Multi-source data fusion three-dimensional modeling system and method

Abstract

The application relates to the technical field of three-dimensional digital geometry reconstruction, and discloses a multi-source data fusion three-dimensional modeling system and method, which comprises a nonlinear geometric energy deformation control unit which constructs a grid energy functional, calculates a grid length distortion rate before updating a grid vertex coordinate, controls the grid vertex to turn into an isotropic relaxation mode when the grid length distortion rate exceeds a safety boundary, cuts off a data geometry approximation term and calculates a first-order neighborhood arithmetic average center point coordinate as a temporary updating position, cuts back to a data-driven mode after the grid length distortion rate returns, and recombines connection relationships in cooperation with a manifold grid topology correction unit to complete self-inspection correction. The application converts geometric residual errors of heterogeneous sources into endogenous energy constraints of deformation evolution, in-situ eliminates local shear distortion, and eliminates surface shadow and topological tearing defects.

Description

Technical Field

[0001] This invention relates to a multi-source data fusion 3D modeling system and method, belonging to the field of 3D digital geometric reconstruction technology. Background Technology

[0002] In current 3D digital geometric reconstruction, a common practice is to fuse multi-source point cloud data by minimizing the grid energy functional. The reconstruction system usually sets data terms and spatial smoothing terms in the grid deformation model to make the grid surface gradually approximate the surface obtained by surveying, while trying to maintain the integrity of the topological manifold structure and the continuity of the surface of the geometric model.

[0003] When performing 3D reconstruction of non-standard, irregularly shaped equipment, the ranging principles and noise distributions of different surveying instruments are not consistent. After registration of multi-source point cloud data, geometric errors may still exist at the seams where different data points intersect. When processing such data in the mesh deformation calculation process, the traction direction of mesh vertices by different sensor data may be inconsistent, easily causing the displacement direction of mesh vertices to swing repeatedly, resulting in local shear deformation on the mesh surface. This shear deformation will propagate along the mesh connection relationship to the surrounding area, thus affecting the vertex normal direction and inducing discontinuities such as seam cracking, mesh self-intersection, and overlapping topological patches in adjacent areas. Traditional The usual approach is to suppress local shear deformation by increasing the weight of the smoothing term, or to forcibly approximate the true boundary by increasing the sampling frequency of the point cloud. However, isotropic smoothing terms do not distinguish the curvature direction and are prone to over-smoothing the originally sharp corners in the model, causing the feature boundaries to shrink in the normal direction and be blunted into arcs. Increasing the point cloud density will increase the computational load of the nonlinear iterative solution of the dense matrix, and it is still difficult to avoid the nonlinear gradient descent step non-convergence problem caused by heterogeneous noise. In areas where the point cloud is missing due to instrument line of sight obstruction, data traction suddenly fails, and the leading mesh vertices are also prone to unbounded deviation along the direction without constraints.

[0004] Therefore, the technical problem to be solved by this invention is how to capture the local mesh edge length distortion state during the mesh evolution process online to adjust the multi-source approximation weights, and reconstruct a self-consistent three-dimensional model of the topological manifold under the premise of integrating anisotropic curvature constraints and damping dissipation rules. Summary of the Invention

[0005] To address the problems in the background art, the technical solution of the present invention is as follows: A multi-source data fusion 3D modeling system, comprising:

[0006] The nonlinear geometric energy deformation control unit constructs a mesh energy functional and calculates the mesh edge distortion rate of the mesh vertex before updating the mesh vertex coordinates. When the mesh edge distortion rate exceeds the safety boundary, it enters an isotropic relaxation mode and calculates the arithmetic mean center coordinates of the first-order neighboring vertices of the mesh vertex as temporary coordinate values.

[0007] The anisotropic curvature tensor evaluation unit extracts the local curvature features of mesh vertices in isotropic relaxation mode. When the maximum principal curvature value exceeds the critical curvature threshold of 2.5, it restricts the mesh vertices to complete projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor in the tangent plane.

[0008] The geometric prior topology filling unit monitors the coordinate iteration change of the grid vertices to establish the geometric deformation field vector. When the data confidence in a local area is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the data approximation term in the grid energy functional of the corresponding area is cut off, and the grid vertices in the local area are controlled to complete the geometric trend filling according to the grid curvature direction.

[0009] The manifold mesh topology correction unit recalculates the mesh edge distortion rate after updating the coordinates of the mesh vertices, and switches back to data-driven mode when the mesh edge distortion rate returns to within the safe boundary.

[0010] Preferably, in the nonlinear geometric energy deformation control unit, the safety boundary is set to 0.8 to 1.2; when the mesh edge length distortion rate is less than 0.8 or greater than 1.2, it is determined that the mesh deformation of the mesh vertex exceeds the limit. The nonlinear geometric energy deformation control unit controls the mesh vertex update mechanism to enter the isotropic relaxation mode and forcibly configures the first data point cloud approximation weight and the second data point cloud approximation weight in the mesh energy functional to 0, so as to block the disturbance effect of the spatial geometric alignment residual between heterogeneous point cloud data on the mesh vertex coordinate update.

[0011] Preferably, in the nonlinear geometric energy deformation control unit, the grid energy functional is constructed by weighting the laser point cloud coordinate constraint term, the dense photogrammetry point cloud coordinate constraint term, and the Laplace second-order differential curvature smoothing term, wherein the numerical distribution of each corresponding weight is determined by the spatial confidence distribution field of the laser point cloud and the dense photogrammetry point cloud.

[0012] Preferably, the multi-source data fusion 3D modeling system also includes a local mesh subdivision unit connected to the anisotropic curvature tensor evaluation unit; in the anisotropic curvature tensor evaluation unit, when the update mechanism of the mesh vertex is in the isotropic relaxation mode, the anisotropic curvature tensor evaluation unit controls the local mesh subdivision unit to complete the one-to-four topological mesh refinement of the over-limit region where the mesh vertex is located, so as to improve the degree of freedom of the digital geometric representation of the over-limit region and restrict the high curvature edge from shrinking or collapsing distortion under the constraint of the Laplace second-order differential curvature smoothing term.

[0013] Preferably, in the anisotropic curvature tensor evaluation unit, when the maximum principal curvature value of the mesh vertex exceeds the critical curvature threshold of 2.5, the anisotropic curvature tensor evaluation unit restricts the mesh vertex to complete projective sliding only in the tangent plane parallel to the mesh vertex along the minimum principal curvature direction of the anisotropic curvature tensor, in order to block the inward normal contraction of the mesh surface caused by the Laplace smoothing mean update.

[0014] Preferably, in the geometric prior topology filling unit, the geometric prior topology filling unit calculates the coordinate iteration change of the grid vertices online during the spatial grid update iteration process. When the confidence level of the local area data is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the multi-source data fusion 3D modeling system is controlled to enter the grid anti-collapse locking state, and the data approximation term in the grid energy functional of the corresponding area is cut off.

[0015] Preferably, the mesh anti-collapse locking state includes: controlling the update step size of the center vertex of the empty mesh region to decrease logarithmically with the search depth by using the mesh vertex update step size decay operator; calling the second-order difference extrapolation operator of the curvature of the neighborhood manifold to control the mesh vertices of the target region to maintain the geometric trend evolution, and completing the geometric topological closure of the data blind zone in accordance with the existing manifold surface, so as to eliminate the mesh's boundaryless drift and inward collapse distortion.

[0016] Preferably, when the manifold mesh topology correction unit performs manifold mesh topology self-check and repair, it includes: performing mesh edge flipping and topology regularization for non-manifold mesh edge-controlled multi-source data fusion 3D modeling systems with a degree less than 3; and performing geometric facet cutting, splitting, and topology separation for non-manifold mesh edge-controlled multi-source data fusion 3D modeling systems shared by more than two faces.

[0017] Preferably, in a multi-source data fusion 3D modeling system, after the multi-source data fusion 3D modeling system recalculates the mesh edge length distortion rate of all connected mesh edges by controlling the multi-source data fusion 3D modeling system every 3 gradient iterations, if it is determined that the mesh edge length distortion rate of all connected mesh edges has returned to within 0.8 to 1.2, the manifold mesh topology correction unit controls the multi-source data fusion 3D modeling system to restore the numerical distribution of the first data point cloud approximation weight and the second data point cloud approximation weight under the data-driven mode.

[0018] A multi-source data fusion 3D modeling method, used to run a multi-source data fusion 3D modeling system, includes the following steps:

[0019] Step S1: The nonlinear geometric energy deformation control unit constructs the mesh energy functional and calculates the mesh edge distortion rate of the mesh vertex before updating the mesh vertex coordinates through gradient descent. When the mesh edge distortion rate exceeds the safety boundary, the unit enters the isotropic relaxation mode and calculates the arithmetic mean center coordinates of the first-order neighboring vertices of the mesh vertex as temporary coordinate values.

[0020] Step S2: The anisotropic curvature tensor evaluation unit extracts the local curvature features of the mesh vertices in the isotropic relaxation mode. When the maximum principal curvature value exceeds the critical curvature threshold of 2.5, the mesh vertices are restricted to complete projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor in the tangent plane.

[0021] Step S3: The geometric deformation field vector is established by monitoring the coordinate iteration change of the grid vertices by the geometric prior topology filling unit. When the data confidence of the local area is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the data approximation term in the grid energy functional of the corresponding area is cut off, and the grid vertices of the local area are controlled to complete the geometric trend filling according to the grid curvature direction.

[0022] Step S4: The manifold mesh topology correction unit recalculates the mesh edge distortion rate after updating the mesh vertex coordinates, and switches back to data-driven mode when the mesh edge distortion rate returns to within the safe boundary.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] 1. In multi-source data fusion 3D modeling, the nonlinear geometric energy deformation control unit introduces mesh edge length distortion rate judgment during mesh vertex update. It can promptly identify local stretching, compression, and shear anomalies generated by multi-source point clouds in the seam region. When the mesh deformation exceeds the safety boundary, the system puts the relevant mesh vertices into an isotropic relaxation mode and cuts off the pulling effect of conflicting point cloud data on the vertex at that point. The vertex first completes local mitigation according to the average position of the neighboring mesh. This can reduce the situation where heterogeneous point cloud residuals continue to pull the mesh surface at the seam, reduce the risk of local shear distortion being transmitted to the surrounding mesh, and thus avoid problems such as surface superposition, mesh self-intersection, and topological tearing. This ensures that the 3D model after multi-source data fusion maintains good continuity in the seam region.

[0025] 2. The anisotropic curvature tensor evaluation unit continues to monitor local curvature changes after the mesh enters the isotropic relaxation mode. It can distinguish between ordinary smooth regions and sharp edge regions. For high curvature edges, the system restricts the mesh vertices to complete projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor, and improves the mesh representation capability of the region by local mesh subdivision. This can reduce the excessive flattening of sharp edges by Laplacian smoothing and avoid normal contraction, blunting or collapse of high curvature edges. This allows the reconstructed model to retain the boundary contours and angular features of the irregular structure while maintaining surface smoothness.

[0026] 3. Geometric prior topology filling elements introduce a mesh anti-collapse locking state in areas where point clouds are missing or have low confidence. This prevents the mesh in blind areas from drifting along unconstrained directions when there is a lack of data. The system cuts off the data approximation terms in the energy functional of the corresponding region mesh and controls the update step size of the mesh vertices in the missing region to gradually decay, so that the mesh update in the blind area is no longer driven by unreliable data. At the same time, the system continues the geometric trend of the existing surface based on the curvature of the neighborhood manifold, so that the missing region can close along the shape of the surrounding surface. This can reduce mesh boundaryless drift, inward collapse and normal overlap and ghosting, and improve the surface continuity of the 3D reconstruction results in the occluded blind area. Attached Figure Description

[0027] Figure 1 This is a flowchart of the mesh geometry deformation and topology self-checking correction process of the present invention;

[0028] Figure 2 This is a functional diagram of the multi-source data fusion modeling system of the present invention.

[0029] The objectives, features, and advantages of this invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0030] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0031] A multi-source data fusion 3D modeling system, comprising:

[0032] The nonlinear geometric energy deformation control unit constructs a mesh energy functional and calculates the mesh edge distortion rate of the mesh vertex before updating the mesh vertex coordinates. When the mesh edge distortion rate exceeds the safety boundary, it enters an isotropic relaxation mode and calculates the arithmetic mean center coordinates of the first-order neighboring vertices of the mesh vertex as temporary coordinate values.

[0033] The anisotropic curvature tensor evaluation unit extracts the local curvature features of mesh vertices in isotropic relaxation mode. When the maximum principal curvature value exceeds the critical curvature threshold of 2.5, it restricts the mesh vertices to complete projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor in the tangent plane.

[0034] The geometric prior topology filling unit monitors the coordinate iteration change of the grid vertices to establish the geometric deformation field vector. When the data confidence in a local area is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the data approximation term in the grid energy functional of the corresponding area is cut off, and the grid vertices in the local area are controlled to complete the geometric trend filling according to the grid curvature direction.

[0035] The manifold mesh topology correction unit recalculates the mesh edge distortion rate after updating the coordinates of the mesh vertices, and switches back to data-driven mode when the mesh edge distortion rate returns to within the safe boundary.

[0036] Preferably, in the nonlinear geometric energy deformation control unit, the safety boundary is set to 0.8 to 1.2; when the mesh edge length distortion rate is less than 0.8 or greater than 1.2, it is determined that the mesh deformation of the mesh vertex exceeds the limit. The nonlinear geometric energy deformation control unit controls the mesh vertex update mechanism to enter the isotropic relaxation mode and forcibly configures the first data point cloud approximation weight and the second data point cloud approximation weight in the mesh energy functional to 0, so as to block the disturbance effect of the spatial geometric alignment residual between heterogeneous point cloud data on the mesh vertex coordinate update.

[0037] Preferably, in the nonlinear geometric energy deformation control unit, the grid energy functional is constructed by weighting the laser point cloud coordinate constraint term, the dense photogrammetry point cloud coordinate constraint term, and the Laplace second-order differential curvature smoothing term, wherein the numerical distribution of each corresponding weight is determined by the spatial confidence distribution field of the laser point cloud and the dense photogrammetry point cloud.

[0038] Preferably, the multi-source data fusion 3D modeling system also includes a local mesh subdivision unit connected to the anisotropic curvature tensor evaluation unit; in the anisotropic curvature tensor evaluation unit, when the update mechanism of the mesh vertex is in the isotropic relaxation mode, the anisotropic curvature tensor evaluation unit controls the local mesh subdivision unit to complete the one-to-four topological mesh refinement of the over-limit region where the mesh vertex is located, so as to improve the degree of freedom of the digital geometric representation of the over-limit region and restrict the high curvature edge from shrinking or collapsing distortion under the constraint of the Laplace second-order differential curvature smoothing term.

[0039] Preferably, in the anisotropic curvature tensor evaluation unit, when the maximum principal curvature value of the mesh vertex exceeds the critical curvature threshold of 2.5, the anisotropic curvature tensor evaluation unit restricts the mesh vertex to complete projective sliding only in the tangent plane parallel to the mesh vertex along the minimum principal curvature direction of the anisotropic curvature tensor, in order to block the inward normal contraction of the mesh surface caused by the Laplace smoothing mean update.

[0040] Preferably, in the geometric prior topology filling unit, the geometric prior topology filling unit calculates the coordinate iteration change of the grid vertices online during the spatial grid update iteration process. When the confidence level of the local area data is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the multi-source data fusion 3D modeling system is controlled to enter the grid anti-collapse locking state, and the data approximation term in the grid energy functional of the corresponding area is cut off.

[0041] Preferably, the mesh anti-collapse locking state includes: controlling the update step size of the center vertex of the empty mesh region to decrease logarithmically with the search depth by using the mesh vertex update step size decay operator; calling the second-order difference extrapolation operator of the curvature of the neighborhood manifold to control the mesh vertices of the target region to maintain the geometric trend evolution, and completing the geometric topological closure of the data blind zone in accordance with the existing manifold surface, so as to eliminate the mesh's boundaryless drift and inward collapse distortion.

[0042] Preferably, when the manifold mesh topology correction unit performs manifold mesh topology self-check and repair, it includes: performing mesh edge flipping and topology regularization for non-manifold mesh edge-controlled multi-source data fusion 3D modeling systems with a degree less than 3; and performing geometric facet cutting, splitting, and topology separation for non-manifold mesh edge-controlled multi-source data fusion 3D modeling systems shared by more than two faces.

[0043] Preferably, in a multi-source data fusion 3D modeling system, after the multi-source data fusion 3D modeling system recalculates the mesh edge length distortion rate of all connected mesh edges by controlling the multi-source data fusion 3D modeling system every 3 gradient iterations, if it is determined that the mesh edge length distortion rate of all connected mesh edges has returned to within 0.8 to 1.2, the manifold mesh topology correction unit controls the multi-source data fusion 3D modeling system to restore the numerical distribution of the first data point cloud approximation weight and the second data point cloud approximation weight under the data-driven mode.

[0044] A multi-source data fusion 3D modeling method, used to run a multi-source data fusion 3D modeling system, includes the following steps:

[0045] Step S1: The nonlinear geometric energy deformation control unit constructs the mesh energy functional and calculates the mesh edge distortion rate of the mesh vertex before updating the mesh vertex coordinates through gradient descent. When the mesh edge distortion rate exceeds the safety boundary, the unit enters the isotropic relaxation mode and calculates the arithmetic mean center coordinates of the first-order neighboring vertices of the mesh vertex as temporary coordinate values.

[0046] Step S2: The anisotropic curvature tensor evaluation unit extracts the local curvature features of the mesh vertices in the isotropic relaxation mode. When the maximum principal curvature value exceeds the critical curvature threshold of 2.5, the mesh vertices are restricted to complete projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor in the tangent plane.

[0047] Step S3: The geometric deformation field vector is established by monitoring the coordinate iteration change of the grid vertices by the geometric prior topology filling unit. When the data confidence of the local area is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the data approximation term in the grid energy functional of the corresponding area is cut off, and the grid vertices of the local area are controlled to complete the geometric trend filling according to the grid curvature direction.

[0048] Step S4: The manifold mesh topology correction unit recalculates the mesh edge distortion rate after updating the mesh vertex coordinates, and switches back to data-driven mode when the mesh edge distortion rate returns to within the safe boundary.

[0049] Example 1: The current 3D reconstruction system for large irregular structures runs on a computing workstation equipped with a graphics processing core. It receives laser point cloud data collected by a 3D laser scanner and dense photogrammetric point cloud data generated by multi-view photogrammetry. Due to the differences between the 3D laser scanner and the photogrammetric sensor in terms of measurement principle, spatial resolution, and noise distribution, the merged point cloud data after spatial rigid registration will exhibit spatial displacement residuals and non-uniform local noise in the joint area. If the data-driven mode is directly used for mesh deformation, the mesh vertices will experience anisotropic stretching and compression of the mesh edges under the action of conflicting geometric tensions. Local shear strain will accumulate, leading to self-reversal of the normal on the mesh surface, topological tearing, and ghosting of non-manifold edges. This causes the finite element mechanical simulation analysis to be interrupted due to the damage to the geometric manifold topology.

[0050] The multi-source data fusion 3D modeling system receives the outer bounding box boundary of the space to be modeled. An initial topological mesh generation unit generates a basic closed initial topological mesh with manifold connectivity according to mesh partitioning rules. The surface vertices of this initial topological mesh are uniformly distributed. Subsequently, a heterogeneous source confidence field construction unit imports laser point cloud data and dense photogrammetric point cloud data, and performs a process on each vertex of the initial topological mesh. Perform spatial neighborhood traversal calculations to extract vertices falling within the grid. Within a sphere's neighborhood centered at a radius of 15 mm, all data points are analyzed. The physical scanning distance, the angle between the scan line and the surface normal, and the local curvature dispersion based on the covariance matrix eigenvalues ​​are recorded for each data point. By calculating the variance of the point set distribution of the laser point cloud data and the dense photogrammetry point cloud data within the sphere's neighborhood, the corresponding inverse variance is determined as the fusion weight, thereby solving for the mesh vertices. Confidence weights corresponding to laser point cloud data With confidence weights corresponding to dense photogrammetric point cloud data A spatial confidence distribution field consisting of discrete vectors is constructed and directly loaded into the vertex index table of the initial topological mesh.

[0051] The nonlinear geometric energy deformation control unit receives the spatial confidence distribution field and constructs a system-level grid energy functional. Grid energy functional The formula is as follows: Under this mathematical constraint, the nonlinear geometric energy deformation control unit uses the explicit Euler integral method to calculate the grid system's energy functional. The global partial derivative matrix under constraints is used to determine the spatial displacement vector of the initial topological mesh vertices along the negative gradient direction. The coordinate update formula is as follows: ,in, For system-level grid energy functionals, For grid vertices Corresponding to the confidence weights of the laser point cloud data, For grid vertices Corresponding to the confidence weights of dense photogrammetric point cloud data, This is the Euclidean distance penalty term for the approximation of the laser point cloud surface from the mesh surface. This is the Euclidean distance penalty term for the approximation of the grid surface to the dense photogrammetric point cloud surface. This is a second-order differential curvature smoothing constraint term based on the grid Laplacian operator. For grid vertices The spatial coordinate displacement vector in a single iteration step The spatial three-dimensional coordinates of the currently processed mesh vertices are represented by the second-order differential curvature smoothing constraint term based on the mesh Laplacian operator, which is implemented by constructing a mesh Laplacian matrix on the triangular mesh. In the actual discretization calculation, this embodiment uses an umbrella-shaped cotangent weight allocation mode to define the mesh Laplacian operator. For any target vertex, all adjacent vertices that form a first-order topological connection with it are traversed first, and the sum of the cotangent values ​​of the two acute or obtuse angles opposite each adjacent edge is calculated as the geometric weight coefficient of the corresponding adjacent vertex. Then, the coordinates of all adjacent vertices are multiplied by their respective weight coefficients and weighted summed, and the product of the target vertex coordinates and the total weight is subtracted to obtain the first-order Laplacian differential coordinate vector of the vertex. On this basis, the processor performs the same Laplacian operator action again on all vertices of the entire mesh, that is, performs two weighted difference iterations on the first-order Laplacian differential coordinate vector to construct a second-order differential curvature smoothing term, which is used to constrain the local bending energy of the mesh surface and smooth high-frequency noise fluctuations.

[0052] In the iterative step before updating the spatial coordinates of the vertex on the mesh surface, the spatial edge topology state machine inside the nonlinear geometric energy deformation control unit processes each mesh vertex within the mesh. Perform topology strain feedforward verification and extract the current mesh vertices online. Read the current spatial Euclidean distance length of all adjacent mesh edges that form a first-order topological connection. And retrieve the historical calibration length of the corresponding edge in the initial topology mesh. Based on this, the mesh edge distortion rate of each connected edge is calculated, and the calculation formula is as follows: ,in, The distortion rate is the mesh side length. This represents the current spatial Euclidean distance length of the mesh edge in the current iteration step. For the historical calibration length of the corresponding edge, when the mesh edge length distortion rate of any connected edge... When the value is less than 0.8 or greater than 1.2, the vertex is determined to be subjected to excessive shear distortion due to geometric alignment residuals. The spatial edge topology state machine completes the state transition, switching the vertex from data-driven mode to isotropic relaxation mode, and forcibly approximating the weights of the first data point cloud in the mesh energy functional. Approximation weights with the second data point cloud Simultaneously, the configuration is set to 0 to sever the spatial tension constraint of the external conflict point cloud on this vertex. Subsequently, the processor calls the Laplace first-order difference operator to calculate the vertex. The arithmetic mean center coordinate of the vertex is calculated from the coordinates of all its adjacent vertices within its first-order topological neighborhood, and this arithmetic mean center coordinate is used as the temporary coordinate value of the vertex in the current iteration step.

[0053] The manifold mesh topology correction element is connected to the nonlinear geometric energy deformation control unit, and the mesh edge distortion rate is recalculated after the vertex coordinates are updated. When the distortion rate of the mesh edge lengths of all connected mesh edges of a vertex When all values ​​return to the safe boundary of 0.8 to 1.2, the spatial edge topology state machine completes the reverse migration, restoring the original weight ratio of the vertex in the mesh energy functional. Every 3 gradient iterations, the system recalculates the mesh edge length distortion rate of all connected mesh edges, and the manifold mesh topology correction unit performs manifold mesh topology self-check and repair on the entire mesh. When a non-manifold mesh edge with a degree less than 3 is detected, the control system completes mesh edge flipping and topology regularization. When a non-manifold mesh edge shared by more than two patches is detected, the control system completes geometric patch cutting and splitting topology separation, and rearranges the mesh surface vertex index connection table.

[0054] If the target vertex is in an isotropic relaxation mode and its region has high-curvature non-standard sharp boundary features, when the maximum principal curvature value of the target vertex exceeds the critical curvature threshold of 2.5, the anisotropic curvature tensor evaluation unit intervenes in the deformation update process of the vertex. This restricts the vertex to projective sliding only along the minimum principal curvature direction of the anisotropic curvature tensor within a tangent plane parallel to the mesh vertex. It also calls the local one-to-four topology mesh refinement operator to increase the mesh topology vertex density in the high-curvature over-limit region. For blind spots caused by complete sensor line-of-sight occlusion, when the confidence level of local region data is lower than... When the safety limit threshold is 0.05 and the direction of the geometric deformation field vector is reversed, the geometric prior topology filling unit control system enters the grid anti-collapse locking state, configures the data approximation term in the grid energy functional of the corresponding missing region to 0, and monitors the displacement iteration velocity vector of each grid vertex in the blind zone online; then, extrapolation is performed by the rigid second-order difference constraint equation based on the curvature of the known manifold boundary in the neighborhood, and the update step size decay operator of the grid vertex update step size controls the update step size of the center vertex of the missing grid region to decrease logarithmically with the search depth, so that the grid vertices in the target region maintain the geometric curvature trend and evolve smoothly.

[0055] During the specific execution of position updates, to avoid direct overlay of vertex coordinates and interference with global gradient descent iterations, when the spatial edge topological state machine switches the target vertex to isotropic relaxation mode, the explicit Euler integral calculation unit of the global energy functional suspends the gradient differentiation and step size accumulation operations for that specific over-limit vertex. The processor directly stops the gradient update and propagation of that specific vertex and writes the arithmetic mean center coordinates calculated by the first-order Laplacian difference operator into the vertex position register of the current iteration step. This operation allows the over-limit vertex to escape the synchronous evolution constraint of the global matrix and complete local elastic relaxation through in-situ overlay of local coordinates. Only after the mesh edge distortion rate of the vertex returns to within the safe boundary and the data-driven mode is restored will the processor reintegrate the vertex into the unified solution loop of global gradient descent, thereby establishing a time-division and domain-division coordination mechanism between local position overlay and global gradient evolution.

[0056] The system undergoes 50 cascaded iterations of multi-scale pyramid calculations from coarse to fine. When the norm of the derivative matrix of the grid energy functional is lower than... When convergence is triggered, the manifold mesh topology correction unit completes the global normal vector alignment update to reduce the local geometric displacement residuals in the non-standard section joint area, reduce the occurrence rate of non-manifold edges, and suppress the self-overlapping shadows of multi-layer surfaces. The geometric sharp features at high curvature corners and the curvature continuity of the surface in the data missing blind area are both preserved, and finally a topologically consistent and tear-free three-dimensional fused geometric model is output.

[0057] Example 2: This verification experiment was conducted on a test platform with a graphics computing core. It was used to verify the operation of the 3D digital geometric reconstruction system maintaining mesh topology consistency when spatial displacement residuals and non-uniform local noise exist in the seam area of ​​the merged point cloud data. The laser point cloud data used in the experiment originated from terrestrial 3D laser scanning mapping, with a spatial resolution boundary of 2.0 mm. The dense photogrammetric point cloud data originated from UAV multi-view photogrammetric geometric reconstruction, with a sampling point density ranging from 500,000 to 800,000 per cubic meter. During the fusion process of non-standard cross-section seams, the geometrical physical parameter required for spatial neighborhood traversal calculation was the spatial radius. Its value directly affects the solution resolution of the spatial confidence distribution field and the spatial radius. The value of is mainly affected by the average point spacing of the test point cloud and the local curvature dispersion of the target surface. When the measured average point spacing of the point cloud increases and the local curvature dispersion exceeds the curvature threshold, in order to avoid the Laplace second-order differential curvature smoothing term from excessively smoothing high-curvature edges during deformation update, the processor controls the spatial radius. It moves closer to the lower limit of its value range.

[0058] In this embodiment, spatial radius The value range is limited to 5.0mm to 30.0mm, with a lower limit of 5.0mm and an upper limit of 30.0mm. The curvature threshold for judgment is set to 1.5. This is for irregularly shaped structures with an average point spacing of 3.0mm and including joint misalignment residuals. The spatial radius... The engineering example setting is 15.0 mm; the mesh side length distortion rate that triggers state machine transition in the nonlinear geometric energy deformation control unit is simultaneously adjusted. The safety boundary range is set to 0.8 to 1.2, the critical curvature threshold in the anisotropic curvature tensor evaluation unit is set to 2.5, and the safety limit threshold in the geometric prior topology filling unit is set to 0.05. This forms a fixed initial state for the experimental system input. In actual operation, the processor monitors the input point cloud data online in real time. When the measured average point spacing of the point cloud increases from 2.0 mm to 5.0 mm, and the local curvature dispersion of the dense point cloud in a local area calculated based on the covariance matrix is ​​greater than 1.5, the processor activates the parameter decay function to control the spatial radius. The spatial radius decreases linearly from the baseline value of 15.0 mm to the lower limit of 5.0 mm. Conversely, when the average point spacing of the point cloud is less than 2.0 mm and the curvature dispersion is not greater than 1.5, the spatial radius... Increasing the height to 30.0mm ensures that denoising smoothing and feature preservation remain consistent within the same digital geometry reconstruction process.

[0059] To determine the synergistic effect of each functional unit in eliminating local shear distortion and resisting noise disturbance, a multidimensional control experiment was organized under the same graphics computing environment. The experiment was divided into three groups: the sample group of this embodiment with complete unit configuration, the partially missing control group 1 with the tensile cutting mechanism of the spatial edge topological state machine removed, the partially missing control group 2 with the anisotropic curvature tensor evaluation unit removed, and the out-of-range control group 3 with the distortion rate boundary range widened to 0.5 to 1.5. The groups were placed under environmental disturbances in the seam area from low to high, and the intermediate state quantities and the performance indicators of the final output geometric mesh were measured online.

[0060] In this example sample, when the input seam alignment residual misalignment is 2.15mm and the injected Gaussian noise power is 10dB, the proportion of vertices in the normal mode is 94.2%, the proportion of isotropic relaxed vertices is 5.8%, the maximum mesh geometric approximation error is 0.23mm, the number of non-manifold edge ghosts is 0, and the convergence steps are 42. When the seam alignment residual misalignment increases to 5.42mm and the injected Gaussian noise power increases to 20dB, the proportion of vertices in the normal mode decreases to 83.5%, the proportion of isotropic relaxed vertices increases to 16.5%, and the maximum mesh geometric approximation error is 0.38mm. The number of non-manifold edge ghosts remained at 0, and the number of convergence steps was 48. When the misalignment of the seam alignment residual further increased to 12.18 mm and the injected Gaussian noise power increased to 30 dB, the proportion of vertices in the normal mode was 67.1%, the proportion of isotropic relaxed vertices was 32.9%, the maximum mesh geometric approximation error was 0.64 mm, the number of non-manifold edge ghosts remained at 0, and the number of convergence steps was 61. The experimental results show that as the seam misalignment and noise intensity increase, more mesh vertices are switched to the isotropic relaxation mode, the local heterogeneous tension is isolated in segments, and the number of non-manifold edge ghosts in the final output does not increase with the increase of perturbation.

[0061] In the partially missing control group 1, after removing the spatial edge topological state machine tension cutting mechanism, with an input seam alignment residual misalignment of 2.15 mm and an injected Gaussian noise power of 10 dB, the vertex ratio in the normal mode is 100.0%, the isotropic relaxed vertex ratio is 0.0%, the maximum mesh geometric approximation error increases to 1.45 mm, the number of non-manifold edge ghosts is 8, and the convergence steps are 126. When the seam alignment residual misalignment is 5.42 mm and the injected Gaussian noise power is 20 dB, the vertex ratio in the normal mode remains 100.0%, the isotropic relaxed vertex ratio remains 0.0%, the maximum mesh geometric approximation error increases to 3.12 mm, and the number of non-manifold edge ghosts is 27. The convergence steps reached 150. When the seam alignment residual misalignment was 12.18 mm and the injected Gaussian noise power was 30 dB, the vertex ratio in the normal mode remained at 100.0%, the isotropic relaxed vertex ratio remained at 0.0%, the maximum mesh geometric approximation error increased to 5.89 mm, and the number of non-manifold edge ghosts was 64. The convergence steps were still 150. In this set of data, the isotropic relaxed vertex ratio was always 0.0%, indicating that without the tension cutting mechanism, the vertices in the seam area could not switch from the data-driven mode to the isotropic relaxed mode in time. The spatial geometric alignment residual between heterogeneous point cloud data continued to affect the mesh vertex coordinate update, which ultimately manifested as a synchronous increase in geometric approximation error and the number of non-manifold edge ghosts.

[0062] In the partially missing control group 2, after removing the anisotropic curvature tensor evaluation unit, under the conditions of a seam alignment residual misalignment of 5.42 mm and an injected Gaussian noise power of 20 dB, the vertex ratio in the normal mode was 85.1%, the isotropic relaxed vertex ratio was 14.9%, the maximum mesh geometric approximation error was 1.84 mm, the number of non-manifold edge ghosts was 4, and the convergence steps were 78. This is comparable to the sample group in this embodiment under the same conditions of a seam alignment residual misalignment of 5.42 mm and an injected Gaussian noise power of 20 dB. Compared to the maximum mesh geometric approximation error of 0.38 mm, the number of non-manifold edge ghosts of 0, and the number of convergence steps of 48 obtained under Gaussian noise power, although some vertices in this control group still enter the isotropic relaxation mode, it lacks the constraint of completing the projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor when the maximum principal curvature value exceeds 2.5. Vertex updates in high curvature regions are prone to deviate from the low contraction direction in the tangent plane, thus increasing the geometric approximation error, the number of non-manifold edge ghosts, and the number of convergence steps.

[0063] In the out-of-range control group three, the grid side length distortion rate was... After widening the trigger range to 0.5 to 1.5, when the input seam alignment residual misalignment is 0.80 mm and the injected Gaussian noise power is 20 dB, the vertex ratio in the normal mode is 98.2%, the isotropic relaxed vertex ratio is 1.8%, the maximum mesh geometric approximation error is 1.12 mm, the number of non-manifold edge ghosts is 5, and the convergence steps are 92. When the input seam alignment residual misalignment is 1.20 mm and the injected Gaussian noise power is 20 dB, the normal mode vertex ratio is 98.2%, the isotropic relaxed vertex ratio is 1.8%, the maximum mesh geometric approximation error is 1.12 mm, the number of non-manifold edge ghosts is 5, and the convergence steps are 92. The scaled vertex ratio is 72.4%, the isotropic relaxed vertex ratio is 27.6%, the maximum mesh geometric approximation error is 0.59mm, the number of non-manifold edge ghosts is 0, the convergence steps are 58, and the two sets of endpoint states show that when the trigger interval is too wide, the state machine is not fully activated under low distortion perturbation. In the normal mode, the vertex ratio reaches 98.2%, the seam residual cannot be released in time, and still produces an approximation error of 1.12mm and 5 non-manifold edge ghosts; while the mesh edge distortion rate When controlled within the safety boundary of 0.8 to 1.2, the sample group in this embodiment did not produce non-manifold edge ghosting under three joint misalignment strengths of 2.15mm, 5.42mm and 12.18mm, and the maximum mesh geometric approximation error was also kept below 0.64mm.

[0064] In the above experiment, the sample group in this embodiment adjusted the neighborhood feature variance weights of the grid vertices by constructing a heterogeneous source confidence field, and calculated the grid energy functional of the generated system as follows: ,in, For system-level grid energy functionals, For grid vertices Corresponding to the confidence weights of the laser point cloud data, For grid vertices Corresponding to the confidence weights of dense photogrammetric point cloud data, This is the Euclidean distance penalty term for the approximation of the laser point cloud surface from the mesh surface. This is the Euclidean distance penalty term for the approximation of the grid surface to the dense photogrammetric point cloud surface. For the second-order differential curvature smoothing constraint term based on the mesh Laplacian operator, in the mesh energy functional Under constraints, the processor uses the explicit Euler integral method to calculate the coordinate displacement vector corresponding to the mesh system: ,in, For grid vertices The spatial coordinate displacement vector in a single iteration step Given the spatial three-dimensional coordinates of the currently processed mesh vertices, as the misalignment of the input seam alignment residual increases from 2.15mm to 12.18mm and the Gaussian noise power increases from 10dB to 30dB, when updating the coordinates according to the above mesh energy functional, the proportion of mesh vertices entering the isotropic relaxation mode increases from 5.8% to 32.9%. Correspondingly, the abnormal shear strain accumulation caused by the anisotropic point cloud tension is cut off, so that the maximum mesh geometric approximation error of the final output is stably controlled below 0.64mm, and the number of non-manifold edge ghosts remains at 0.

[0065] Distortion rate of grid side length The calculation formula for determining this is as follows: ,in, The distortion rate is the mesh side length. This represents the current spatial Euclidean distance length of the mesh edge in the current iteration step. For the historical calibration length of the corresponding edge, when this ratio falls outside the safety boundary of 0.8 to 1.2, the nonlinear geometric energy deformation control unit controls the mesh vertices to enter an isotropic relaxation mode and approximates the first data point cloud with weights. Approximation weights with the second data point cloud Configured to 0; when the manifold mesh topology correction element recalculates all connected mesh edges and confirms the mesh edge length distortion rate. When the value returns to within the range of 0.8 to 1.2, the system restores the numerical distribution of the first data point cloud approximation weight and the second data point cloud approximation weight under the data-driven mode. As can be seen from the sample group and the control group data in this embodiment, the safety boundary of 0.8 to 1.2 can enable local mesh vertices to escape the tension of conflicting point clouds and return to the stable mesh deformation process when the joint misalignment residual expands and the noise increases, thereby maintaining the continuity of the three-dimensional geometric model after the topological manifold self-check correction.

[0066] Example 3: After receiving the outer bounding box boundary of the space to be modeled, the current initial topology mesh generation unit divides the outer bounding box into uniform octree manifolds along the three-dimensional spatial axis. In the intersecting boundary region, based on the surface change characteristics where the curvature dispersion is greater than a preset threshold, it calls the local subdivision meshing rule to generate the initial topology mesh. The workstation's memory stores the input laser point cloud data and dense photogrammetry point cloud data. The nonlinear geometric energy deformation control unit calculates the mesh side length distortion rate in gradient iteration. When the value is less than 0.8 or greater than 1.2, the spatial edge topology state machine switches the target vertex to an isotropic relaxation mode, and the processor reads the mesh vertex. The first-order neighborhood connection relationship is obtained, and the coordinate sequence of adjacent vertices stored in the register is retrieved. The first-order Laplacian difference arithmetic mean of these adjacent vertex coordinates is calculated and summed. The obtained average center point spatial three-dimensional coordinates are written in place to the temporary projection address of the mesh vertex to cut off the external conflict geometric tension. This local relaxation update mechanism no longer introduces external point cloud input, and blocks mesh distortion error in the scenario of geometric misalignment caused by real manufacturing error.

[0067] The anisotropic curvature tensor evaluation unit retrieves curvature feature data of vertices in isotropic relaxation mode online. When the calculated maximum principal curvature value of a mesh vertex exceeds 2.5, the processor constructs the local tangent plane basis vector of that vertex in its internal register. It then determines the minimum principal curvature direction using the cross product of the local mesh surface normal and the maximum principal curvature direction, and constructs the anisotropic curvature tensor matrix accordingly. This restricts the mesh vertex to slide its projection within the local tangent plane only along the minimum principal curvature direction of the anisotropic curvature tensor during deformation updates. Simultaneously, it calls the local one-to-four topology mesh refinement operator to add three new topological vertices around the vertex and rearranges the local index connection table to increase... In regions with high curvature exceeding limits, the topology vertex density of the mesh is determined. Every three gradient iterations, the manifold mesh topology correction unit performs manifold mesh topology self-checks and repairs on the entire mesh. When a non-manifold mesh edge with a degree less than 3 is detected, the processor controls the system to complete mesh edge flipping and topology regularization by modifying the pointers in the vertex index connection table. When a non-manifold mesh edge shared by more than two faces is detected, the processor performs shearing and splitting on the geometric faces sharing the non-manifold mesh edge, separating the overlapping and shared faces from the same edge connection relationship, thus completing the geometric face shearing and splitting topology separation. This eliminates the defects of mesh ridge normal shrinkage and edge smoothing and blunting induced by multi-source heterogeneous noise at sharp contours.

[0068] When constructing the anisotropic curvature tensor matrix, the processor first performs a cross product calculation between the normal vector of the local mesh surface and the maximum principal curvature direction vector to determine the minimum principal curvature direction vector orthogonal to it. Then, it performs an outer product operation between the maximum principal curvature direction vector and its own transpose to generate the first direction matrix, and performs an outer product operation between the minimum principal curvature direction vector and its own transpose to generate the second direction matrix. Next, it multiplies the measured maximum principal curvature value with the first direction matrix, and multiplies the measured minimum principal curvature value with the second direction matrix. Finally, it sums the two product results to construct the complete anisotropic curvature tensor matrix. This second-order tensor matrix is ​​used to characterize the anisotropic curvature distribution of the target vertex in the tangent plane and provides geometric operator support for restricting the vertex to complete projective sliding along the minimum principal curvature direction.

[0069] In the multi-source data loss blind zone where the sensor line of sight is completely blocked by the internal web or stiffening ribs of heavy steel structure nodes, when the confidence level of local area data of a local mesh patch drops below 0.05 and the direction of the geometric deformation field vector is reversed, the geometric prior topology filling unit control system enters the mesh anti-collapse locking state. The processor forcibly configures the data approximation term in the mesh energy functional to 0, and simultaneously monitors the displacement iteration velocity vector of each mesh vertex in the blind zone online. Starting from the known manifold boundary curvature at the edge of the blind zone, a rigid second-order difference constraint equation based on the neighborhood Laplace operator is constructed, and the calculated extrapolated displacement increment is multiplied by the number of iterations. Logarithmically decaying topological velocity damping dissipation operator The calculation formula for the topological velocity damping dissipation operator is as follows: ,in, For topological velocity damping dissipation operators, The number of gradient iteration steps to fill the current blind zone is determined by the processor during a single coordinate update, which combines the previous step size with the topological velocity damping dissipation operator. Multiplication is used to control the spatial displacement increment of vertices, so that the update step size of the center vertex of the missing region decays and converges with the approach depth, thus preventing mesh surface penetration collapse and normal overlap caused by lack of data traction.

[0070] The construction and solution process of the rigid second-order difference constraint equation based on the neighborhood Laplacian operator is as follows: Within the data-missing blind zone, the processor establishes multiple advancing topology layers inward along the edge of the blind zone; for the current blind zone vertex to be filled, first obtain the first-order Laplacian coordinate vector of its adjacent vertices on the existing known manifold boundary. This vector represents the initial curvature direction of the edge region. In order to maintain the continuous extrapolation of the geometric trend, the Laplacian coordinates of the vertices at the leading edge of the blind zone filling and the Laplacian coordinates of the vertices on the outer known boundary satisfy the second-order difference identity, that is, the current vertex... The difference between the Laplacian difference rate of change and the Laplacian difference rate of change of adjacent known vertices is kept at 0. The processor establishes a system of linear equations containing the spatial coordinates of the vertices to be filled, sets the coefficients of the second-order Laplacian operator matrix of the unknown vertices inside the blind zone to a fixed constant, and uses the coordinates of the known vertices on the edge as boundary conditions for relaxation iteration. By solving this rigid second-order difference equation, the manifold curvature is linearly extrapolated inward, so that the generated empty mesh conforms to the tangent direction and curvature direction of the existing surface on the periphery, realizing the geometric topological closure of the blind zone data at the boundary.

[0071] In the cascaded deformation solution, when the norm of the derivative matrix of the overall energy functional of the mesh is lower than... When the iteration terminates, after the manifold mesh topology correction unit completes the global normal vector alignment update, the occurrence rate of non-manifold edges in the output 3D geometric model converges to 0%, the self-overlapping shadows of multi-layer surfaces are completely eliminated, and the geometrical sharp features at high curvature corners and the curvature continuity of surfaces in data-missing blind areas are preserved. This is achieved by using the mesh edge length distortion rate... As a feedforward variable, and with step size decay control implemented under grid anti-collapse locking in the extreme defect region, the system coordinates local approximation and manifold topological consistency during the digital geometric grid manifold evolution generation stage, providing manifold grid completeness support for the mechanical analysis and deformation monitoring of large-scale non-standard industrial twins.

[0072] Example 4: This example combines Figures 1 to 2 This paper describes a multi-source data fusion 3D modeling system and method, such as... Figure 1As shown, step S1 involves constructing a mesh energy functional using a nonlinear geometric energy deformation control unit. Before updating the mesh vertex coordinates via gradient descent, the mesh edge distortion rate of the mesh vertices is calculated. When the mesh edge distortion rate exceeds the safety boundary, an isotropic relaxation mode is entered. The arithmetic mean center coordinates of the first-order neighboring vertices of the mesh vertex are calculated as temporary coordinate values. Step S2 involves extracting the local curvature features of the mesh vertices in the isotropic relaxation mode using an anisotropic curvature tensor evaluation unit. When the maximum principal curvature value exceeds the critical curvature threshold of 2.5, the minimum curvature value of the mesh vertex along the anisotropic curvature tensor in the tangent plane is restricted. In step S3, the geometric prior topology filling unit monitors the coordinate iteration change of the mesh vertices to establish the geometric deformation field vector. When the data confidence in a local area is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the data approximation term in the energy functional of the corresponding area mesh is cut off, and the mesh vertices in the local area are controlled to complete the geometric trend filling according to the mesh curvature direction. In step S4, the manifold mesh topology correction unit recalculates the mesh edge length distortion rate after updating the coordinates of the mesh vertices. When the mesh edge length distortion rate returns to within the safety boundary, the data-driven mode is switched back.

[0073] like Figure 2 As shown, the multi-source data fusion 3D modeling method consists of nonlinear geometric energy deformation control, anisotropic curvature tensor evaluation, geometric prior topology filling, and manifold mesh topology correction. Nonlinear geometric energy deformation control includes constructing the mesh energy functional and calculating the mesh edge distortion rate. Anisotropic curvature tensor evaluation includes extracting local curvature features and restricting mesh vertex projection sliding. Geometric prior topology filling includes monitoring coordinate iteration changes and completing the filling based on the mesh curvature. Manifold mesh topology correction includes recalculating the mesh edge distortion rate and switching back to the data-driven mode.

[0074] Example 5: The current data fusion 3D modeling system runs on a computing workstation equipped with a graphics processing core. A ground-based 3D laser scanner and a UAV multi-view photogrammetry core mechanism are fixed within a multi-scale geometric calibration field. The computing workstation collects control point data with known spatial coordinates in the multi-scale geometric calibration field, constructs an initial spatial rigid registration matrix, and retrieves sensor physical error correction parameters stored in non-volatile memory. Subsequently, based on the point density difference between the laser point cloud data and the dense photogrammetry point cloud data, the initial registration residual feature vector is calculated. Simultaneously, the temperature drift compensation damping constant is retrieved from the historical statistical database. The initial registration residual feature vector is input into the initial calibration loop of the nonlinear geometric energy deformation control unit, running the system-level grid energy functional... Before, in-situ cancellation of the initial topology mesh vertices Spatial coordinate zero-point drift error.

[0075] In this embodiment, the temperature drift compensation damping constant is used to correlate changes in the external ambient temperature with the structural thermal deformation of the surveying instrument. The historical statistical database pre-stores the fuselage axis drift of the 3D laser scanner at different operating temperatures. When the temperature sensor in the calibration field detects a 1°C temperature change, the processor retrieves the corresponding temperature drift compensation damping constant and maps the thermal displacement of the fuselage axis into a 3D spatial coordinate correction. Before calculating the mesh deformation, the initial topological mesh vertices are directly adjusted. The spatial coordinates are corrected by homogeneous coordinate translation. Through this coordinate pre-correction, the physical temperature drift residual at the hardware level is reflected in the mesh energy functional. The deviations are offset before the iterative calculation begins, thus preventing thermodynamic measurement biases from coupling to the subsequent evolution of the pure geometric deformation field.

[0076] When a multi-source data fusion 3D modeling system scans large-scale, non-standard equipment, the historical statistical database advances along the sampling time axis, using a sliding time window to eliminate point cloud noise distribution data that has exceeded the specified time period. The processor periodically calculates the time decay coefficient of the temperature drift compensation damping constant based on the real-time displacement feedback of the newly merged point cloud data in the seam area. When the mesh energy functional... When the norm of the derivative matrix oscillates and fails to converge after 5 consecutive cycles, the adaptive parameter reconstruction loop locks the current deformation displacement field. The computation workstation then drives the control point sensor of the multi-scale geometric calibration field to scan the spatial points on the surface of the heavy steel structure node, and overwrites the original memory address with the newly calculated geometric residual weight calibration value in situ, so that the initial mesh generation unit can reconstruct the vertex manifold connection relationship and reorganize the connection structure when a large noise interference occurs.

[0077] Example 6: The current calibration control mechanism uses the three-dimensional digital manifold topology geometry of a stainless steel calibration entity containing a standard stepped cross-section as the benchmark verification source input. The processor loads the known surface reflection curvature of the calibration entity and the discrete point cloud noise features collected by multi-view sensors into the storage area. The heterogeneous source confidence field construction unit retrieves the average local variance at historical calibration points, and the processor verifies the state switching time delay constant of the spatial edge topology state machine in isotropic relaxation mode. In this example, the mesh side length distortion rate... The trigger interval for determination is limited to a critical value between 0.8 and 1.2. The calibration control mechanism constructs a spatial registration misalignment in the transition zone of the calibration entity to determine the step size of a single positional derivative update and to verify the mesh energy functional after 50 consecutive multi-scale pyramid cascade deformation updates of the initial topology mesh. The gradient matrix decreases polarity.

[0078] Geometric prior topology filling cells retrieve the gradient iteration steps corresponding to the current blind zone filling. Topological velocity damping dissipation operator The processor monitors the mesh vertices within the blind zone under the dead state of the adaptive parameter reconstruction loop. The displacement iteration velocity vector, when the mesh energy functional is calculated. The norm of the partial derivative matrix is ​​lower than When the system triggers a convergence termination judgment, the calibration control mechanism then solves the residual convergence criterion formula to determine the global displacement convergence residual of the grid. The formula for the residual convergence criterion is expressed as: ,in, For the global displacement convergence residual of the grid, For grid vertices The spatial coordinate displacement vector in a single iteration step To calibrate the total number of mesh vertices in the local curvature exceeding limit region of the entity, the manifold mesh topology correction unit performs a global manifold topology feature self-check on the entire mesh after the update. The results show that the occurrence rate of non-manifold edges in the output manifold mesh geometric model remains at 0, and the maximum mesh geometric approximation error at the high curvature sharp boundary converges at a low position. This completes the pre-deployment functional verification for digital twin geometric 3D reconstruction of non-standard heavy steel structure nodes.

[0079] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0080] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A multi-source data fusion 3D modeling system, characterized in that, include: A nonlinear geometric energy deformation control unit is constructed to generate a mesh energy functional and calculate the mesh edge length distortion rate of the mesh vertices before updating the mesh vertex coordinates. When the mesh edge distortion rate exceeds the safety boundary, it enters the isotropic relaxation mode and calculates the arithmetic mean center coordinates of the first-order neighboring vertices of the mesh vertex as temporary coordinate values. The anisotropic curvature tensor evaluation unit extracts the local curvature features of mesh vertices in isotropic relaxation mode. When the maximum principal curvature value exceeds the critical curvature threshold of 2.5, it restricts the mesh vertices to complete projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor in the tangent plane. The geometric prior topology filling unit monitors the coordinate iteration change of the grid vertices to establish the geometric deformation field vector. When the data confidence in a local area is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the data approximation term in the grid energy functional of the corresponding area is cut off, and the grid vertices in the local area are controlled to complete the geometric trend filling according to the grid curvature direction. The manifold mesh topology correction unit recalculates the mesh edge distortion rate after updating the coordinates of the mesh vertices, and switches back to data-driven mode when the mesh edge distortion rate returns to within the safe boundary.

2. The multi-source data fusion 3D modeling system according to claim 1, characterized in that, In the nonlinear geometric energy deformation control unit, the safety boundary is set to 0.8 to 1.

2. When the mesh edge distortion rate is less than 0.8 or greater than 1.2, the mesh deformation of the mesh vertex is determined to be out of limit. The nonlinear geometric energy deformation control unit controls the mesh vertex update mechanism to enter the isotropic relaxation mode and forcibly configures the first data point cloud approximation weight and the second data point cloud approximation weight in the mesh energy functional to 0, so as to block the disturbance effect of the spatial geometric alignment residual between heterogeneous point cloud data on the mesh vertex coordinate update.

3. The multi-source data fusion 3D modeling system according to claim 1, characterized in that, In the nonlinear geometric energy deformation control unit, the grid energy functional is constructed by weighting the laser point cloud coordinate constraint term, the dense photogrammetry point cloud coordinate constraint term, and the Laplace second-order differential curvature smoothing term. The numerical distribution of each corresponding weight is determined by the spatial confidence distribution field of the laser point cloud and the dense photogrammetry point cloud.

4. The multi-source data fusion 3D modeling system according to claim 1, characterized in that, The multi-source data fusion 3D modeling system also includes a local mesh subdivision unit connected to the anisotropic curvature tensor evaluation unit. In the anisotropic curvature tensor evaluation unit, when the update mechanism of the mesh vertex is in the isotropic relaxation mode, the anisotropic curvature tensor evaluation unit controls the local mesh subdivision unit to complete the one-to-four topological mesh refinement of the over-limit region where the mesh vertex is located, so as to improve the degree of freedom of the digital geometric representation of the over-limit region and restrict the high curvature edge from shrinking or collapsing distortion under the constraint of the Laplace second-order differential curvature smoothing term.

5. The multi-source data fusion 3D modeling system according to claim 1, characterized in that, In the anisotropic curvature tensor evaluation unit, when the maximum principal curvature value of the mesh vertex exceeds the critical curvature threshold of 2.5, the anisotropic curvature tensor evaluation unit restricts the mesh vertex to complete projective sliding only in the tangent plane parallel to the mesh vertex along the minimum principal curvature direction of the anisotropic curvature tensor, in order to block the inward normal contraction of the mesh surface caused by the Laplace smoothing mean update.

6. The multi-source data fusion 3D modeling system according to claim 1, characterized in that, In the geometric prior topology filling unit, the geometric prior topology filling unit calculates the coordinate iteration change of the grid vertices online during the spatial grid update iteration process. When the confidence of the local area data is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the multi-source data fusion 3D modeling system is controlled to enter the grid anti-collapse locking state, and the data approximation term in the grid energy functional of the corresponding area is cut off.

7. A multi-source data fusion 3D modeling system according to claim 6, characterized in that, The mesh anti-collapse locking state includes: controlling the update step size of the center vertex of the missing mesh region to decrease logarithmically with the search depth by using the mesh vertex update step size decay operator; calling the second-order difference extrapolation operator of the curvature of the neighborhood manifold to control the mesh vertices of the target region to maintain the geometric trend evolution, and completing the geometric topological closure of the data blind zone in accordance with the existing manifold surface, so as to eliminate the mesh's boundaryless drift and inward collapse distortion.

8. The multi-source data fusion 3D modeling system according to claim 1, characterized in that, When performing manifold mesh topology self-checking and repair, the manifold mesh topology correction unit includes: performing mesh edge flipping and topology regularization for non-manifold meshes with a degree less than 3 using a multi-source data fusion 3D modeling system with edge control; and performing geometric facet cutting, splitting, and topology separation for non-manifold meshes with edge control shared by more than two faces using a multi-source data fusion 3D modeling system.

9. A multi-source data fusion 3D modeling system according to claim 1, characterized in that, In the manifold mesh topology correction unit, after recalculating the mesh edge length distortion rate of all connected mesh edges by controlling the multi-source data fusion 3D modeling system every 3 gradient iterations, if it is determined that the mesh edge length distortion rate of all connected mesh edges has returned to within 0.8 to 1.2, the manifold mesh topology correction unit controls the multi-source data fusion 3D modeling system to restore the numerical distribution of the first data point cloud approximation weight and the second data point cloud approximation weight under the data-driven mode.

10. A multi-source data fusion 3D modeling method, used to run the multi-source data fusion 3D modeling system of claim 1, characterized in that, Includes the following steps: Step S1: The nonlinear geometric energy deformation control unit constructs the mesh energy functional and calculates the mesh edge distortion rate of the mesh vertex before updating the mesh vertex coordinates through gradient descent. When the mesh edge distortion rate exceeds the safety boundary, the unit enters the isotropic relaxation mode and calculates the arithmetic mean center coordinates of the first-order neighboring vertices of the mesh vertex as temporary coordinate values. Step S2: The anisotropic curvature tensor evaluation unit extracts the local curvature features of the mesh vertices in the isotropic relaxation mode. When the maximum principal curvature value exceeds the critical curvature threshold of 2.5, the mesh vertices are restricted to complete projective sliding along the minimum principal curvature direction of the anisotropic curvature tensor in the tangent plane. Step S3: The geometric deformation field vector is established by monitoring the coordinate iteration change of the grid vertices by the geometric prior topology filling unit. When the data confidence of the local area is lower than the safety limit threshold of 0.05 and the direction of the geometric deformation field vector is reversed, the data approximation term in the grid energy functional of the corresponding area is cut off, and the grid vertices of the local area are controlled to complete the geometric trend filling according to the grid curvature direction. Step S4: The manifold mesh topology correction unit recalculates the mesh edge distortion rate after updating the mesh vertex coordinates, and switches back to data-driven mode when the mesh edge distortion rate returns to within the safe boundary.

Images