Bridge appearance defect detection system based on three-dimensional point cloud reconstruction

Abstract

This invention relates to the field of 3D modeling technology, specifically to a bridge appearance defect detection system based on 3D point cloud reconstruction. The system includes: a point cloud spatial partitioning module, which constructs a 3D bounding box based on the bridge point cloud, calculates a local geometric degradation index, and compares the local geometric degradation index with a flatness threshold and a degradation threshold to obtain an initial 3D polygon mesh. In this invention, by filtering the normal vector angles of density feature patch regions and extracting polygon contours, local geometric abrupt change regions are separated from the overall structure. Furthermore, a dual filtering process is performed using the rate of curvature change and surface area values, ensuring that defect regions possess clear characteristic expressions in both spatial scale and morphological changes. This improves the accuracy of defect localization and ensures that bridge appearance defects remain identifiable even in complex environments.

Description

Technical Field

[0001] This invention relates to the field of 3D modeling technology, and in particular to a bridge appearance defect detection system based on three-dimensional point cloud reconstruction. Background Technology

[0002] 3D point cloud reconstruction is a branch of the 3D modeling technology field. It involves acquiring spatial discrete point data through laser scanning equipment, structured light equipment, or multi-view image acquisition systems, and then processing the point cloud data such as registration, filtering, feature extraction, surface reconstruction, and mesh generation to restore the 3D geometric shape and spatial structure of the target object.

[0003] Existing technologies primarily focus on the acquisition and geometric reconstruction of point cloud data, emphasizing the improvement of point cloud quality through registration and filtering, and the generation of an overall 3D model through surface reconstruction. However, in practice, the goal is usually overall geometric reconstruction, limiting the ability to express local structural differences. When point clouds contain occlusions or acquisition blind spots, the generated 3D model often exhibits holes or broken areas. These areas lack further repair mechanisms, resulting in incomplete model structures. For example, occluded areas at the bottom or sides of bridges are prone to forming gaps during scanning, affecting the accuracy of subsequent analysis. Furthermore, in the defect identification stage, existing technologies often rely on single geometric features for judgment, lacking comprehensive analysis of normal changes, curvature changes, and area constraints. This makes edge structures or normal construction details easily misidentified as defects. Therefore, improvements are needed. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a bridge appearance defect detection system based on three-dimensional point cloud reconstruction.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a bridge appearance defect detection system based on three-dimensional point cloud reconstruction includes:

[0006] The point cloud spatial subdivision module constructs a three-dimensional bounding box based on the bridge point cloud, calculates the local geometric degradation index, and compares the local geometric degradation index with the flatness threshold and the degradation threshold respectively to obtain the initial three-dimensional polygon mesh.

[0007] The mesh feature extraction module extracts the hole boundaries on the initial three-dimensional polygonal mesh, performs closure processing to obtain the boundary loop, generates bridge symmetry axis parameters, projects the boundary loop onto the opposite side of the initial three-dimensional polygonal mesh along the bridge symmetry axis parameters to extract the corresponding facets, and obtains mirror repair mesh facets.

[0008] The mesh topology stitching module calculates the distance difference between the outer edge vertex and the hole edge vertex based on the initial three-dimensional polygonal mesh and the mirrored repair mesh patch, obtains the edge geometric gap difference value, extracts the difference coordinate value inside the mirrored repair mesh patch, constructs the Laplacian operator matrix equation, and obtains the three-dimensional mesh of the closed bridge.

[0009] The appearance defect location module extracts the area where the density feature patch is located in the three-dimensional mesh of the closed bridge, calculates the angle between the normal vectors of all adjacent patches in the area, connects the adjacent patches whose normal vector angle is greater than the angle judgment threshold to the outside to obtain the polygon feature contour line, calculates the rate of curvature change and surface area of ​​the polygon feature contour line, and filters to obtain the coordinate set of bridge appearance defects.

[0010] Preferably, the step of obtaining the initial three-dimensional polygon mesh is as follows:

[0011] A 3D bounding box is constructed based on the bridge point cloud and an octree is performed. The point cloud coordinates within each leaf node are located one by one. The normal vector of the fitting plane, the eigenvalue of the covariance matrix, and the offset of each point from the local fitting plane are extracted. The geometric parameters within the same leaf node are collected to form a geometric parameter group of the leaf node.

[0012] Calculate the local geometric degradation index based on the set of geometric parameters of the leaf nodes;

[0013] Each local geometric degradation index is determined by interval matching with the flatness threshold and degradation threshold. Leaf nodes with values ​​less than the flatness threshold are extracted as scale patches. Leaf nodes within the flatness threshold and degradation threshold intervals are further subdivided. Leaf nodes with values ​​greater than or equal to the degradation threshold are subdivided to the highest level and extracted as density feature patches. The scale patches and density feature patches are combined to form an initial three-dimensional polygon mesh.

[0014] Preferably, the steps for obtaining the bridge's axis of symmetry parameters are as follows:

[0015] Based on the initial three-dimensional polygonal mesh, open edges that are shared only by a single mesh face are retrieved one by one. Open edge chains are spliced ​​according to the connection order of the endpoints of the open edges. The closure relationship between the first and last vertices of the open edge chain is verified. For open edge chains with discontinuities between the first and last vertices, closed edges are added. The order of all vertices on the closed edges is rearranged according to a unified wrapping direction to form a boundary loop.

[0016] Traverse all vertices on the initial three-dimensional polygonal mesh, record the mapping coordinates of the vertex normal vector on the unit sphere point by point, extract the principal curvature of the vertex point by point, construct the covariance matrix of the Gaussian mapping lattice according to all the mapping coordinates, extract the eigenvectors of the Gaussian mapping lattice covariance matrix as principal components, and calibrate the bridge symmetry axis parameters according to the directional distribution of the principal components.

[0017] Preferably, the step of obtaining the mirror-repaired mesh patch is as follows:

[0018] Next, the boundary ring is projected onto the opposite side of the initial three-dimensional polygon mesh along the bridge symmetry axis parameter. The mesh surfaces within the coverage area of ​​the projection landing point are screened to form a corresponding rigid transformation group. The vertex coordinates, bridge symmetry axis parameters, and spatial pose offset relationship of the corresponding face in the rigid transformation group are read. The rigid transformation matrix from the vertex of the corresponding face to the position inside the boundary ring is established point by point. The corresponding face is projected and copied into the boundary ring according to the rigid transformation matrix. Then, the position of the copied mesh surface boundary is matched and the vertex order is sorted to obtain the mirror repair mesh surface.

[0019] Preferably, the step of obtaining the edge geometric gap difference value is as follows:

[0020] Based on the initial 3D polygonal mesh and the mirrored repair mesh patch, locate all hole edge vertices in the hole region where the boundary ring is located, locate all outer edge vertices on the outer contour of the mirrored repair mesh patch, arrange the hole edge vertices in sequence according to the boundary ring's wrapping direction, arrange the outer edge vertices in sequence according to the connection order of the outer contour of the mirrored repair mesh patch, read the 3D coordinates of each group of outer edge vertices and hole edge vertices one by one, calculate the spatial distance difference between each group of corresponding vertices, and write all spatial distance differences into the boundary ring position according to the vertex order to generate edge geometric gap difference values.

[0021] Preferably, the steps for obtaining the three-dimensional mesh of the closed bridge are as follows:

[0022] Extract the coordinate offsets of all vertices inside the mirror-repaired mesh relative to their adjacent vertices. Arrange each set of coordinate offsets according to the adjacency relationship of the vertices inside the mirror-repaired mesh and write them into the differential coordinate values ​​inside the mirror-repaired mesh. Construct a Laplacian operator matrix equation based on the connection relationship between the hole edge vertices, the outer edge vertices, and the vertices inside the mirror-repaired mesh. Write the differential coordinate values ​​inside the mirror-repaired mesh into the constant term constraint position of the Laplacian operator matrix equation. Then write the edge geometric gap differential value into the boundary constraint position of the Laplacian operator matrix equation. Solve the coordinate components corresponding to each vertex row by row to generate the edge transition coordinates.

[0023] The positions of the hole edge vertices corresponding to the edge transition coordinates are read one by one. The original three-dimensional coordinates of the corresponding hole edge vertices in the initial three-dimensional polygon mesh are replaced with the edge transition coordinates. The boundary vertices of the mirror repair mesh are reconnected according to the updated hole edge vertex positions. The connection order of the adjacent patches of the hole edge vertices is corrected. The splicing relationship of the triangular patches of the boundary of the mirror repair mesh is corrected. The coordinate breaks and topological breaks at the hole edge positions are eliminated to obtain the closed bridge three-dimensional mesh.

[0024] Preferably, the step of obtaining the polygonal feature contour line is as follows:

[0025] Based on the closed bridge 3D mesh, all mesh faces within the area where the density feature facet is located are located. Adjacent facets are extracted one by one according to the shared edge relationship. The normal vector component of each mesh facet is read. The angle between the normal vectors of each pair of adjacent facets is calculated. Adjacent facets with a normal vector angle greater than the angle judgment threshold are retained as candidate connection pairs. The first and last connections are performed according to the continuous distribution order of the candidate connection pairs in the outer contour direction to eliminate duplicate connection edges, retain the outer envelope boundary, and generate polygon feature contour lines.

[0026] Preferably, the steps for obtaining the coordinate set of bridge appearance defects are as follows:

[0027] Extract all grid faces within the area enclosed by the polygon feature contour lines one by one, count the change range of curvature values ​​of adjacent grid faces one by one, accumulate the area values ​​of each grid face according to the enclosed range of the polygon feature contour lines, and write the curvature change rate and surface area value of the corresponding enclosed area respectively. Keep the enclosed areas where the curvature change rate is greater than the curvature threshold and the surface area value falls within the peeling area range, and remove the remaining enclosed areas to form a polygon set.

[0028] Based on the polygon set, the three-dimensional spatial coordinates of all vertices of each polygon are read one by one. The position of the center vertex is calculated according to the spatial distribution of the three-dimensional spatial coordinates of all vertices within the same polygon. The three-dimensional spatial coordinates of the center vertex corresponding to each polygon are extracted and written into the defect location result sequence in the traversal order of the polygon set. Duplicate coordinate items in the defect location result sequence are checked and deduplicated to obtain the bridge appearance defect coordinate set.

[0029] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0030] In this invention, the point cloud of a bridge is spatially bounding box divided and different regions are graded by combining the local geometric degradation index. This transforms the point cloud from its original discrete state into a polygonal representation with structural hierarchy, enhancing the ability to distinguish local abnormal regions while maintaining overall geometric continuity. Furthermore, by closing the hole boundaries and extracting boundary loops, and combining the inherent symmetry information of the bridge structure, the bridge's symmetry axis parameters are generated. This enables the mapping and mirror repair of the opposite side panels of missing regions, filling in areas with gaps or missing data while maintaining structural consistency. Subsequently, by jointly constraining the distance difference between edge vertices and the internal differential coordinate values, a coordinate transition relationship in continuous space is constructed, making the connection between the repaired region and the original mesh smoother and avoiding geometric distortion caused by abrupt boundaries. On this basis, by filtering the normal vector angle of the density feature patch region and extracting the polygon contour, local geometric abrupt regions are separated from the overall structure. This is further filtered by combining the rate of curvature change and the surface area value, ensuring that the defective region has clear feature expression in terms of spatial scale and morphological changes. This improves the accuracy of defect localization and ensures that bridge appearance defects remain identifiable in complex environments. Attached Figure Description

[0031] Figure 1 This is a system flowchart of the present invention;

[0032] Figure 2 A 3D bounding box and octree partition diagram of the bridge point cloud;

[0033] Figure 3 This is a simulation diagram of the scatter plot distribution of the local geometric degradation index. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0035] Please see Figure 1-3 The present invention provides a technical solution: a bridge appearance defect detection system based on three-dimensional point cloud reconstruction, comprising:

[0036] The point cloud spatial subdivision module constructs a 3D bounding box based on the bridge point cloud, calculates the local geometric degradation index, and compares the local geometric degradation index with the flatness threshold and degradation threshold respectively to obtain the initial 3D polygon mesh.

[0037] The mesh feature extraction module extracts the hole boundaries on the initial 3D polygon mesh, performs closure processing to obtain the boundary loop, generates the bridge symmetry axis parameters, projects the boundary loop onto the opposite side of the initial 3D polygon mesh along the bridge symmetry axis parameters to extract the corresponding facets, and obtains the mirrored repair mesh facets.

[0038] The mesh topology stitching module calculates the distance difference between the outer edge vertex and the hole edge vertex based on the initial 3D polygon mesh and the mirrored repair mesh patch, obtains the edge geometric gap difference value, extracts the difference coordinate value inside the mirrored repair mesh patch, constructs the Laplacian operator matrix equation, and obtains the 3D mesh of the closed bridge.

[0039] The appearance defect location module extracts the area where the density feature patch is located in the three-dimensional mesh of the closed bridge, calculates the angle between the normal vectors of all adjacent patches in the area, connects the adjacent patches whose normal vector angle is greater than the angle judgment threshold to obtain the polygon feature contour line, calculates the rate of curvature change and surface area of ​​the polygon feature contour line, and filters to obtain the coordinate set of bridge appearance defects.

[0040] The steps to obtain the initial 3D polygon mesh are as follows:

[0041] A 3D bounding box is constructed based on the bridge point cloud and an octree is performed. The point cloud coordinates within each leaf node are located one by one. The normal vector of the fitting plane, the eigenvalue of the covariance matrix, and the offset of each point from the local fitting plane are extracted. The geometric parameters within the same leaf node are collected to form a geometric parameter group of the leaf node.

[0042] Based on the geometric parameter set of the leaf nodes, the local geometric degradation index is calculated using the following formula:

[0043] ;

[0044] in, Let be the local geometric degradation index of the current leaf node. The first eigenvalue of the covariance matrix of the current leaf node, sorted in ascending order. The second eigenvalue of the covariance matrix of the current leaf node, sorted in ascending order. The third eigenvalue of the covariance matrix of the current leaf node after sorting it from smallest to largest. This represents the number of points within the current leaf node that participate in the offset calculation. The first leaf node in the current leaf node The offset of a point from the local fitting plane. The value is the square of the third eigenvalue of the covariance matrix of the current leaf node;

[0045] Each local geometric degradation index is determined by interval matching with the flatness threshold and degradation threshold. Leaf nodes with values ​​less than the flatness threshold are extracted as scale patches. Leaf nodes within the flatness threshold and degradation threshold intervals are further subdivided. Leaf nodes with values ​​greater than or equal to the degradation threshold are subdivided to the highest level and extracted as density feature patches. The scale patches and density feature patches are then combined to form the initial three-dimensional polygon mesh.

[0046] Specifically, the process involves acquiring the 3D point cloud data of the bridge, traversing all point cloud coordinates to find the maximum and minimum values ​​along the three coordinate axes. For example, if a bridge's point cloud ranges from 0 to 150 meters on the X-axis, from -10 meters to 10 meters on the Y-axis, and from 0 to 30 meters on the Z-axis, a 3D bounding box with a length of 150 meters, a width of 20 meters, and a height of 30 meters is constructed using these extreme boundaries. This bounding box is used as the initial root node. An octree is then used to partition the space according to the rule of dividing each of the three dimensions into two equal halves. Each partition generates 8 subspaces, and the number of point clouds contained in each subspace is calculated. The maximum node capacity is set to 50 points. When the number of points in a space exceeds 50, the partitioning continues to the next level. As the level deepens, the spatial size continuously shrinks until the number of points is less than or equal to 50 or the maximum partitioning depth of 10 levels is reached. The node where partitioning stops is designated as a leaf node. For each leaf node whose physical size has shrunk to approximately 0.5 meters, the 3D coordinates of all point clouds contained within it are read, and the maximum value is calculated. The centroid coordinates of these points are obtained by summing their three-dimensional coordinates and dividing by 40, for example, within a leaf node with 40 points. Centering is performed on all points with the centroid as the origin, constructing a three-dimensional covariance matrix composed of these 40 points. Eigenvalue decomposition is performed on this covariance matrix to obtain three eigenvalues ​​representing the spatial distribution variance and their corresponding eigenvectors. These three eigenvalues ​​are sorted in ascending order, and the eigenvector corresponding to the smallest eigenvalue is the normal vector of the local fitting plane of the leaf node. The spatial equation of the local fitting plane is determined based on the normal vector and the centroid coordinates. The 40 points within the leaf node are traversed, and the perpendicular distance from each point to the local fitting plane is calculated, serving as the offset of each point from the local fitting plane. The three components of the normal vector extracted from the current leaf node, the three sorted eigenvalues ​​of the covariance matrix, and the offsets corresponding to each of the 40 points are summarized and sequentially written into a one-dimensional array. This process is used to aggregate the geometric parameters within the same leaf node, forming a leaf node geometric parameter group.

[0047] In the formula for calculating the local geometric deterioration index, the ratio of the first eigenvalue to the third eigenvalue of the covariance matrix is ​​introduced to evaluate the flatness of the main body of the point cloud on a macro scale. At the same time, the fourth central moment of the offset of the point from the fitting plane is introduced as a penalty term. The fourth moment has a strong ability to capture abnormal deviation points that are far from the plane, and can lock the tiny depressions or protrusions on the bridge surface. The secondary principal eigenvalue and the square of the principal eigenvalue are used for dimension unification and normalization, which weakens the influence of the uneven point cloud scanning density on the calculation results and accurately quantifies the degree of local geometric deterioration in complex point cloud environments.

[0048] The first eigenvalue of the covariance matrix of the current leaf node, sorted from smallest to largest, represents the variance of the point cloud in the direction of the normal vector of the local fitting plane in point cloud geometric feature analysis. It indicates the thickness distribution range of the point cloud deviating from the fitting plane, with units of square meters. The specific steps to obtain this parameter are: read the 3D coordinates of all points within the current leaf node, calculate the centroid coordinates of these points, decentralize all points using the centroid as the origin to construct a 3D covariance matrix, and use the Jacobi iteration algorithm to perform eigenvalue decomposition on this 3D covariance matrix to obtain three non-negative variance eigenvalues. These three eigenvalues ​​are then sorted in ascending order, and the smallest value in the first position is the first eigenvalue. For example, if the physical size of a leaf node on the surface of a bridge is approximately 0.5 meters and the internal distribution is relatively flat, the three eigenvalues ​​obtained after covariance matrix decomposition are 0.0004, 0.01, and 0.04. After sorting in ascending order, the smallest value is extracted as the first eigenvalue, thus obtaining the first eigenvalue of the covariance matrix of the current leaf node sorted from smallest to largest. It equals 0.0004;

[0049] The second eigenvalue of the current leaf node covariance matrix, sorted from smallest to largest, represents the variance distribution of the point cloud along the secondary principal component direction within the local fitting plane. It reflects the dispersion span of the local point cloud in the width direction, also measured in square meters. Its acquisition process is the same as the first eigenvalue: by centering all point cloud coordinates within the leaf node and constructing a three-dimensional covariance matrix, matrix decomposition using the Jacobi iterative algorithm yields three eigenvalues. These three eigenvalues ​​are then sorted in ascending order, and the value in the middle position is extracted as the second eigenvalue. For example, after performing covariance matrix decomposition on the aforementioned leaf node point cloud and obtaining eigenvalues ​​0.0004, 0.01, and 0.04, they are sorted in ascending order, and the value 0.01, located in the second position, is selected. This yields the second eigenvalue of the current leaf node covariance matrix, sorted from smallest to largest. It equals 0.01;

[0050] The third eigenvalue of the covariance matrix of the current leaf node, sorted from smallest to largest, represents the variance of the point cloud along the principal component directions within the local fitting plane. It reflects the maximum extension of the point cloud along its length, measured in square meters. This value is obtained based on the eigenvalue decomposition results of the same covariance matrix. After constructing the covariance matrix for the point cloud coordinate set within the leaf node, three eigenvalues ​​are obtained through eigenvalue decomposition. These three eigenvalues ​​are then sorted in ascending order, and the largest value is extracted as the third eigenvalue. For example, if the three eigenvalues ​​obtained from the aforementioned decomposition are 0.0004, 0.01, and 0.04, the largest value at the end (0.04) is extracted after sorting to obtain the third eigenvalue of the covariance matrix of the current leaf node, sorted from smallest to largest. It equals 0.04;

[0051] : The number of points within the current leaf node that participate in the offset calculation. This parameter is a dimensionless count value used for averaging when calculating higher-order moments. For example, by iterating and counting the point cloud records inside a leaf node after octree partitioning, the counter eventually stops at 40, thus obtaining the number of points within the current leaf node that participate in the offset calculation. Equals 40;

[0052] The first leaf node in the current leaf node The offset of a point from the local fitted plane reflects the vertical distance of a single discrete point from the local bridge surface geometry, measured in meters. It is used to capture small local bumps or depressions. For example, if the current normal vector is along the Z-axis, the centroid coordinates are at the origin, and the plane equation is Z = 0, and the Z-axis value of the first point is 0.02, the vertical distance from this point to the plane is calculated to be 0.02 meters, thus yielding the offset of that point. The value equals 0.02. This process is repeated for all 40 points to obtain their offsets. The formula requires calculating the sum of the fourth powers of the offsets, converting the units to fourth power meters. The fourth power of each point's offset is calculated separately and then summed. For example, the fourth power of a single point is approximately 0.00000016. The sum of the offsets from all 40 points is then obtained. It equals 0.0000064;

[0053] : The square of the third eigenvalue of the covariance matrix of the current leaf node. This parameter serves as the normalization denominator in the formula, ensuring comparability of calculation results for regions of different sizes. The steps to obtain it are: directly read the third eigenvalue already calculated and extracted in the previous steps, and perform a series of squared multiplications on this value. For example, if the previous steps have already obtained the third eigenvalue of the covariance matrix of the current leaf node sorted from smallest to largest... Equal to 0.04, substituting it into the squaring operation, i.e., multiplying 0.04 by 0.04, yields the square of the third eigenvalue of the covariance matrix of the current leaf node. It equals 0.0016;

[0054] Calculations based on parameters:

[0055] Will , , , , , Substitute into the formula;

[0056] Calculate the mean of the fourth power of the offset. ;

[0057] Calculate the square root of the mean. ;

[0058] Calculate the ratio in the first part of the formula. ;

[0059] Calculate the ratio in the second part of the formula. ;

[0060] Calculate the final local geometric degradation index. ;

[0061] The results show that the local geometric degradation index of the current leaf node is 0.0125. This small value indicates that the point cloud distribution inside the current leaf node is relatively flat, with a very low degree of dispersion from the local fitting plane. It does not exhibit geometrical abrupt changes such as cracks or peeling, and can be classified as a flat and healthy bridge surface area. This result will serve as the basis for subsequent threshold interval determination and mesh subdivision operations.

[0062] The same octree partitioning and index calculation were performed on healthy point cloud data from 10 known defect-free smooth bridge deck areas and 10 point cloud data containing cracks and spalling defects. Statistical calculations showed that the local geometric deterioration index of smooth areas ranged from 0.005 to 0.015, while that of defective areas ranged from 0.08 to 0.15. The maximum index of smooth areas (0.015) was taken as the smoothness threshold, and the minimum index of defective areas (0.08) was taken as the deterioration threshold. The local geometric deterioration index corresponding to each leaf node was extracted. For example, if the local geometric deterioration index of the current leaf node was 0.0125, this value was compared with the smoothness threshold (0.015) and the deterioration threshold (0.08). Since the current index (0.0125) was less than the smoothness threshold (0.015), the area where the leaf node was located was determined to be smooth and without structural damage. The point cloud and its local fitted plane are extracted into large-area scale patches. For leaf nodes with a local geometric degradation index of 0.05, the value is between 0.015 and 0.08, indicating that there is potential irregular deformation in the region but it has not reached the degree of degradation. The region is further subdivided into 8 subspaces according to the rule of dividing the bounding box corresponding to the node into 8 subspaces. For leaf nodes with a local geometric degradation index of 0.12, the value is greater than 0.08, indicating that the region contains obvious surface defect features. The region is subdivided three times according to the rule of dividing into eight layers to reach the highest level of 10 layers. The small leaf nodes containing point cloud coordinates after subdivision are retained as density feature patches. All scale patches generated in flat areas and density feature patches generated in degraded areas are combined and assembled in the three-dimensional coordinate system according to the topological adjacency relationship of vertex coordinates to form an initial three-dimensional polygon mesh.

[0063] The steps for obtaining the bridge's axis of symmetry parameters are as follows:

[0064] Based on the initial 3D polygon mesh, open edges that are shared by only a single mesh face are retrieved one by one. Open edge chains are spliced ​​according to the connection order of the endpoints of the open edges. The closure relationship between the first and last vertices of the open edge chain is verified. For open edge chains with discontinuities between the first and last vertices, closed edges are added. The order of all vertices on the closed edges is rearranged according to a unified wrapping direction to form a boundary loop.

[0065] Traverse all vertices on the initial 3D polygonal mesh, record the mapping coordinates of the vertex normal vector on the unit sphere point by point, extract the principal curvature of the vertex point by point, construct the covariance matrix of the Gaussian mapping lattice according to all the mapping coordinates, extract the eigenvectors of the Gaussian mapping lattice covariance matrix as principal components, and calibrate the bridge symmetry axis parameters according to the directional distribution of the principal components.

[0066] Specifically, the initial 3D polygon mesh is extracted, and the edge data structure of all triangle faces in the mesh is traversed. The number of triangle faces associated with each edge is counted, and edges with an associated face count of 1 are marked as open edges. For example, if an edge belongs only to triangle A and is not shared by an adjacent triangle B, it is extracted. Then, the coordinates of the two endpoints of all marked open edges are read, a starting edge is set, and its last vertex is used as the current matching point. The next open edge with the matching point as the starting vertex is found in the remaining open edges. This first-to-last-connected open edge chain is formed by sequentially splicing multiple open edges. The splicing stops when no next adjacent open edge is found. Then, the 3D coordinates of the first and last vertices of the open edge chain are extracted, the 3D Euclidean distance between these two vertices is calculated, and this distance is compared with the closure tolerance threshold. The closure tolerance threshold is based on the mesh. The average side length is set. For example, the average length of all sides of the initial 3D polygon mesh is calculated to be 0.05 meters. 10% of this, i.e., 0.005 meters, is taken as the closure tolerance threshold. If the distance between the first vertex and the last vertex is 0.02 meters, which is greater than the threshold of 0.005 meters, it is determined that there is a discontinuity between them. For such an unclosed open edge chain, a new ruled line segment connecting the first vertex and the last vertex is created as a supplementary closed edge. This is added to the open edge chain to form a closed topological outer loop. Then, any vertex on the closed loop is selected as the starting point. The normal vector direction of the closed loop is calculated using the right-hand rule of vector cross product. According to the geometric principle that the calculated normal vector direction is consistent with the outward normal vector of the mesh surface, all vertices on the closed loop are rearranged in a counterclockwise unified wrapping direction to eliminate the vertex sequence disorder and interlacing flipping problems that occurred in the original splicing process, forming a boundary loop.

[0067] Based on the initial 3D polygon mesh obtained above, all discrete vertices on the mesh surface are traversed. For each vertex, all connected polygon faces are found, and the normal vectors of these adjacent faces are extracted. A weighted average of these normal vectors, using the area of ​​the adjacent faces as weights, is calculated to obtain the vertex's own normal vector. The starting point of the obtained vertex normal vector is then translated to the origin of the 3D coordinate system, so that its endpoint falls on a unit sphere with a radius of 1. The X, Y, and Z coordinate components of this endpoint in 3D space are read and used as the Gaussian mapping coordinates of the vertex on the unit sphere, and the values ​​are recorded. Simultaneously, the minute shape of the local surface is estimated based on the spatial positional relationship between the vertex and its surrounding vertices. The curvature value corresponding to the direction with the greatest curvature of the vertex surface is calculated as the vertex's principal curvature. After traversing all vertices on the mesh, the Gaussian mapping coordinates of all vertices are collected to form a lattice, and all... The overall centroid position of the mapped coordinates is determined by subtracting the centroid coordinate from each mapped coordinate for centering. A third-order Gaussian mapped point lattice covariance matrix is ​​constructed using the centered coordinate components. The singular value decomposition algorithm is then used to decompose this covariance matrix, extracting the eigenvectors corresponding to the largest singular values ​​as the first principal components and the eigenvectors corresponding to the second largest singular values ​​as the second principal components. Since the Gaussian mapping of surface normal vectors of structures like bridges, which have slender extension characteristics, is usually distributed along a great circle perpendicular to the extension direction of the bridge, the annular normal direction of the most concentrated distribution of this Gaussian mapped point lattice on a unit sphere is extracted. Combined with the previously extracted vertex principal curvature directions, eigenvectors parallel to the direction of the minimum change in the principal curvature of the mesh are selected. The pointing angle and offset distance of these eigenvectors in the three-dimensional space system are used to derive the parameters of the three-dimensional spatial symmetry plane equation of the bridge. The bridge symmetry axis parameters are then calibrated based on the directional distribution of the principal components.

[0068] The steps to obtain the mirror-repaired mesh patch are as follows:

[0069] Next, the boundary ring is projected onto the opposite side of the initial 3D polygon mesh along the bridge's symmetry axis parameters. Mesh surfaces within the coverage area of ​​the projection landing point are selected to form corresponding rigid transformation groups. The vertex coordinates, bridge symmetry axis parameters, and spatial pose offset relationships of the corresponding facets in the rigid transformation groups are read. A rigid transformation matrix from the vertex of the corresponding facet to the position inside the boundary ring is established point by point. The corresponding facet is projected and copied into the boundary ring according to the rigid transformation matrix. Then, the copied mesh surface boundary is aligned and the vertex order is adjusted to obtain a mirror-repaired mesh facet.

[0070] Specifically, the bridge's symmetry axis parameters and boundary ring obtained earlier are acquired. A spatial mirror symmetry plane based on the bridge's symmetry axis parameters is constructed in the 3D coordinate system. The 3D coordinates of all vertices contained in the boundary ring are mathematically mirrored and reflected according to this mirror symmetry plane to obtain the coordinates of the mirror landing points projected onto the opposite side of the initial 3D polygonal mesh. The 2D bounding box region of the polygonal mesh enclosed by these mirror landing points is calculated. All mesh patches of the initial 3D polygonal mesh on the opposite side region are traversed. The centroid coordinates of the three vertices of each patch are checked to see if they fall within the 2D bounding box region. Patches with their centroids within the coverage area are extracted, summarized, and grouped to form corresponding patch rigid transformation groups. The spatial coordinates of the three vertices of each patch in the corresponding patch rigid transformation group are read. Combining the bridge's symmetry axis parameters and the spatial 3D pose offset relationship before and after mirroring, a 3D rotation matrix and a 3D translation matrix are constructed for each corresponding patch vertex. The fourth-order rigid transformation matrix is ​​used to obtain the specific parameters of the rigid transformation matrix by solving the least squares solution of the corresponding point pairs before and after the mirror image. For example, the vertices of the opposite side panel are rotated 180 degrees by reverse mirroring and translated by the corresponding offset distance along the axis of symmetry to obtain the specific matrix values. The corresponding facets selected by the established rigid transformation matrix are subjected to matrix multiplication and addition operations on the coordinates of each vertex. The corresponding facets are projected from the opposite side of the bridge and copied to the position inside the hole space where the boundary ring is located. At this time, there may be a small gap between the edge of the copied mesh and the boundary ring. The outer edge vertices of the copied mesh are traversed, and the Euclidean space distance between each vertex and the nearest vertex on the boundary ring is calculated. The edge vertices of the copied mesh are moved, snapped, and aligned to the coordinate position of the nearest vertex of the boundary ring. Finally, the spatial index of all the vertices after the meshing is renumbered and the connection relationship of the facets is updated according to the original counterclockwise wrapping direction of the boundary ring to obtain the mirrored repaired mesh facets.

[0071] The steps for obtaining the edge geometric gap difference value are as follows:

[0072] Based on the initial 3D polygonal mesh and the mirrored repair mesh patch, locate all the edge vertices of the holes in the hole region where the boundary ring is located, locate all the outer edge vertices on the outer contour of the mirrored repair mesh patch, arrange the hole edge vertices in order according to the direction of the boundary ring, arrange the outer edge vertices in order according to the connection order of the outer contour of the mirrored repair mesh patch, read the 3D coordinates of each group of outer edge vertices and hole edge vertices one by one, calculate the spatial distance difference between each group of corresponding vertices, and write all the spatial distance differences into the boundary ring position in vertex order to generate edge geometric gap difference values.

[0073] Specifically, the initial 3D polygonal mesh and the mirrored repair mesh obtained in the previous steps are extracted. The half-side data structure of the initial 3D polygonal mesh is traversed to find the topological edges surrounding the hole region on the bridge surface that are associated only by a single polygonal patch. All endpoints on these edges are extracted as hole edge vertices. Simultaneously, the independently generated mirrored repair mesh is traversed, and the same rules are used to find its unclosed outer contour line. All vertices on the contour line are extracted as outer edge vertices. Any vertex on the hole edge is selected as the sorting starting point. Using the surface normal vector determined in the previous steps and the right-hand screw rule, the counter-clockwise traversal direction of the boundary loop is determined. Adjacent hole edge vertices are traced along this direction, and their corresponding index numbers are recorded, completing the sequential arrangement of hole edge vertices. Using a consistent starting point determination rule and a counter-clockwise circling direction, the vertex index numbers are traced and recorded sequentially along the outer contour of the mirrored repair mesh. Arrange the vertices of the outer edge in order. Considering that the number of vertices of the outer edge and the vertices of the hole edge are exactly the same after the mirror transformation, establish a one-to-one mapping relationship according to their respective arranged order. Extract the first pair of mapped vertices, read the three-dimensional space X, Y, Z coordinate values ​​of the corresponding outer edge vertex, and read the three-dimensional space X, Y, Z coordinate values ​​of the corresponding hole edge vertex. Calculate the subtraction results on the same coordinate axes, that is, subtract the X coordinate of the hole edge vertex from the X coordinate of the outer edge vertex, and subtract the Y coordinate and Z coordinate in the same way to obtain the three-dimensional vector difference containing the three coordinate components. Calculate the remaining mapped vertex pairs in the same way, and calculate the spatial distance difference between each pair of corresponding vertices. Finally, according to the vertex index order on the boundary ring, allocate a contiguous memory space, and fill all the calculated three-dimensional vector differences into the memory space in the order of the corresponding vertices to generate the edge geometric gap difference value.

[0074] The steps for obtaining the 3D mesh of a closed bridge are as follows:

[0075] Extract the coordinate offsets of all vertices inside the mirror-repaired mesh relative to their adjacent vertices. Arrange each set of coordinate offsets according to the adjacency relationship of the vertices inside the mirror-repaired mesh and write them into the differential coordinate values ​​inside the mirror-repaired mesh. Construct a Laplacian operator matrix equation based on the connection relationship between the vertices of the hole edge, the outer edge, and the vertices inside the mirror-repaired mesh. Write the differential coordinate values ​​inside the mirror-repaired mesh into the constant term constraint position of the Laplacian operator matrix equation. Then write the differential values ​​of the edge geometric gaps into the boundary constraint position of the Laplacian operator matrix equation. Solve the coordinate components corresponding to each vertex row by row to generate the edge transition coordinates.

[0076] Read the vertex number of the hole edge corresponding to the edge transition coordinate one by one, replace the original three-dimensional coordinates of the corresponding hole edge vertex in the initial three-dimensional polygon mesh with the edge transition coordinate, reconnect the boundary vertices of the mirror repair mesh according to the updated hole edge vertex positions, correct the connection order of the adjacent facets of the hole edge vertex, correct the splicing relationship of the triangular facets of the boundary of the mirror repair mesh, eliminate the coordinate break and topological break at the hole edge position, and obtain the three-dimensional mesh of the closed bridge.

[0077] Specifically, the mirror-patched mesh created in the previous steps is invoked. All non-edge discrete vertices within the mesh are traversed. For each currently traversed internal vertex, all adjacent vertices connected by mesh edges are found. The 3D coordinates of the current vertex and all adjacent vertices are read. The 3D coordinates of all adjacent vertices are summed and divided by the number of adjacent vertices to obtain the geometric mean coordinates. The original 3D coordinates of the current vertex are subtracted from this geometric mean coordinate to obtain the spatial vector difference containing X, Y, and Z dimensions, which is used as the coordinate offset of the vertex relative to its adjacent vertices. After completing the calculation for all internal vertices, a one-dimensional array corresponding to the vertex index number is created. The coordinate offsets of each group are stored in the array in numerical order. The internal difference coordinate values ​​of the mirror-patched mesh are written, and a system is constructed based on graph theory. A square matrix is ​​constructed, where the number of rows and columns equals the total number of vertices at the edge of the hole and the vertices of the repair patch. All vertex pairs are traversed; if the two vertices represented by the row and column numbers are adjacent, a negative one is entered at that position; otherwise, a zero is entered. The diagonal positions of the square matrix are filled with the total number of adjacent connecting edges of the corresponding vertex, i.e., the degree of the vertex. This constructs the Laplacian operator matrix equation. The column vector region on the right side of this matrix equation is extracted. The internal difference coordinate values ​​recorded in the aforementioned one-dimensional array are filled into the constant term constraint positions of the corresponding internal vertex numbers. The previously obtained edge geometric gap difference values ​​are filled into the boundary constraint positions of the corresponding hole edge vertex numbers. The Gaussian elimination algorithm is used to solve this sparse linear system of equations containing constraints, calculating the value of each element in the unknown vector in turn. The solved three-dimensional values ​​are assigned to the X, Y, and Z coordinate dimensions to generate edge transition coordinates.

[0078] Read the edge transition coordinates and corresponding hole edge vertex index numbers from the previous steps. Locate the original hole edge vertex data segments in the memory data block of the initial 3D polygon mesh based on these index numbers. Erase the original X, Y, and Z 3D coordinate data of the original hole edge vertices and overwrite the newly generated edge transition coordinate data into the memory address according to the corresponding index numbers, completing the coordinate position replacement and update. Then, traverse the hole boundary edges of the initial 3D polygon mesh and the outer boundary edges of the mirrored repair mesh facets, searching for adjacent edges formed by pairs of vertices with the same name. Point the adjacency pointer of the mirrored repair mesh facet boundary edge to the interior of the corresponding hole boundary edge of the initial 3D polygon mesh. The initial mesh is then modified so that the adjacency pointers of the hole boundary edges are reversed to point to the inner facets of the mirror-repaired mesh facets. The two meshes are then connected and merged. Following the direction of the original boundary ring, the arrangement order of the polygonal facets around each vertex after connection is checked sequentially. If the order of the three vertices of the same face does not conform to the counterclockwise rule of the outward normal, the order of the last two vertices is swapped. The connection order of the adjacent facets of the hole edge vertices is corrected. The connection area is checked for any hanging edges or overlapping facets. Redundant triangular facets sharing three vertices are removed. The splicing relationship of the triangular facets at the boundary of the mirror-repaired mesh facets is corrected. The coordinate breaks and topological breaks at the hole edge are eliminated, resulting in a closed bridge 3D mesh.

[0079] The steps for obtaining the polygon feature contour are as follows:

[0080] Based on the 3D mesh of the closed bridge, all mesh faces within the region where the density feature facet is located are located. Adjacent facets are extracted one by one according to the shared edge relationship. The normal vector components of each mesh facet are read. The angle between the normal vectors of each pair of adjacent facets is calculated. Adjacent facets with a normal vector angle greater than the angle judgment threshold are retained as candidate connection pairs. The first and last connections are performed according to the continuous distribution order of the candidate connection pairs in the outer contour direction to eliminate duplicate connection edges, retain the outer envelope boundary, and generate polygon feature contour lines.

[0081] Specifically, the spatial coordinate range of the density feature patches marked in the 3D mesh of the closed bridge is read. All mesh faces are retrieved within this coordinate range, and the three vertex indices of each mesh face are read. The vertex indices of all mesh faces are compared. If two mesh faces have two identical vertex indices, they are determined to share an edge, and these are extracted as a pair of adjacent patches. The components of the normal vectors of each pair of adjacent patches in the three spatial dimensions are read. The cosine value is obtained by calculating the dot product of the two normal vectors and dividing it by the product of their respective moduli. Then, the spatial angle between the two normal vectors is calculated using the inverse cosine function. The angle between the normal vectors is calculated for 50 pairs of adjacent patches in the normally flat bridge area. The average angle of the flat surface is approximately 5 degrees. Considering the allowable tolerance for minor surface ripples, an additional 3 degrees is added to this. Angle determination threshold, i.e., setting the angle determination threshold to 8 degrees, compares the calculated angle between the normal vectors of each pair of adjacent facets with 8 degrees. When the angle between the normal vectors is greater than 8 degrees, it indicates that there is a significant geometric transformation between the two adjacent facets, and they are retained as candidate connection pairs. After traversing all adjacent facets to complete the determination, the shared edges of all candidate connection pairs are extracted. The endpoints of these shared edges are found in 3D space, and the shared edges are sequentially spliced ​​together along a single direction surrounding the outer contour of these density feature facets in the manner of connecting the endpoints of the shared edges end to end. In this splicing process, if multiple duplicate shared edges are found between two vertices, only the outermost one is retained, and all redundant internal overlapping edges are deleted. The outermost closed line is used as the outer envelope boundary to generate the polygon feature contour line.

[0082] The steps for obtaining the coordinate set of bridge appearance defects are as follows:

[0083] Extract all grid faces within the area enclosed by the polygon feature contour lines one by one, count the change range of the curvature values ​​of adjacent grid faces one by one, accumulate the area values ​​of each grid face according to the enclosed range of the polygon feature contour lines, and write the curvature change rate and surface area value of the corresponding enclosed area respectively. Keep the enclosed areas where the curvature change rate is greater than the curvature threshold and the surface area value falls within the peeling area range, and remove the remaining enclosed areas to form a polygon set.

[0084] Based on the polygon set, read the three-dimensional spatial coordinates of all vertices of each polygon one by one. Calculate the position of the center vertex according to the spatial distribution of the three-dimensional spatial coordinates of all vertices within the same polygon. Extract the three-dimensional spatial coordinates of the corresponding center vertex of each polygon. Write them into the defect location result sequence in the traversal order of the polygon set. Verify duplicate coordinate items in the defect location result sequence and remove duplicates to obtain the bridge appearance defect coordinate set.

[0085] Specifically, the process involves reading the polygonal feature contour line generated in the previous steps, locating all internal mesh faces completely enclosed by this contour line within the closed bridge 3D mesh, extracting the principal curvature of each vertex of these internal mesh faces, calculating the absolute value of the difference between the principal curvature of the same mesh face and its adjacent mesh faces, summing the curvature differences between all adjacent faces and dividing by the number of adjacent faces to obtain the change range of the corresponding curvature value for that mesh face, then extracting the 3D coordinates of the three vertices of each mesh face within the area enclosed by the polygonal feature contour line, calculating the area of ​​each triangular mesh face using Heron's formula, summing the areas of all mesh faces within the enclosed area to obtain the total surface area value of the entire enclosed area, and taking the average of the change range of the curvature values ​​of each mesh face within the area as the curvature change rate of the corresponding enclosed area, and writing it together with the total surface area value into a memory record. This process is then applied to 100 typical crack and spalling defect areas in the bridge's historical inspection records. The curvature change rate was statistically analyzed, and the average curvature change rate was calculated to be 0.15. 0.15 was set as the curvature threshold. Simultaneously, the actual physical area of ​​100 typical spalling defects was statistically analyzed, revealing that their areas were mainly concentrated between 0.01 square meters and 0.5 square meters. This range was set as the spalling area interval. The curvature change rate and total surface area of ​​the current enclosed area were read. For example, if the curvature change rate of the current enclosed area was 0.2 and the total surface area was 0.08 square meters, the comparison showed that 0.2 was greater than the set curvature threshold of 0.15, and 0.08 square meters fell within the spalling area interval of 0.01 to 0.5 square meters. This area was determined to be a real surface defect area and retained. Conversely, if an area had an area less than 0.01 square meters or a curvature change rate less than 0.15, it was determined to be noise interference or normal fluctuation and was removed. All retained areas meeting the conditions were integrated together to form a polygon set.

[0086] Read the previously acquired polygon set. For each polygon in the set, iterate through all the outer contour boundary points and all the mesh face vertices contained within it, extracting the X, Y, and Z axis 3D spatial coordinates of each vertex. Count the total number of vertices contained within the current polygon, sum the X-axis coordinates of all vertices, and divide the sum by the total number of vertices to obtain the X-axis coordinate of the center position. Similarly, sum the Y and Z axis coordinates of all vertices and divide by the total number of vertices to obtain the Y and Z axis coordinates of the center position. Combine these three coordinate components to obtain the coordinates of the center vertex of the polygon. Establish a one-dimensional coordinate data list according to the retrieval order of the polygon set, and store the 3D spatial coordinates of the center vertex of each polygon. The coordinates are sequentially filled into this data list as the defect location result sequence. After all coordinates are written, a judgment distance threshold is set. For example, according to the spatial resolution accuracy requirements of bridge inspection, two points with a spatial distance of less than 0.05 meters are judged as the same physical defect. 0.05 meters is set as the repetition judgment threshold. All three-dimensional coordinates in the defect location result sequence are traversed, and the three-dimensional Euclidean distance between any two coordinate points is calculated. When it is found that the distance between two coordinate points is 0.02 meters, which is less than the set repetition judgment threshold of 0.05 meters, it is determined that these two coordinate items represent the same bridge appearance defect duplicate record. The duplicate coordinate item at the end of the sequence is removed, and only the unique coordinate item at the beginning is retained. In this way, the deduplication and sorting operation of all coordinates in the sequence is completed, and the bridge appearance defect coordinate set is obtained.

Claims

1. A bridge appearance defect detection system based on 3D point cloud reconstruction, characterized in that, The system includes: The point cloud spatial subdivision module constructs a three-dimensional bounding box based on the bridge point cloud, calculates the local geometric degradation index, and compares the local geometric degradation index with the flatness threshold and the degradation threshold respectively to obtain the initial three-dimensional polygon mesh. The mesh feature extraction module extracts the hole boundaries on the initial three-dimensional polygonal mesh, performs closure processing to obtain the boundary loop, generates bridge symmetry axis parameters, projects the boundary loop onto the opposite side of the initial three-dimensional polygonal mesh along the bridge symmetry axis parameters to extract the corresponding facets, and obtains mirror repair mesh facets. The mesh topology stitching module calculates the distance difference between the outer edge vertex and the hole edge vertex based on the initial three-dimensional polygonal mesh and the mirrored repair mesh patch, obtains the edge geometric gap difference value, extracts the difference coordinate value inside the mirrored repair mesh patch, constructs the Laplacian operator matrix equation, and obtains the three-dimensional mesh of the closed bridge. The appearance defect location module extracts the area where the density feature patch is located in the three-dimensional mesh of the closed bridge, calculates the angle between the normal vectors of all adjacent patches in the area, connects the adjacent patches whose normal vector angle is greater than the angle judgment threshold to the outside to obtain the polygon feature contour line, calculates the rate of curvature change and surface area of ​​the polygon feature contour line, and filters to obtain the coordinate set of bridge appearance defects.

2. The bridge appearance defect detection system based on three-dimensional point cloud reconstruction according to claim 1, characterized in that, The steps for obtaining the initial 3D polygon mesh are as follows: A 3D bounding box is constructed based on the bridge point cloud and an octree is performed. The point cloud coordinates within each leaf node are located one by one. The normal vector of the fitting plane, the eigenvalue of the covariance matrix, and the offset of each point from the local fitting plane are extracted. The geometric parameters within the same leaf node are collected to form a geometric parameter group of the leaf node. Calculate the local geometric degradation index based on the set of geometric parameters of the leaf nodes; Each local geometric degradation index is determined by interval matching with the flatness threshold and degradation threshold. Leaf nodes with values ​​less than the flatness threshold are extracted as scale patches. Leaf nodes within the flatness threshold and degradation threshold intervals are further subdivided. Leaf nodes with values ​​greater than or equal to the degradation threshold are subdivided to the highest level and extracted as density feature patches. The scale patches and density feature patches are combined to form an initial three-dimensional polygon mesh.

3. The bridge appearance defect detection system based on three-dimensional point cloud reconstruction according to claim 1, characterized in that, The steps for obtaining the bridge's axis of symmetry parameters are as follows: Based on the initial three-dimensional polygonal mesh, open edges that are shared only by a single mesh face are retrieved one by one. Open edge chains are spliced ​​according to the connection order of the endpoints of the open edges. The closure relationship between the first and last vertices of the open edge chain is verified. For open edge chains with discontinuities between the first and last vertices, closed edges are added. The order of all vertices on the closed edges is rearranged according to a unified wrapping direction to form a boundary loop. Traverse all vertices on the initial three-dimensional polygonal mesh, record the mapping coordinates of the vertex normal vector on the unit sphere point by point, extract the principal curvature of the vertex point by point, construct the covariance matrix of the Gaussian mapping lattice according to all the mapping coordinates, extract the eigenvectors of the Gaussian mapping lattice covariance matrix as principal components, and calibrate the bridge symmetry axis parameters according to the directional distribution of the principal components.

4. The bridge appearance defect detection system based on three-dimensional point cloud reconstruction according to claim 1, characterized in that, The steps for obtaining the mirror-repaired mesh patch are as follows: Next, the boundary ring is projected onto the opposite side of the initial three-dimensional polygon mesh along the bridge symmetry axis parameter. The mesh surfaces within the coverage area of ​​the projection landing point are screened to form a corresponding rigid transformation group. The vertex coordinates, bridge symmetry axis parameters, and spatial pose offset relationship of the corresponding face in the rigid transformation group are read. The rigid transformation matrix from the vertex of the corresponding face to the position inside the boundary ring is established point by point. The corresponding face is projected and copied into the boundary ring according to the rigid transformation matrix. Then, the position of the copied mesh surface boundary is matched and the vertex order is sorted to obtain the mirror repair mesh surface.

5. The bridge appearance defect detection system based on three-dimensional point cloud reconstruction according to claim 1, characterized in that, The steps for obtaining the edge geometric gap difference value are as follows: Based on the initial 3D polygonal mesh and the mirrored repair mesh patch, locate all hole edge vertices in the hole region where the boundary ring is located, locate all outer edge vertices on the outer contour of the mirrored repair mesh patch, arrange the hole edge vertices in sequence according to the boundary ring's wrapping direction, arrange the outer edge vertices in sequence according to the connection order of the outer contour of the mirrored repair mesh patch, read the 3D coordinates of each group of outer edge vertices and hole edge vertices one by one, calculate the spatial distance difference between each group of corresponding vertices, and write all spatial distance differences into the boundary ring position according to the vertex order to generate edge geometric gap difference values.

6. The bridge appearance defect detection system based on three-dimensional point cloud reconstruction according to claim 1, characterized in that, The steps for obtaining the three-dimensional mesh of the closed bridge are as follows: Extract the coordinate offsets of all vertices inside the mirror-repaired mesh relative to their adjacent vertices. Arrange each set of coordinate offsets according to the adjacency relationship of the vertices inside the mirror-repaired mesh and write them into the differential coordinate values ​​inside the mirror-repaired mesh. Construct a Laplacian operator matrix equation based on the connection relationship between the hole edge vertices, the outer edge vertices, and the vertices inside the mirror-repaired mesh. Write the differential coordinate values ​​inside the mirror-repaired mesh into the constant term constraint position of the Laplacian operator matrix equation. Then write the edge geometric gap differential value into the boundary constraint position of the Laplacian operator matrix equation. Solve the coordinate components corresponding to each vertex row by row to generate the edge transition coordinates. The positions of the hole edge vertices corresponding to the edge transition coordinates are read one by one. The original three-dimensional coordinates of the corresponding hole edge vertices in the initial three-dimensional polygon mesh are replaced with the edge transition coordinates. The boundary vertices of the mirror repair mesh are reconnected according to the updated hole edge vertex positions. The connection order of the adjacent patches of the hole edge vertices is corrected. The splicing relationship of the triangular patches of the boundary of the mirror repair mesh is corrected. The coordinate breaks and topological breaks at the hole edge positions are eliminated to obtain the closed bridge three-dimensional mesh.

7. The bridge appearance defect detection system based on three-dimensional point cloud reconstruction according to claim 1, characterized in that, The steps for obtaining the polygonal feature contour line are as follows: Based on the closed bridge 3D mesh, all mesh faces within the area where the density feature facet is located are located. Adjacent facets are extracted one by one according to the shared edge relationship. The normal vector component of each mesh facet is read. The angle between the normal vectors of each pair of adjacent facets is calculated. Adjacent facets with a normal vector angle greater than the angle judgment threshold are retained as candidate connection pairs. The first and last connections are performed according to the continuous distribution order of the candidate connection pairs in the outer contour direction to eliminate duplicate connection edges, retain the outer envelope boundary, and generate polygon feature contour lines.

8. The bridge appearance defect detection system based on three-dimensional point cloud reconstruction according to claim 1, characterized in that, The steps for obtaining the coordinate set of bridge appearance defects are as follows: Extract all grid faces within the area enclosed by the polygon feature contour lines one by one, count the change range of curvature values ​​of adjacent grid faces one by one, accumulate the area values ​​of each grid face according to the enclosed range of the polygon feature contour lines, and write the curvature change rate and surface area value of the corresponding enclosed area respectively. Keep the enclosed areas where the curvature change rate is greater than the curvature threshold and the surface area value falls within the peeling area range, and remove the remaining enclosed areas to form a polygon set. Based on the polygon set, the three-dimensional spatial coordinates of all vertices of each polygon are read one by one. The position of the center vertex is calculated according to the spatial distribution of the three-dimensional spatial coordinates of all vertices within the same polygon. The three-dimensional spatial coordinates of the center vertex corresponding to each polygon are extracted and written into the defect location result sequence in the traversal order of the polygon set. Duplicate coordinate items in the defect location result sequence are checked and deduplicated to obtain the bridge appearance defect coordinate set.

Images