

[0050] It should be noted that this application is specifically designed for "imperfect" data commonly found in real-world engineering projects. Therefore, the obtained 3D point cloud data is usually noisy (with scattered points that deviate from the real surface), uneven in density (sparse in flat areas and dense in areas with rich features), and may contain some missing data (due to occlusion or reflection).

[0051] S202. Input the 3D point cloud data into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface.

[0052] For example, the three-dimensional point cloud data obtained in step S201 is input into a pre-trained deep implicit function network. Through the inference calculation of the network, a closed manifold triangular mesh surface is fitted from the input data that may have defects, that is, a complete, smooth and watertight triangular mesh surface.

[0053] The deep implicit function network employs a dual-path architecture combining global point cloud features and local spatial coordinates, and undergoes end-to-end training using physical regularization loss. The specific training process includes: constructing a large-scale training dataset containing various typical automotive surface shapes and their corresponding 3D point cloud data. The point cloud data is preprocessed to simulate common issues in actual 3D scanning, such as noise interference, uneven density distribution, and missing data. During training, the network simultaneously learns the global shape features and local geometric details of the point cloud through the dual-path architecture. The feature extraction path extracts multi-scale feature representations of the point cloud, while the function fitting path maps spatial coordinates to corresponding geometric state scalars based on these features. In terms of optimization objectives, in addition to minimizing reconstruction accuracy loss, a physical regularization loss based on the differential geometric properties of the surface is introduced. This loss constrains the surface curvature and smoothness to ensure that the implicit function learned by the network conforms to the physical laws of the real surface.

[0054] It should be noted that the closed manifold triangular mesh surface in this step refers to the generated triangular mesh surface having complete geometric closure, without any holes, cracks, or non-manifold structures, thus meeting the basic requirements for geometric model integrity in subsequent CAD / CAE / CAM processes. This characteristic is achieved thanks to the powerful geometric priors and physical constraints learned during the training process of the deep implicit function network, enabling it to infer the complete surface morphology from defective input data.

[0055] S203. Input the closed manifold triangular mesh surface into the physically constrained NURBS transformation network, and iteratively optimize the NURBS surface parameters by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints; wherein, the physical constraint loss function includes data fitting loss, smoothness loss and continuity loss.

[0056] Among them, data fitting loss: used to measure the geometric difference between the generated NURBS surface and the input triangular mesh surface, ensuring that the reconstructed surface is faithful to the original data in terms of geometry; smoothness loss: based on the principle of minimizing the energy of surface curvature, by suppressing unnecessary fluctuations and wrinkles of the surface, the generated NURBS surface meets the requirements of automotive aerodynamics and aesthetic appearance compliance; continuity loss: when the surface is composed of multiple NURBS surface patches, this loss forces adjacent surface patches to meet the specified geometric continuity conditions at the splicing boundary, ensuring a smooth transition of the entire surface.

[0057] It should be noted that NURBS is the standard surface representation in the field of computer-aided design, which achieves a precise mathematical description of the surface through parameters such as control points, weights, and node vectors; engineering constraints refer to the specific requirements for surface quality in the automotive industry, including but not limited to indicators such as surface smoothness, continuity, and manufacturability.

[0058] Correspondingly, the NURBS transformation network based on physical constraints automatically quantifies the gradient information of the physical constraint loss function with respect to each parameter of NURBS through differentiable optimization techniques, and continuously adjusts the relevant surface description parameters based on the gradient descent principle until the optimal surface solution that satisfies all engineering constraints is found. This optimization mechanism based on physical constraints effectively overcomes the shortcomings of traditional pure data-driven methods in terms of engineering applicability, and realizes intelligent transformation from geometric models to engineering models.

[0059] S204. Perform quality assessment on the NURBS parametric surface. When a target region with a fitting error exceeding the preset engineering tolerance is identified, trigger an adaptive iterative optimization process for the target region until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0060] In some embodiments, the quality assessment of the NURBS parametric surface includes: determining the geometric error between the NURBS parametric surface and the closed manifold triangular mesh surface, and generating a corresponding error distribution chromatogram; based on the error distribution chromatogram, identifying local regions where the fitting error exceeds a preset engineering tolerance as target optimization regions, and obtaining a quality assessment result containing the identification information of the target optimization regions; wherein, the target optimization regions include at least one of edge regions where the curvature value exceeds a preset curvature threshold and splicing areas where multiple surfaces intersect.

[0061] For example, the spatial distance between the NURBS parametric surface and the closed manifold triangular mesh surface generated in step S202 is calculated, and the geometric error between the two is determined based on the spatial distance. This geometric error data is then visualized as a distribution chromatogram, where different colors represent different error levels, thus visually displaying the distribution of fitting accuracy in different regions of the surface. Accordingly, based on the error distribution chromatogram, an automated region identification algorithm is used to detect local regions where the fitting error exceeds the preset engineering tolerance, and these regions are marked as target optimization regions.

[0062] When generating the distribution chromatogram, the geometric error data between the NURBS surface and the triangular mesh surface is used to map the error values ​​into visual colors through linear or nonlinear color mapping rules. The specific error threshold range and color mapping scheme can be configured according to the engineering accuracy requirements and are not limited here.

[0063] It should be noted that engineering tolerance refers to the allowable range of precision set according to automotive industry standards. Different tolerance grades are usually adopted according to the functional requirements of different areas, that is, areas with different tolerance grades correspond to different engineering tolerances. Error distribution chromatogram is a technical means of converting numerical error data into a visual image through color mapping, which facilitates the rapid location of problem areas. Edge areas with curvature values ​​exceeding the preset curvature threshold usually correspond to abrupt curvature changes such as body feature lines and rib lines. The splicing area where multiple curved surfaces intersect is prone to the problem of decreased fitting accuracy due to the complex topological connection relationships involved.

[0064] Correspondingly, the output includes quality assessment results containing the identification information of the target optimization regions. This result records the spatial location and error characteristics of all target regions requiring optimization in the form of a data structure. Furthermore, when a target region with a fitting error exceeding the preset engineering tolerance is identified in the quality assessment results, an adaptive iterative optimization process is automatically triggered for these target regions. This optimization process employs a local refinement strategy, specifically including: within the identified target optimization regions, automatically adjusting the parameter configuration of the optimization algorithm and implementing targeted accuracy improvement measures. These measures are adaptively selected based on error distribution characteristics and regional geometric properties, aiming to achieve accuracy improvement with optimal computational cost.

[0065] It should be noted that adaptive iterative optimization refers to an intelligent processing mechanism in which the system automatically adjusts the optimization strategy based on the quality assessment results. This mechanism achieves cyclical improvement of the surface quality by establishing a closed-loop control process of "assessment-optimization-verification". In each iteration, the error distribution in the current optimized state is recalculated, and the optimization strategy is dynamically updated until the termination condition is met. The termination condition for the optimization process is set as follows: the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance requirements. This condition ensures that the final output surface model meets the precision standards of the automotive industry in all regions. The system automatically determines whether the optimization target has been achieved by monitoring the global error data in real time, thereby realizing intelligent control of the entire quality optimization process.

[0066] The automotive surface fitting method based on deep implicit functions and physical constraints provided in this application achieves end-to-end automatic fitting from raw 3D point cloud data to a high-quality parametric CAD model through the collaborative work of a deep implicit function network and a physically constrained NURBS transformation network. This effectively solves the problems of traditional solutions relying on human experience and having low automation levels. Specifically, the deep implicit function network can robustly reconstruct a complete watertight mesh from noisy and incomplete point clouds, significantly improving the ability to handle complex topologies and data defects. Meanwhile, the physically constrained NURBS transformation network, by fusing data fitting loss, smoothness loss, and continuity loss into a physical constraint loss function, ensures that the generated NURBS parametric surface maintains both geometric accuracy and meets engineering physical constraints, effectively overcoming the physical unreliability of output results from traditional pure data-driven solutions. Furthermore, the adaptive iterative optimization mechanism achieves intelligent control of fitting accuracy through automatic identification and local refinement of error-prone areas, thereby outputting an industry-standard NURBS surface that can be directly used in CAE or CAM processes, significantly improving the efficiency, quality, and automation level of automotive surface design.

[0067] In some embodiments, the deep implicit function network includes a point cloud feature encoder and an implicit function decoder. Inputting 3D point cloud data into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface includes: extracting global shape feature vectors and local geometric features from the 3D point cloud data using the point cloud feature encoder; constructing a 3D query mesh within the spatial bounding box of the 3D point cloud data; inputting the global shape feature vectors and the spatial coordinates of each vertex in the 3D query mesh into the implicit function decoder to obtain the geometric state scalars corresponding to each vertex; wherein the geometric state scalars include signed distance values, unsigned distance values, or occupancy probability values; and using the moving cube algorithm to extract zero isosurfaces based on the geometric state scalars corresponding to each vertex of the 3D query mesh to generate a closed manifold triangular mesh surface.

[0068] The point cloud feature encoder employs a point cloud neural network such as PointNet++ or Dynamic Graph Convolutional Neural Network (DGCNN) to extract global shape features and local geometric features from the 3D point cloud data. The implicit function decoder defines an implicit function that, for a given 3D spatial coordinate (x, y, z) and its corresponding features, outputs a scalar value s. This implicit function decoder can be configured to different operating modes: in the Signed Distance Function (SDF) mode, the scalar s output by the implicit function decoder represents the shortest distance from the point to the target surface, with the sign indicating the inside-outside relationship; in the Occupancy Network mode, the scalar s output by the implicit function decoder is a probability between 0 and 1, representing the probability that the point is occupied by the surface. It should be noted that this deep implicit function network is trained under supervision by using a large number of automotive surface point clouds and their corresponding complete surface models. The loss functions used in the training process include reconstruction accuracy loss (such as L1 or L2 loss) and implicit smoothness constraints. The fully trained network can adaptively fit complete, smooth and engineering-quality-compliant triangular mesh surfaces from "imperfect" point clouds containing noise and missing data, demonstrating strong robustness and generalization ability.

[0069] Accordingly, the original 3D point cloud, for example The input is fed into a deep implicit function network, and the point cloud P is processed by the point cloud feature encoder in the deep implicit function network, where, Let N represent the i-th point in the point cloud set P, and let N represent the total number of points in the point cloud set P. The encoder extracts the global shape feature vector of the point cloud. and the local geometric features of each point The implicit function lookup section defines an implicit function decoder. (Typically a multilayer perceptron (MLP)); for any query point q=(x, y, z) in space, the output of this function is the SDF, i.e. , where s is the signed distance from point q to the target surface; or the output is an occupancy function, i.e. ,in , where is the probability that point q is occupied by the surface. Further, a dense 3D mesh is constructed within the point cloud bounding box, and a trained implicit function decoder is used. Predict the geometric state scalars such as SDF values ​​or occupancy probabilities for each mesh vertex; use the moving cube algorithm to extract zero isosurfaces from these scalar fields, thereby generating a closed manifold triangular mesh surface S'. This stage, through deep learning, learns to "imagine" and "complete" the missing geometric parts, is insensitive to noise, and can adaptively handle various complex topologies.

[0070] This application employs a deep implicit function network (LAN) that includes a point cloud feature encoder and an implicit function decoder. The LNA extracts global and local geometric features, enabling the network to deeply understand the structural priors and detailed features of the vehicle surface, laying the foundation for subsequent accurate reconstruction. By constructing a query mesh within a spatial bounding box and using the LNA decoder to predict geometric state scalars, a unified and continuous mathematical representation of arbitrarily complex topological surfaces is achieved, effectively overcoming the dependence of traditional schemes on surface segmentation and stitching. Furthermore, the extraction of zero isosurfaces from the geometric state scalar field based on the moving cube algorithm ensures the topological correctness and closure of the generated mesh, effectively reducing holes and non-manifold structures. Finally, the overall process transforms surface fitting into an implicit function optimization and query problem through deep learning, significantly improving robustness to actual scanning defects such as noise, missing data, and uneven density, achieving efficient and high-quality conversion from "imperfect" point clouds to closed manifold triangular mesh surfaces that can be directly used in subsequent CAD processes.

[0071] In some embodiments, a closed manifold triangular mesh surface is input into a physically constrained NURBS transformation network. The NURBS surface parameters are iteratively optimized by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints. This includes: extracting the geometric features of the closed manifold triangular mesh surface, including boundary curves, feature lines, and curvature information. The boundary curves include the outer contour and internal opening boundaries of the closed manifold triangular mesh surface. The feature lines are used to identify areas where the closed manifold triangular mesh surface undergoes drastic geometric changes. The curvature information includes principal curvature and Gaussian curvature. The closed manifold triangular mesh surface and the extracted geometric features are input into a preset parameter prediction network. The parameter prediction network predicts the initial values ​​of the NURBS surface parameters, including the positions, weights, and node vector distributions of the control point meshes. Based on the predicted initial values ​​of the NURBS surface parameters, a differentiable physical constraint loss function is constructed. By minimizing the physical constraint loss function, the control point meshes, weights, and node vectors are iteratively optimized to generate a NURBS parametric surface that conforms to engineering constraints.

[0072] For example, the geometric features of the manifold triangular mesh surface are obtained by analyzing the adjacency relationships of the edges of the closed manifold triangular mesh surface, extracting boundary curves, extracting feature lines based on discrete curvature estimation and edge detection algorithms, and extracting curvature information using a discrete differential geometric operator based on the triangular mesh. Correspondingly, the closed manifold triangular mesh surface and its extracted geometric features are input into a pre-defined parameter prediction network. This parameter prediction network, for example, is built based on a convolutional neural network or graph neural network architecture and can learn the complex mapping relationship from mesh features to NURBS surface parameters. The network outputs initial predicted values ​​of the NURBS surface parameters, including the spatial location of control point meshes, the weight values ​​of each control point, and the distribution of node vectors. Further, based on the initial values ​​of the NURBS surface parameters output by the parameter prediction network, a differentiable physical constraint loss function is constructed. This loss function is a weighted combination of multiple constraint terms with explicit engineering significance. Among them, the data fitting loss ensures that the generated NURBS surface and the input triangular mesh surface are highly consistent in geometry, minimizing the spatial distance between them; the smoothness loss is based on the principle of curvature energy minimization (such as thin plate energy), which suppresses unnecessary fluctuations and wrinkles in the surface, so that the generated NURBS surface meets the stringent requirements of automotive aerodynamics and aesthetic appearance compliance; the continuity loss: when the surface is composed of multiple NURBS surface patches, it forces adjacent surface patches to meet the specified geometric continuity conditions (such as G1 or G2 continuity) at the splicing boundary, ensuring a smooth transition of the entire surface.

[0073] Accordingly, by minimizing the physical constraint loss function, the parameters of the NURBS surface, such as the control point grid, weights, and node vectors, are iteratively optimized. This optimization process is carried out within a differentiable framework, which can automatically calculate gradients and update parameters, ultimately generating a NURBS parametric surface that conforms to engineering constraints. By quantifying these engineering constraints in the form of a loss function and embedding them into the optimization process, the output results are ensured to have both mathematical accuracy and engineering practicality.

[0074] In some embodiments, the physical constraint loss function satisfies the following formula:

[0075]

[0076] Where L_data is the data fitting loss, used to minimize the geometric distance between the NURBS surface and the closed manifold triangular mesh surface; L_fairness is the smoothness loss; and L_continuity is the continuity loss, used to force adjacent NURBS surface patches to achieve the target geometric continuity at the splicing boundary. , and These are the weighting coefficients for the corresponding losses.

[0077] For example, the loss of smoothness Energy minimization based on curvature, such as the energy of a thin plate, aims to make the surface as smooth as possible to reduce unnecessary fluctuations. Its formula can generally be approximated as: ,in Principal curvature; if the surface is composed of multiple pieces, continuity loss. The specified target geometric continuity must be satisfied at the forced splicing boundary. The target geometric continuity includes, but is not limited to, G1 continuity or G2 continuity.

[0078] It should be noted that the weighting coefficients , and These coefficients are used to balance the relative importance of data fitting accuracy, smoothness requirements, and continuity constraints. The specific values ​​of these coefficients can be determined through grid search, empirical assignment, or priority analysis based on task requirements. Their numerical configuration directly affects the trade-off between geometric fidelity and physical plausibility in the final generated NURBS surface. Those skilled in the art can adaptively adjust the above weighting coefficients through conventional optimization methods or experimental verification according to different emphases on surface accuracy, smoothness, and continuity in specific application scenarios. Here, the weighting coefficients... , and The value of is not limited.

[0079] In this embodiment, by automatically extracting geometric features such as boundary curves, feature lines, and curvature information, the topological structure and geometric characteristics of the surface can be understood, providing sufficient semantic information support for parametric transformation. By using a parameter prediction network to generate initial values ​​for NURBS parameters, the limitations of relying on manual experience to set initial parameters in traditional methods are effectively overcome, significantly improving processing efficiency and ensuring the convergence of the optimization process. By constructing a physical constraint loss function that includes data fitting, smoothness, and continuity losses, key quality requirements in engineering practice are transformed into quantifiable optimization objectives, ensuring that the generated NURBS surface simultaneously meets the requirements of geometric accuracy, smoothness, and continuity. In addition, through iterative optimization of the control point mesh, weights, and node vectors using a differentiable optimization framework, the automated fine adjustment of the parametric surface is realized, so that the output results maintain geometric consistency with the original mesh and meet the strict standards of the automotive industry for Class A surfaces, thereby achieving seamless integration between deep learning models and industrial CAD processes.

[0080] In some embodiments, triggering an adaptive iterative optimization process for the target region includes: performing an adaptive sampling operation within the corresponding target optimization region based on the target optimization region identification information in the quality assessment results to obtain encrypted sampling information. The adaptive sampling operation includes encrypting sampling points and / or increasing the control point grid density of the NURBS parametric surface within the parameter domain; feeding the encrypted sampling information back to the NURBS transformation network and performing an iterative optimization step of minimizing the physical constraint loss function on the NURBS surface parameters until the fitting error of all regions of the NURBS parametric surface is lower than the engineering tolerance.

[0081] For example, based on the target optimization region identification information recorded in the quality assessment results, the surface region requiring optimization is precisely located. For each target optimization region, an adaptive sampling operation is performed, specifically including: within the parameter domain of the NURBS parametric surface, increasing the sampling points in the identified high-error regions to obtain more accurate geometric error information by increasing the sampling density within the parameter domain; or, adaptively increasing the control point grid density of the NURBS parametric surface according to the error distribution characteristics, providing greater freedom for local surface optimization. These two sampling strategies can be used individually or in combination to ensure that the system can select the optimal optimization resource configuration scheme based on specific error characteristics. It should be noted that the dynamic adjustment rules for the sampling density and control point grid density within the parameter domain are automatically determined based on the error distribution characteristics in the quality assessment results. Specifically, the required sampling point densification ratio and control point enhancement scheme can be adaptively determined according to the geometric characteristics of the target optimization region (such as curvature distribution and error gradient) and the preset engineering accuracy requirements. Those skilled in the art can implement the above-mentioned density adjustment strategy by setting a threshold rule or constructing an adaptive adjustment function according to the accuracy requirements and computing resource constraints in specific applications. All specific implementations of such intelligent optimization mechanisms based on error feedback should be considered within the protection scope of this invention.

[0082] Accordingly, the encrypted sampling information is fed back to the physically constrained NURBS transformation network. Based on this new sampling data, the iterative optimization of the NURBS surface parameters by minimizing the physical constraint loss function is re-executed within the target optimization region. It should be noted that this optimization process is local; parameters are adjusted only for the identified problem areas, while other compliant surface areas remain unchanged, thus significantly improving optimization efficiency. Specifically, this adaptive iterative optimization process employs a closed-loop control mechanism. After each local optimization, a quality assessment is performed to check whether the fitting error of the target region has decreased to within the corresponding engineering tolerance. This "assessment-optimization-verification" cycle continues until the fitting error of all regions of the NURBS parametric surface, including the initially identified high-curvature edges and complex transition areas, meets the corresponding engineering tolerance requirements.

[0083] In this embodiment, the positioning mechanism based on quality assessment results ensures the precise allocation of optimization resources, focusing on only substandard local areas, reducing the waste of computational resources caused by global re-optimization, and significantly improving optimization efficiency. Secondly, through an adaptive sampling strategy that densifies sampling points in the parameter domain and increases the grid density of control points, the optimization granularity can be dynamically adjusted according to the error distribution characteristics, providing stronger detail depiction capabilities for high-error areas and effectively solving the problem of insufficient fitting accuracy of traditional methods at complex geometric features such as edges and transition areas. Furthermore, by feeding back the densified sampling information to the NURBS conversion network to form a closed-loop optimization, the system has the ability to self-diagnose and refine, continuously approaching the optimal solution through multiple "evaluation-optimization" cycles. In addition, this mechanism ensures that all areas of the output surface meet strict engineering tolerance requirements, taking into account local detail quality while ensuring overall accuracy, greatly reducing the cost of manual model repair in the later stage, and providing reliable quality assurance for achieving end-to-end automation from scanned data to production-grade CAD models.

[0084] Figure 3 Another flowchart illustrating the vehicle surface fitting method based on deep implicit functions and physical constraints, provided as an exemplary embodiment of this application. Figure 3 As shown, the method for fitting car surfaces based on deep implicit functions and physical constraints includes the following steps:

[0085] S301. Construct a large-scale training dataset based on a 3D point cloud dataset containing various typical automotive surface shapes.

[0086] S302. Based on a large-scale training dataset, train the deep implicit function network to obtain a trained deep implicit function network.

[0087] S303. Obtain the 3D point cloud data of the target vehicle;

[0088] S304. Input the 3D point cloud data into the trained deep implicit function network to generate a closed manifold triangular mesh surface.

[0089] Specifically, a point cloud feature encoder extracts global shape feature vectors and local geometric features from 3D point cloud data; a 3D query mesh is constructed within the spatial bounding box of the 3D point cloud data; the global shape feature vectors and the spatial coordinates of each vertex in the 3D query mesh are input into an implicit function decoder to obtain the geometric state scalars corresponding to each vertex; wherein, the geometric state scalars include signed distance values, unsigned distance values, or occupancy probability values; based on the geometric state scalars corresponding to each vertex of the 3D query mesh, a moving cube algorithm is used to extract zero isosurfaces to generate closed manifold triangular mesh surfaces.

[0090] S305. Extract the geometric features of the closed manifold triangular mesh surface. The geometric features include boundary curves, feature lines, and curvature information.

[0091] S306. Input the closed manifold triangular mesh surface and the extracted geometric features into the preset parameter prediction network. The parameter prediction network predicts the initial values ​​of the NURBS surface parameters, which include the position, weight, and distribution of the node vectors of the control point mesh.

[0092] The parameter prediction network is constructed, for example, based on a convolutional neural network or a graph neural network architecture.

[0093] S307. Based on the initial values ​​of the predicted NURBS surface parameters, construct a differentiable physical constraint loss function.

[0094] For example, the physical constraint loss function satisfies the following formula:

[0095]

[0096] Where L_data is the data fitting loss, used to minimize the geometric distance between the NURBS surface and the closed manifold triangular mesh surface; L_fairness is the smoothness loss; and L_continuity is the continuity loss, used to force adjacent NURBS surface patches to achieve the target geometric continuity at the splicing boundary. , and These are the weighting coefficients for the corresponding losses.

[0097] S308. By minimizing the physical constraint loss function, the control point mesh, weights, and node vectors are iteratively optimized to generate a NURBS parametric surface that conforms to engineering constraints.

[0098] S309. Perform a quality assessment on the NURBS parametric surface and obtain the quality assessment results.

[0099] S310. Determine whether there is a target area in the quality assessment results where the fitting error exceeds the preset engineering tolerance.

[0100] If so, execute S311;

[0101] If not, proceed with S312.

[0102] S311. Trigger an adaptive iterative optimization process for the target region until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0103] Specifically, based on the target optimization region identification information in the quality assessment results, an adaptive sampling operation is performed within the corresponding target optimization region to obtain encrypted sampling information. The adaptive sampling operation includes decrypting sampling points and / or increasing the control point grid density of the NURBS parametric surface within the parameter domain. The encrypted sampling information is then fed back to the NURBS transformation network to perform iterative optimization of the NURBS surface parameters by minimizing the physical constraint loss function, until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0104] S312, Output the CAD model of the NURBS parametric surface.

[0105] In summary, this application has at least the following advantages:

[0106] I. By collaborating with a deep implicit function network and a physically constrained NURBS transformation network, end-to-end automatic fitting from raw 3D point cloud data to high-quality parametric CAD models is achieved, effectively solving the problems of traditional solutions relying on manual experience and having low automation levels. Specifically, the deep implicit function network can robustly reconstruct complete watertight meshes from noisy and incomplete point clouds, significantly improving the ability to handle complex topologies and data defects. Meanwhile, the physically constrained NURBS transformation network, through a physical constraint loss function that integrates data fitting loss, smoothness loss, and continuity loss, ensures that the generated NURBS parametric surfaces maintain geometric accuracy while satisfying engineering physical constraints, effectively overcoming the physical unreliability of output results from traditional pure data-driven solutions. Furthermore, the adaptive iterative optimization mechanism achieves intelligent control of fitting accuracy through automatic identification and local refinement of error-prone areas, thereby outputting industry-standard NURBS surfaces that can be directly used in CAE or CAM processes, significantly improving the efficiency, quality, and automation level of automotive surface design.

[0107] Second, by employing a deep implicit function network that includes a point cloud feature encoder and an implicit function decoder, the network extracts global and local geometric features using the point cloud feature encoder, enabling it to deeply understand the structural priors and detailed features of the automotive surface, laying the foundation for subsequent accurate reconstruction. By constructing a query mesh within a spatial bounding box and using the implicit function decoder to predict geometric state scalars, a unified and continuous mathematical representation of arbitrarily complex topological surfaces is achieved, effectively overcoming the dependence of traditional schemes on surface segmentation and stitching. In addition, the zero isosurface extracted from the geometric state scalar field based on the moving cube algorithm ensures the topological correctness and closure of the generated mesh, effectively reducing holes and non-manifold structures. Furthermore, the overall process transforms surface fitting into an implicit function optimization and query problem through deep learning, significantly improving robustness to actual scanning defects such as noise, missing data, and density inhomogeneity, achieving efficient and high-quality conversion from "imperfect" point clouds to closed manifold triangular mesh surfaces that can be directly used in subsequent CAD processes.

[0108] Third, by automatically extracting geometric features such as boundary curves, feature lines, and curvature information, the topological structure and geometric properties of the surface can be understood, providing sufficient semantic information support for parametric transformation. By using a parameter prediction network to generate initial values ​​for NURBS parameters, the limitations of relying on manual experience to set initial parameters in traditional methods are effectively overcome, significantly improving processing efficiency and ensuring the convergence of the optimization process. Furthermore, by constructing a physical constraint loss function that includes data fitting, smoothness, and continuity losses, key quality requirements in engineering practice are transformed into quantifiable optimization objectives, ensuring that the generated NURBS surface simultaneously meets the requirements of geometric accuracy, smoothness, and continuity. In addition, through iterative optimization of the control point mesh, weights, and node vectors using a differentiable optimization framework, the automated fine adjustment of the parametric surface is achieved, ensuring that the output results maintain geometric consistency with the original mesh and meet the strict standards of the automotive industry for Class A surfaces, thus achieving seamless integration between the deep learning model and the industrial CAD process.

[0109] Fourth, the positioning mechanism based on quality assessment results ensures the precise allocation of optimization resources, focusing on only substandard local areas and reducing the waste of computational resources caused by global re-optimization, thus significantly improving optimization efficiency. Secondly, the adaptive sampling strategy, which involves increasing the density of sampling points in the parameter domain and the grid density of control points, enables dynamic adjustment of the optimization granularity based on error distribution characteristics. This provides stronger detail representation capabilities for high-error areas, effectively solving the problem of insufficient fitting accuracy in complex geometric features such as edges and transition zones using traditional methods. Furthermore, by feeding back the encrypted sampling information to the NURBS transformation network to form a closed-loop optimization, the system possesses self-diagnosis and refinement capabilities, continuously approaching the optimal solution through multiple "evaluation-optimization" cycles. In addition, this mechanism ensures that all areas of the output surface meet strict engineering tolerance requirements, maintaining overall accuracy while also considering local detail quality, significantly reducing the cost of manual model repair in the later stages, and providing reliable quality assurance for end-to-end automation from scanned data to production-grade CAD models.

[0110] The following are embodiments of the apparatus described in this application, which can be used to execute the embodiments of the method described in this application. For details not disclosed in the apparatus embodiments of this application, please refer to the embodiments of the method described in this application.

[0111] Figure 4 This is a schematic diagram of a vehicle surface fitting device based on deep implicit functions and physical constraints, provided as an exemplary embodiment of this application. Figure 4 As shown, the vehicle surface fitting device 40 based on deep implicit functions and physical constraints includes an acquisition module 41, a first processing module 42, a second processing module 43, and a quality assessment module 44, wherein:

[0112] The acquisition module 41 is used to acquire the three-dimensional point cloud data of the target vehicle;

[0113] The first processing module 42 is used to input three-dimensional point cloud data into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface.

[0114] The second processing module 43 is used to input the closed manifold triangular mesh surface into the physically constrained NURBS transformation network, and iteratively optimize the NURBS surface parameters by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints; wherein, the physical constraint loss function includes data fitting loss, smoothness loss and continuity loss;

[0115] The quality assessment module 44 is used to assess the quality of the NURBS parametric surface. When a target area with a fitting error exceeding the preset engineering tolerance is identified, an adaptive iterative optimization process for the target area is triggered until the fitting error of all areas of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0116] In one possible implementation, the deep implicit function network includes a point cloud feature encoder and an implicit function decoder; the first processing module 42 can be specifically used to: extract global shape feature vectors and local geometric features of the 3D point cloud data through the point cloud feature encoder; construct a 3D query mesh within the spatial bounding box of the 3D point cloud data; input the global shape feature vectors and the spatial coordinates of each vertex in the 3D query mesh to the implicit function decoder to obtain the geometric state scalars corresponding to each vertex; wherein, the geometric state scalars include signed distance values, unsigned distance values, or occupancy probability values; based on the geometric state scalars corresponding to each vertex of the 3D query mesh, the moving cube algorithm is used to extract the zero isosurface to generate a closed manifold triangular mesh surface.

[0117] In one possible implementation, the second processing module 43 can be specifically used to: extract the geometric features of a closed manifold triangular mesh surface, the geometric features including boundary curves, feature lines, and curvature information, wherein the boundary curves include the outer contour and internal opening boundaries of the closed manifold triangular mesh surface, the feature lines are used to identify the regions of drastic geometric changes in the closed manifold triangular mesh surface, and the curvature information includes principal curvature and Gaussian curvature; input the closed manifold triangular mesh surface and the extracted geometric features into a preset parameter prediction network, and predict the initial values ​​of NURBS surface parameters through the parameter prediction network, the NURBS surface parameters including the position, weight, and node vector distribution of the control point mesh; construct a differentiable physical constraint loss function based on the predicted initial values ​​of the NURBS surface parameters; and iteratively optimize the control point mesh, weight, and node vector by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints.

[0118] In one possible implementation, the physical constraint loss function satisfies the following formula:

[0119]

[0120] Where L_data is the data fitting loss, used to minimize the geometric distance between the NURBS surface and the closed manifold triangular mesh surface; L_fairness is the smoothness loss; and L_continuity is the continuity loss, used to force adjacent NURBS surface patches to achieve the target geometric continuity at the splicing boundary. , and These are the weighting coefficients for the corresponding losses.

[0121] In one possible implementation, the quality assessment module 44 can be specifically used to: determine the geometric error between the NURBS parametric surface and the closed manifold triangular mesh surface, and generate a corresponding error distribution chromatogram; based on the error distribution chromatogram, identify local areas where the fitting error exceeds a preset engineering tolerance as target optimization areas, and obtain a quality assessment result containing the identification information of the target optimization areas; wherein, the target optimization areas include at least one of the edge regions where the curvature value exceeds a preset curvature threshold and the splicing area where multiple surfaces intersect.

[0122] In one possible implementation, the quality assessment module 44 can also be used to: perform an adaptive sampling operation within the corresponding target optimization region based on the target optimization region identification information in the quality assessment result, to obtain encrypted sampling information. The adaptive sampling operation includes encrypting sampling points and / or increasing the control point grid density of the NURBS parametric surface within the parameter domain; feeding the encrypted sampling information back to the NURBS transformation network, and performing an iterative optimization step of minimizing the physical constraint loss function on the NURBS surface parameters until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

[0123] The vehicle surface fitting device based on deep implicit functions and physical constraints provided in this application embodiment can execute the technical solution shown in the above-described vehicle surface fitting method embodiment based on deep implicit functions and physical constraints. Its implementation principle and beneficial effects are similar, and will not be repeated here.

[0124] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to this application.

[0125] It should be further noted that although the steps in the flowchart are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowchart may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0126] It should be noted that the above-described device embodiments are merely illustrative, and the device of this application can be implemented in other ways. Furthermore, it should be understood that the division of the various modules in the above device is only a logical functional division; in actual implementation, they can be fully or partially integrated into a single physical entity, or they can be physically separated. These modules can all be implemented in software via processing element calls; they can all be implemented in hardware; or some modules can be implemented by processing element calls to software, while others are implemented in hardware. For example, the quality assessment module can be a separately established processing element, or it can be integrated into a chip in the above device. Alternatively, it can be stored as program code in the memory of the above device, and its function can be called and executed by a processing element of the above device. The implementation of other modules is similar. In addition, these modules can be fully or partially integrated together, or they can be implemented independently. The processing element here can be an integrated circuit with signal processing capabilities. During implementation, each step of the above method or each of the above modules can be completed through integrated logic circuits in the hardware of the processor element or through software instructions.

[0127] For example, these modules can be one or more integrated circuits configured to implement the above methods, such as one or more Application Specific Integrated Circuits (ASICs), one or more Digital Signal Processors (DSPs), or one or more Field Programmable Gate Arrays (FPGAs). As another example, when a module is implemented using processing element scheduler code, the processing element can be a general-purpose processor, such as a Central Processing Unit (CPU) or other processor capable of calling program code. Furthermore, these modules can be integrated together as a System-On-a-Chip (SOC).

[0128] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., Digital Video Discs, DVDs), or semiconductor media (e.g., solid-state disks (SSDs)).

[0129] Figure 5 A schematic diagram of the structure of an electronic device provided as an exemplary embodiment of this application. For example... Figure 5 As shown, the electronic device 50 in this embodiment includes:

[0130] At least one processor 51; and a memory 52 communicatively connected to the at least one processor;

[0131] The memory 52 stores instructions that can be executed by at least one processor 51 to cause the electronic device to perform the method as described in any of the above embodiments.

[0132] Alternatively, the memory 52 can be either standalone or integrated with the processor 51.

[0133] The memory 52 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0134] The processor 51 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application. Specifically, when implementing the automotive surface fitting method based on deep implicit functions and physical constraints described in the foregoing method embodiments, the electronic device may be, for example, an electronic device with processing capabilities such as a server.

[0135] Optionally, the electronic device may also include a communication interface 53. In specific implementations, if the communication interface 53, memory 52, and processor 51 are implemented independently, they can be interconnected via a bus to complete communication. 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., but this does not imply that there is only one bus or one type of bus.

[0136] Optionally, in a specific implementation, if the communication interface 53, memory 52 and processor 51 are integrated on a single chip, then the communication interface 53, memory 52 and processor 51 can communicate through an internal interface.

[0137] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0138] This application also provides a computer-readable storage medium storing computer-executable instructions. When the computer-executable instructions are executed, they are used to implement the method steps as described in the above method embodiments. The specific implementation methods and technical effects are similar and will not be repeated here.

[0139] The aforementioned computer-readable storage media can be implemented from 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.

[0140] 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 a vehicle surface fitting device based on deep implicit functions and physical constraints.

[0141] This application also provides a computer program product, including a computer program, which, when executed, implements the method steps as described in the above method embodiments. The specific implementation and technical effects are similar and will not be repeated here.

[0142] In the above embodiments, the descriptions of each embodiment have their own emphasis. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments. The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification.

[0143] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of this application are indicated by the following claims.

[0144] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for fitting automotive surfaces based on deep implicit functions and physical constraints, characterized in that, include: Obtain the 3D point cloud data of the target vehicle; The three-dimensional point cloud data is input into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface; The closed manifold triangular mesh surface is input into a physically constrained non-uniform rational B-spline NURBS transformation network. The NURBS surface parameters are iteratively optimized by minimizing the physical constraint loss function to generate a NURBS parametric surface that meets engineering constraints. The physical constraint loss function includes data fitting loss, smoothness loss, and continuity loss. The quality of the NURBS parametric surface is evaluated. When a target region with a fitting error exceeding the preset engineering tolerance is identified, an adaptive iterative optimization process is triggered for the target region until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

2. The method for fitting automotive surfaces based on deep implicit functions and physical constraints according to claim 1, characterized in that, The deep implicit function network includes a point cloud feature encoder and an implicit function decoder; the step of inputting the 3D point cloud data into the pre-trained deep implicit function network to generate a closed manifold triangular mesh surface includes: The global shape feature vector and local geometric features of the 3D point cloud data are extracted using the point cloud feature encoder. Construct a 3D query grid within the spatial bounding box of the 3D point cloud data; The global shape feature vector and the spatial coordinates of each vertex in the 3D query mesh are input to the implicit function decoder to obtain the geometric state scalar corresponding to each vertex; wherein, the geometric state scalar includes a signed distance value, an unsigned distance value, or an occupancy probability value; Based on the geometric state scalars corresponding to each vertex of the three-dimensional query mesh, the zero isosurface is extracted using the moving cube algorithm to generate the closed manifold triangular mesh surface.

3. The method for fitting automotive surfaces based on deep implicit functions and physical constraints according to claim 1, characterized in that, The process of inputting the closed manifold triangular mesh surface into a physically constrained non-uniform rational B-spline NURBS transformation network, and iteratively optimizing the NURBS surface parameters by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints includes: Extract the geometric features of the closed manifold triangular mesh surface. The geometric features include boundary curves, feature lines, and curvature information. The boundary curves include the outer contour and internal opening boundary of the closed manifold triangular mesh surface. The feature lines are used to identify regions where the closed manifold triangular mesh surface undergoes drastic geometric changes. The curvature information includes principal curvature and Gaussian curvature. The closed manifold triangular mesh surface and the extracted geometric features are input into a preset parameter prediction network. The parameter prediction network predicts the initial values ​​of the NURBS surface parameters, which include the position, weight, and distribution of the node vectors of the control point mesh. Based on the initial values ​​of the predicted NURBS surface parameters, a differentiable physical constraint loss function is constructed. By minimizing the physical constraint loss function, the control point mesh, the weights, and the node vectors are iteratively optimized to generate a NURBS parametric surface that conforms to engineering constraints.

4. The method for fitting automotive surfaces based on deep implicit functions and physical constraints according to claim 3, characterized in that, The physical constraint loss function satisfies the following formula: Wherein, L_data is the data fitting loss, used to minimize the geometric distance between the NURBS surface and the closed manifold triangular mesh surface; L_fairness is the smoothness loss; L_continuity is the continuity loss, used to force adjacent NURBS surface patches to achieve the target geometric continuity at the splicing boundary; , and These are the weighting coefficients for the corresponding losses.

5. The method for fitting automotive surfaces based on deep implicit functions and physical constraints according to any one of claims 1 to 4, characterized in that, The quality assessment of the NURBS parametric surface includes: Determine the geometric error between the NURBS parametric surface and the closed manifold triangular mesh surface, and generate the corresponding error distribution chromatogram; Based on the error distribution chromatogram, local regions where the fitting error exceeds the preset engineering tolerance are identified as target optimization regions, and a quality assessment result containing the identification information of the target optimization regions is obtained. The target optimization region includes at least one of the following: an edge region with a curvature value exceeding a preset curvature threshold and a splicing area where multiple curved surfaces intersect.

6. The method for fitting automotive surfaces based on deep implicit functions and physical constraints according to claim 5, characterized in that, The process of triggering adaptive iterative optimization for the target region includes: Based on the target optimization region identification information in the quality assessment results, an adaptive sampling operation is performed within the corresponding target optimization region to obtain encrypted sampling information. The adaptive sampling operation includes encrypting sampling points within the parameter domain of the NURBS parametric surface and / or increasing the control point grid density of the NURBS parametric surface. The encrypted sampling information is fed back to the NURBS transformation network, and the step of iteratively optimizing the NURBS surface parameters by minimizing the physical constraint loss function is performed until the fitting error of all regions of the NURBS parameterized surface is lower than the corresponding engineering tolerance.

7. A vehicle surface fitting device based on deep implicit functions and physical constraints, characterized in that, include: The acquisition module is used to acquire the 3D point cloud data of the target vehicle. The first processing module is used to input the three-dimensional point cloud data into a pre-trained deep implicit function network to generate a closed manifold triangular mesh surface. The second processing module is used to input the closed manifold triangular mesh surface into a physically constrained non-uniform rational B-spline NURBS transformation network, and iteratively optimize the NURBS surface parameters by minimizing the physical constraint loss function to generate a NURBS parametric surface that conforms to engineering constraints; wherein, the physical constraint loss function includes data fitting loss, smoothness loss and continuity loss; The quality assessment module is used to assess the quality of the NURBS parametric surface. When a target region with a fitting error exceeding the preset engineering tolerance is identified, an adaptive iterative optimization process is triggered for the target region until the fitting error of all regions of the NURBS parametric surface is lower than the corresponding engineering tolerance.

8. An electronic device, characterized in that, include: A processor, and a memory communicatively connected to the processor; The memory is used to store computer-executed instructions; The processor is configured to execute the computer execution instructions to implement the method as described in any one of claims 1 to 6.

9. 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 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed, it implements the method as described in claim 6.

Images