Bridge prefabricated part-oriented handheld image point cloud and laser point cloud fusion detection method

Abstract

The invention discloses a handheld image point cloud and laser point cloud fusion detection method for a bridge prefabricated part. The method comprises the following steps: synchronously obtaining a laser point cloud and a sequence image of the prefabricated part; reconstructing the sequence image to obtain an image point cloud; performing normalization by taking the laser point cloud as a reference to obtain the laser point cloud and the image point cloud with consistent physical scales; initial attitude alignment of the two sets of point clouds is completed, rigid fine registration is carried out, the laser point clouds are corrected, and fused point clouds are obtained; generating a standard virtual point cloud; performing semantic segmentation on the fused point cloud, and outputting a point cloud set of each sub-component; extracting line-plane features of the key points; calculating the key geometric dimension of the prefabricated part, comparing the key geometric dimension with the standard virtual point cloud, and judging whether the dimension deviation meets a preset deviation or not; and evaluating the flatness quality of the surface of the component, and generating a three-dimensional full-field deviation chromatogram which covers the whole surface of the component. According to the invention, rapid, comprehensive and high-precision detection of the geometric dimension and the appearance flatness of the prefabricated part is realized.

Description

Technical Field

[0001] This invention belongs to the field of bridge inspection technology, and in particular relates to a handheld image point cloud and laser point cloud fusion inspection method for prefabricated bridge components. Background Technology

[0002] With the rapid development of prefabricated bridge construction technology, the geometrical accuracy and appearance quality of precast components (such as box girders and T-beams) directly determine the alignment and structural safety of the completed bridge. Traditional inspection methods mainly rely on manual tools such as measuring tapes and straightedges or total stations for single-point measurements, which suffers from low efficiency, large human error, and incomplete and discrete data, making it difficult to meet the high-precision and automated inspection requirements of modern bridge industrial production.

[0003] In recent years, 3D laser scanning technology has been increasingly applied to engineering inspection due to its non-contact and high-precision characteristics. However, while stationary 3D laser scanners offer high accuracy, they are bulky, cumbersome to set up, and prone to blind spots when scanning complex components. On the other hand, handheld 3D laser scanners, while flexible and efficient, are typically based on Simultaneous Localization and Mapping (SLAM) algorithms, which can lead to trajectory drift and accumulated errors when scanning large components, and their point cloud noise is relatively high. For precast bridge components with extremely strict dimensional deviation control requirements (millimeter level), relying solely on handheld laser point clouds is often insufficient in accuracy and cannot be directly used for precise assessment of key geometric dimensions. Furthermore, laser point clouds lack texture information, making it difficult to intuitively identify minute cracks or appearance defects. Summary of the Invention

[0004] Purpose of the invention: The purpose of this invention is to provide a handheld image point cloud and laser point cloud fusion detection method for precast bridge components, which enables rapid, comprehensive and high-precision detection of the geometric dimensions and appearance flatness of precast components.

[0005] Technical Solution: To achieve the above objectives, the present invention provides a handheld image point cloud and laser point cloud fusion detection method for precast bridge components, comprising the following steps: Step 1: Plan the scanning path according to the geometric dimensions of the precast component, and use a handheld 3D laser scanner with a built-in camera to perform a full-coverage scan of the precast component, simultaneously acquiring the laser point cloud of the precast component. and image sequences ; Step 2: Use the structure-of-motion reconstruction algorithm and multi-view stereo vision algorithm to process the image sequence. Reconstruction is performed to obtain an image point cloud with RGB color information. ; Step 3: Process the laser point cloud and image point cloud Preprocessing was performed separately, and laser point clouds were used. Based on the geometric feature distribution, the scale scaling factor is calculated to transform the image point cloud. Normalized to the true physical scale, laser point clouds with consistent physical scale are obtained. and image point cloud ; Step 4: Based on the Fast Point Feature Histogram (FPFH) algorithm, complete the initial pose alignment of the two sets of point clouds to obtain the coarsely registered point cloud; Step 5: Based on the coarsely registered point cloud obtained in Step 4, rigid fine registration is first performed using the point-to-surface iterative nearest point algorithm. Then, a non-rigid drift correction model based on the embedded deformation map is constructed. The deformation field is solved using the global constraints of the image point cloud to correct the cumulative drift of the laser point cloud, resulting in the fused point cloud. ; Step 6: Based on the design drawings and parameters of the prefabricated components, establish a 3D parametric model and generate a standard virtual point cloud. ; Step 7: Based on the geometric features of the point cloud, fuse the point cloud. Semantic segmentation is performed to divide the data into point cloud sets for each sub-component: top plate, bottom plate, web plate, and flange plate. The point cloud clusters for each sub-component are then output. ; Step 8: For the point cloud sets of each sub-component obtained in Step 7, fit the best geometric plane or quadratic surface using the global least squares method, and extract key point, line and surface features. Step 9: Based on the key point, line, and surface features extracted in Step 8, calculate the key geometric dimensions of the prefabricated components and compare them with the standard virtual point cloud model in Step 6. Compare the measurements to determine if the dimensional deviations meet the preset deviations. ; Step 10: Using the best fitted geometric plane or quadratic surface in Step 8 as a reference, calculate the vertical distance residual from each point in the point cloud of each sub-component to the corresponding reference plane. Evaluate the flatness quality of the component surface through residual statistical analysis, map the flatness analysis results of all sub-components back to the three-dimensional parametric model, and generate a three-dimensional full-field deviation chromatogram covering the entire surface of the component.

[0006] Optionally, step 2 specifically includes the following steps: Step 2.1: Process the acquired image sequence Feature extraction and matching are performed. The scale-invariant feature transformation algorithm is used to detect key feature points in the image, and corresponding feature point pairs are established between adjacent images by calculating the Euclidean distance of the feature descriptors. Step 2.2: Sparse reconstruction is performed based on the Structure for Motion Reconstruction (SMR) algorithm. The intrinsic and extrinsic parameters of the camera are recalculated and globally optimized using pairs of identical feature points. The camera pose parameters and 3D feature point coordinates are then jointly optimized using the bundle adjustment algorithm. The bundle adjustment algorithm aims to minimize the reprojection error by iteratively solving the objective function. The minimum value is found to obtain the optimal camera parameters. and the optimal 3D feature point coordinates The globally optimized set of 3D feature points constitutes a high-precision sparse point cloud. , In the formula, For the number of images, The number of feature points, For the first The feature point at the th ... Visibility weights in an image For the first The feature point at the th ... The coordinates of the observed pixels on the image. For camera projection model, For the first Camera pose parameters For the first The coordinates of the feature points; Step 2.3: Perform dense reconstruction based on multi-view stereo vision algorithm. Using the camera pose parameters and sparse point cloud determined in Step 2.2, calculate the depth map for each image. Then, eliminate mismatches by checking depth consistency and epipolar constraints. Backproject and fuse all depth maps to generate dense image point cloud data with RGB color information. .

[0007] Optionally, step 3 specifically includes the following steps: Step 3.1: Traverse the laser point cloud data separately and image point cloud data Each point in Searching points based on kd-tree data structure of nearest neighbor Calculation points arrive The average distance between the nearest neighbors : , in, This represents the Euclidean distance between two points; This indicates the number of nearest neighbors; Step 3.2: Calculate the average distance of all points in both the laser point cloud and the image point cloud. Given the distribution of values, calculate the corresponding mean values ​​for each. and standard deviation : , in The number of points in a point element; Step 3.3: Set the standard deviation multiple threshold For each point Make a judgment if the conditions are met. If the point is not an outlier, it will be marked as an outlier and removed, while the remaining valid point cloud will be retained. Step 3.4: Construct a 3D voxel mesh from the denoised effective point cloud and set the voxel side length parameters. Calculate the 3D bounding box of the point cloud and divide the bounding box space into multiple voxel units. All points in the laser point cloud data and image point cloud data are mapped to the corresponding voxel units respectively; Step 3.5: For each non-empty voxel cell containing point cloud data, calculate the centroid coordinates of all points within the voxel cell. The centroid is then used as the sole representative point of the voxel unit, replacing all original points within the voxel, to obtain the preprocessed laser point cloud. and image point cloud Centroid coordinates The calculation formula is: , in, This represents the number of points contained within the current voxel unit. The first voxel The coordinates of the points; Step 3.6: Assemble the preprocessed laser point cloud With image point cloud The centroids of all voxel units were calculated, and the preprocessed laser point cloud was calculated. With image point cloud The ratio of the 3D bounding box or principal component eigenvalues ​​is used to obtain the scale scaling factor. Isotropic scaling of the image point cloud using a scaling factor. To make the image point cloud Capable of laser point cloud Consistent physical scale, outputting laser point clouds with consistent physical scale. and image point cloud .

[0008] Optionally, step 4 specifically includes the following steps: Step 4.1: Process laser point clouds of consistent physical scales separately. and image point cloud For each point in the point cloud, its normal vector is calculated using principal component analysis. and points in its neighborhood Establish the Dabu marker frame and calculate the three angular characteristic elements between them. ,in The normal vector of the neighboring point is in Projected components on the axis, The normal vector of the line connecting two points is the vector at the center point. The projection on The yaw angle of the normal vector of a neighboring point on the plane; Step 4.2: Based on the three angular feature elements calculated in Step 4.1 Each of their respective value ranges is divided into a preset number of feature intervals; statistical points The frequency distribution of point pairs formed by the point and all its neighboring points within each feature interval is calculated, and the frequency distributions from the three angles are concatenated to generate a point. Simplified point feature histogram And by using weighted neighborhood features, the points are calculated Fast point feature histogram : , in, For point The number of neighborhood points used for feature calculation For the first in this neighborhood One point, For point Its first neighborhood points Euclidean distance weights between them; For the corresponding neighborhood points Simplified point feature histogram; Step 4.3: Calculate laser point clouds at a consistent physical scale and image point cloud The Euclidean distance between the FPFH feature descriptors of each point is used to construct the feature correspondence. The sampling consistency algorithm is used for iterative filtering. During the iteration process, the geometric consistency between corresponding points is constrained, and incorrect corresponding point pairs are eliminated to obtain the correct corresponding point pairs after filtering. Step 4.4: Based on the selected correct corresponding point pairs, solve for the rigid body transformation matrix using singular value decomposition. laser point cloud Transform to image point cloud In the coordinate system, the initial pose alignment of the two sets of data is completed to obtain the point cloud after coarse registration.

[0009] Optionally, step 5 specifically includes the following steps: Step 5.1: Using image point clouds To achieve this, the point-to-surface ICP algorithm is used to analyze laser point clouds. Iterative optimization is performed to eliminate the rigid pose deviation between the two, resulting in a rigidly registered laser point cloud. The least squares objective function based on the point-to-surface distance is... The expression is: , in, For a three-dimensional rigid body rotation matrix, It is a translation vector; This represents the number of valid correspondence pairs established during the iteration process. laser point cloud The first in One source point, For image point clouds Zhongyu Matching target points; For target point The unit normal vector at that location, Represents the dot product of vectors; Step 5.2: Construct an embedded deformation map for the rigidly registered laser point cloud. Perform spatial uniform downsampling, select A set of control nodes For rigidly registered laser point clouds Each point in In the set Search for it in the middle Find the nearest neighbor control nodes and calculate the point. Received its first Nearest neighbor control nodes Influence on skin weight ,in, The set number of nearest neighbor control nodes. For point To its The maximum Euclidean distance among the nearest neighbor control nodes is used to ensure a smooth transition of weights in the deformation space. Step 5.3: Establish the non-rigid drift correction energy function and define a data fitting term. and regularization smoothing term Total energy objective function This is used to solve for the local affine transformation parameters of each control node: , in, These are preset weighting coefficients for the data fitting term, used to control the degree to which the deformed point cloud is integrated into the target point cloud; These are preset weight coefficients for the regularization smoothing term, used to control the rigidity and continuity of the control graph topology. Step 5.4: Construct data fitting constraints in the image point cloud. Searching for laser points Corresponding point The constrained laser point cloud is brought closer to the image point cloud to correct geometric drift; data fitting term The expression is: , in, To establish the total number of valid matching point pairs, For laser points In image point cloud The target matching point coordinates are shown in the figure; assuming each control node... Each has an independent local affine transformation, which includes a rotation matrix. and a translation vector Laser point cloud after rigid registration any point in Coordinates after deformation It is determined by the transformations of its neighboring nodes: , in, Point In the control chart The set of nearest neighbor control nodes; The skin weights are calculated in step 5.2. The physical meaning of this formula is: through an interpolation algorithm, the local rotations and translations of sparse control nodes are smoothly transferred to the dense laser point cloud, thereby simulating a non-rigid bending correction effect.

[0010] Step 5.5: Construct regularization smoothing constraints to ensure consistent transformations between adjacent control nodes, maintain the continuity of the point cloud topology, and prevent overfitting or tearing. The regularization smoothing term... The expression is: , in, This represents the total number of control nodes selected in step 5.2; To control the topology of the graph and the nodes Adjacent control node sets; and Adjacent control nodes Spatial coordinates and translation vector parameters; Step 5.6: Minimize using the Gauss-Newton method Solve for the optimal local affine transformation parameters of all control nodes, i.e., the optimal rotation matrix of each control node. With translation vector Using the linear interpolation algorithm and the skin weights calculated in step 5.2, the solved optimal transformation parameters are applied to the rigidly registered laser point cloud. Each point in the sparse control node is processed to smoothly propagate the deformation of the sparse control node to the global point cloud, ultimately outputting a fused point cloud after non-rigid drift correction. .

[0011] Optionally, step 6 specifically includes the following steps: Step 6.1: Define the parametric geometric features of the precast bridge components, obtain the key design structural parameters of the precast components, and define the set of geometric parameters used to describe the cross-sectional shape as... ; Step 6.2: Based on key design structural parameters, reconstruct the standard 3D solid model using parametric modeling algorithms. First, establish a local modeling coordinate system based on the cross-sectional parameter set. Calculate the coordinates of each corner point of the cross section in a two-dimensional plane. And connect the corner points in topological order to form a closed two-dimensional cross-sectional profile. ; Redefine the central axis vector of the component Its modulus is equal to the total length of the input component. Finally, the two-dimensional cross-sectional profile was... Vector along the central axis A spatial stretching operation is performed to generate a closed 3D geometric solid. The edges of the solid are then processed according to preset chamfer radius parameters to obtain the final standard 3D solid model. ; Step 6.3: For the standard 3D solid model Triangular mesh discretization is performed to generate a closed triangular mesh model; a Poisson disk sampling algorithm is used to uniformly resample the surface of the triangular mesh model to generate a standard virtual point cloud. .

[0012] Optionally, step 7 specifically includes the following steps: Step 7.1: For the fused point cloud Calculate the local normal vector at each point and set the RANSAC distance threshold. Normal consistency threshold and minimum number of cluster points ; Step 7.2: Employ the serialized random sampling consistency algorithm in the fused point cloud. In the intermediate loop iteration, three points are randomly sampled each time to fit a planar model, and the points that satisfy the distance threshold are extracted. and normal consistency threshold Intra-points as independent point cloud clusters And then peel it off from the origin cloud until the number of remaining points is less than ; Step 7.3: Process the extracted independent point cloud clusters Perform plane fitting to obtain its normal vector and plane intercept Calculate the normal vector Angle with the global coordinate system Z-axis Set verticality threshold and levelness threshold ,like Then it is determined to be a horizontal component. If so, it is classified as a vertical component; Step 7.4: For horizontal components, based on the intercept parameter in the plane equation... Sort the height information of the representatives, and then... The largest one is marked as the top plate. The smallest component is labeled as the base plate, and the rest as flange plates; for vertical components, they are labeled as web plates, thus completing the semantic segmentation of each part of the component, obtaining the point cloud set of each sub-component, and outputting the point cloud cluster of each sub-component. .

[0013] Optionally, step 8 specifically includes the following steps: Step 8.1: For each semantically segmented sub-component point cloud cluster output from Step 7 Establish the general form of the plane equation The overall least squares method is used for parameter estimation, and an objective function is constructed to minimize the sum of squared perpendicular distances from all interior points to the fitting plane. : , in, For point cloud clusters The number of points in the middle, For point cloud clusters The Middle The coordinates of each point are obtained; the optimal plane parameters are obtained by solving the eigenvalue or singular value decomposition of the objective function. , , , ; Step 8.2: Based on the planar parameters of each sub-component obtained by fitting in Step 8.1, identify two adjacent but non-parallel sub-components, solve the planar equations of the two sub-components, and solve the equation of the intersection of the two planes as the measured edge feature lines of the precast component. Step 8.3: Identify three adjacent but non-coplanar sub-components, establish the plane equations of the three sub-components, and solve for the coordinates of the intersection point of the three surfaces, which will be used as the measured key corner points of the precast component.

[0014] Optionally, step 9 specifically includes the following steps: Step 9.1: Extract the sub-component planar parameters and key corner coordinates calculated in Step 8. For the component height and web thickness, determine them by the distance between the corresponding approximate parallel planes on both sides of the component. , in, Reference plane The normal vector parameters; For plane The intercept parameter; Opposite planes constrained by a plane Equivalent intercept parameters under the same normal vector direction; for each dimension of component length, top plate width, and bottom plate width, determined by key corner points at the end or edge of the component, and the coordinates of the key corner points defining the two ends of this dimension are extracted. and Calculate the Euclidean distance between two points: , Step 9.2: Convert the standard virtual point cloud from Step 6 into a virtual point cloud. Design values ​​of key structural parameters Compared with the measured geometric dimensions The comparison is performed, and the allowable deviation thresholds for each dimension are set according to the preset quality acceptance specifications. The dimensional deviations are then calculated. , if the deviation If the dimension is within the allowable deviation threshold range, it is considered qualified; otherwise, it is considered unqualified.

[0015] Optionally, step 10 specifically includes the following steps: Step 10.1: Traverse each planar sub-component point cloud cluster extracted in Step 7. Obtain the optimal plane equation parameters for fitting the sub-component in step 8. , , and ; Step 10.2: Calculate the point cloud clusters of sub-components Each point in Vertical distance residual to the fitting plane of the sub-component The calculation formula is: , Based on the calculated residuals The root mean square error (RMSE) of the sub-component surface is statistically analyzed to quantitatively evaluate the surface flatness quality of that area. The calculation formula is as follows: , in Point cloud cluster for this sub-component The total number of point clouds in the cloud; Step 10.3: Based on the precast component quality acceptance specifications, set the allowable deviation threshold for the surface flatness of the components. First, the entire sub-component... With threshold Perform macroscopic comparison and judgment: if If so, the overall flatness of the sub-component is deemed acceptable; if If the flatness of the sub-component exceeds the limit, then the overall flatness of the sub-component is determined to be excessive; secondly, the residuals at each point are... With threshold Compare: If The surface flatness at that point is deemed acceptable; if The surface flatness at that point is determined to be out of tolerance. Finally, the flatness analysis results of all sub-components are mapped back to the three-dimensional model to generate a three-dimensional full-field deviation chromatogram covering the entire surface of the component, which intuitively represents the unevenness mass distribution of each part of the beam surface.

[0016] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: This invention utilizes a handheld 3D laser scanner to simultaneously acquire laser point clouds and sequential images of components, achieving high-precision fusion of heterogeneous data based on a coarse-to-fine registration strategy. It utilizes globally optimized image point cloud constraints and corrects accumulated errors during the handheld scanning process, improving the overall geometric accuracy of the point cloud model. Simultaneously, this invention achieves automated semantic segmentation of various parts of prefabricated components, automatically identifying substructures such as the top slab and web through a serialized random sampling consistency algorithm. Based on the comparison between the fused point cloud and the parametric standard model, key geometric dimensions are extracted and surface flatness residuals are statistically analyzed, achieving full automation from data acquisition to quality assessment. This invention enables rapid, comprehensive, and high-precision detection of the dimensions and appearance quality of prefabricated bridge components, providing effective data support for the quality control of prefabricated bridge construction. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the process of the present invention; Figure 2 This is a schematic diagram of the original point cloud collected in this invention; Figure 3 This is a schematic diagram of laser point cloud noise reduction in this invention; Figure 4 This is a schematic diagram of the image point cloud in this invention; Figure 5 This is a schematic diagram of the laser point cloud in this invention; Figure 6 This is a schematic diagram of the cumulative error chromatogram after point cloud fusion in this invention; Figure 7 This is a standard virtual point cloud of a precast pier component in this invention; Figure 8 This is a schematic diagram of the point cloud slice of the steel strand anchorage hole of a precast pier component in this invention; Figure 9 This is a feature fitting diagram of the steel strand anchorage hole of a precast pier component in this invention. Figure 10 This is a visual schematic diagram of the flatness deviation of a precast pier component in this invention. Detailed Implementation

[0018] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0019] like Figure 1 As shown, a precast pier component in a precast beam yard is selected as the object of illustration. The present invention provides a handheld image point cloud and laser point cloud fusion detection method for precast bridge components, comprising the following steps: Step 1: Plan the scanning path according to the geometric dimensions of the precast component, and use a handheld 3D laser scanner with a built-in camera to perform a full-coverage scan of the precast component, simultaneously acquiring the laser point cloud of the precast component. and image sequences .

[0020] Step 1 specifically includes the following steps: Step 1.1: Based on the geometric dimensions and stacking layout of the components to be tested in the precast beam yard, plan the data acquisition path of the handheld 3D laser scanner with built-in camera to ensure that the scanning view covers all visible feature surfaces of the components; Step 1.2: Scan the component under test according to the planned data acquisition path, and simultaneously capture a sequence of images using the built-in camera. The acquired images must maintain a preset overlap to meet the needs of subsequent 3D reconstruction; Step 1.3: Using the inertial measurement unit (IMU) and simultaneous localization and mapping (SLAM) algorithm built into the 3D laser scanner, the continuously acquired local scan slices are uniformly transformed to the global coordinate system to generate the complete original laser point cloud of the component. ,like Figure 2 As shown.

[0021] Step 2: Use the Structure of Motion Recovery (SFM) algorithm and the Multi-View Stereo Vision (MVS) algorithm to process the sequence image data. Reconstruction is performed to obtain an image point cloud with RGB color information. ,like Figure 4 As shown.

[0022] Step 2 specifically includes the following steps: Step 2.1: Process the acquired image sequence Feature extraction and matching are performed. The scale-invariant feature transform algorithm SIFT is used to detect key feature points in the image, and corresponding feature point pairs are established between adjacent images by calculating the Euclidean distance of the feature descriptors. Step 2.2: Sparse reconstruction based on the Structure of Motion (SFM) algorithm. To achieve camera pose accuracy superior to real-time SLAM estimation and eliminate accumulated errors, the intrinsic and extrinsic parameters of the camera are recalculated and globally optimized using pairs of corresponding feature points. Furthermore, the camera pose parameters and 3D feature point coordinates are jointly optimized using the bundle adjustment algorithm. The bundle adjustment algorithm aims to minimize the reprojection error by iteratively solving the objective function. The minimum value is found, thus yielding the optimal camera parameters. and the optimal 3D feature point coordinates The globally optimized set of 3D feature points constitutes a high-precision sparse point cloud. , In the formula, For the number of images, The number of feature points, For the first The feature point at the th ... Visibility weights in an image For the first The feature point at the th ... The coordinates of the observed pixels on the image. For camera projection model, For the first Camera pose parameters For the first The coordinates of the feature points; Step 2.3: Dense reconstruction is performed based on the Multi-View Stereo Vision (MVS) algorithm. Using the camera pose parameters and sparse point cloud determined in Step 2.2, a depth map is calculated for each image. Mismatches are eliminated through depth consistency verification and epipolar constraints. All depth maps are back-projected and fused to generate dense image point cloud data with RGB color information. .

[0023] Step 3: Process the laser point cloud data and image point cloud data Statistical filtering for noise reduction and voxel downsampling for preprocessing are performed separately, such as Figure 3 As shown, and using laser point cloud data Based on the geometric feature distribution, a scaling factor is calculated for the image point cloud data. Normalize to the true physical scale to obtain laser point cloud data with a consistent physical scale. and image point cloud data ,like Figure 5 As shown.

[0024] Step 3 specifically includes the following steps: Step 3.1: Traverse the laser point cloud data separately and image point cloud data Each point in Searching points based on kd-tree data structure of nearest neighbor Calculation points arrive The average distance between the nearest neighbors : , in, This represents the Euclidean distance between two points; This indicates the number of nearest neighbors; Step 3.2: Calculate the average distance of all points in both the laser point cloud data and the image point cloud data. Given the distribution of values, calculate the corresponding mean values ​​for each. and standard deviation : , in The number of points in a point element; Step 3.3: Set the standard deviation multiple threshold For each point Make a judgment if the conditions are met. If the point is not an outlier, it will be marked as an outlier and removed, while the remaining valid point cloud will be retained. Step 3.4: Construct a 3D voxel mesh from the denoised effective point cloud and set the voxel side length parameters. Calculate the 3D bounding box of the point cloud and divide the bounding box space into multiple voxel units. All points in the laser point cloud data and image point cloud data are mapped to the corresponding voxel units respectively; Step 3.5: For each non-empty voxel cell containing point cloud data, calculate the centroid coordinates of all points within the voxel cell. This centroid is used as the sole representative point of the voxel unit, replacing all original points within the voxel. , in, This represents the number of points contained within the current voxel unit. The first voxel The coordinates of the points; Obtain the preprocessed laser point cloud and image point cloud ; Step 3.6: Assemble the preprocessed laser point cloud With image point cloud The centroids of all voxel units were calculated, and the preprocessed laser point cloud was calculated. With image point cloud The ratio of the 3D bounding box or principal component eigenvalues ​​is used to obtain the scale scaling factor. Isotropic scaling of the image point cloud using a scaling factor. To make the image point cloud Capable of laser point cloud Consistent physical scale, outputting laser point clouds with consistent physical scale. and image point cloud .

[0025] Step 4: Based on the Fast Point Feature Histogram (FPFH) algorithm, find laser point clouds with consistent physical scales. and image point cloud Calculate the initial transformation matrix for the corresponding feature points. The initial pose alignment of the two sets of point clouds is completed, resulting in coarsely registered point clouds.

[0026] Step 4 specifically includes the following steps: Step 4.1: Process laser point clouds of consistent physical scales separately. and image point cloud For each point in the point cloud, its normal vector is calculated using Principal Component Analysis (PCA). and points in its neighborhood Establish the Darboux Frame and calculate the three angular feature elements between it and the frame. To eliminate the influence of initial pose differences, where The normal vector of the neighboring point is in Projected components on the axis, The normal vector of the line connecting two points is the vector at the center point. The projection on The yaw angle of the normal vector of a neighboring point on the plane; Step 4.2: Based on the three angular feature elements calculated in Step 4.1 Each of their respective value ranges is divided into a preset number of feature intervals; statistical points The frequency distribution of point pairs formed by the point and all its neighboring points within each feature interval is calculated, and the frequency distributions from the three angles are concatenated to generate a point. Simplified point feature histogram And by using weighted neighborhood features, the points are calculated Fast point feature histogram : , in, For point The number of neighborhood points used for feature calculation For the first in this neighborhood One point, For point Its first neighborhood points Euclidean distance weights between them; For the corresponding neighborhood points Simplified point feature histogram; Step 4.3: Calculate laser point clouds at a consistent physical scale and image point cloud The Euclidean distance between the FPFH feature descriptors of each point is used to construct the feature correspondence. The sampling consensus algorithm SAC-IA is used for iterative filtering. During the iteration process, the geometric consistency between corresponding points is constrained, and incorrect corresponding point pairs are eliminated to obtain the correct corresponding point pairs after filtering. Step 4.4: Based on the selected correct corresponding point pairs, solve for the rigid body transformation matrix using Singular Value Decomposition (SVD). laser point cloud Transform to image point cloud In the coordinate system, the initial pose alignment of the two sets of data is completed to obtain the point cloud after coarse registration.

[0027] Step 5: Based on the coarsely registered point cloud obtained in Step 4, rigid fine registration is first performed using the point-to-surface iterative nearest point algorithm. Then, a non-rigid drift correction model based on the embedded deformation map is constructed. The deformation field is solved using the global constraints of the image point cloud to correct the cumulative drift of the laser point cloud, resulting in the fused point cloud data. ,like Figure 6 As shown.

[0028] Step 5 specifically includes the following steps: Step 5.1: Using image point clouds To achieve this, the Point-to-Plane (ICP) algorithm is used to analyze laser point clouds. Iterative optimization is performed to eliminate the rigid pose deviation between the two, resulting in a rigidly registered laser point cloud. The least squares objective function based on the point-to-surface distance is... as follows: , in, For a three-dimensional rigid body rotation matrix, It is a translation vector; This represents the number of valid correspondence pairs established during the iteration process. laser point cloud The first in One source point, For image point clouds Zhongyu Matching target points; For target point The unit normal vector at that location, Represents the dot product of vectors; Step 5.2: Constructing an embedded deformation map. To efficiently deform the massive laser point cloud, we do not directly manipulate each point, but instead construct a sparse control map; this is done on the rigidly registered laser point cloud. Perform spatial uniform downsampling, select A set of control nodes For rigidly registered laser point clouds Each point in In the set Search for it in the middle Find the nearest neighbor control nodes and calculate the point. Received its first Nearest neighbor control nodes Influence on skin weight ,in, The set number of nearest neighbor control nodes. For point To its The maximum Euclidean distance among the nearest neighbor control nodes is used to ensure a smooth transition of weights in the deformation space. Step 5.3: Establish the non-rigid drift correction energy function and define a data fitting term. and regularization smoothing term Total energy objective function This is used to solve for the local affine transformation parameters of each control node: , in, These are preset weighting coefficients for the data fitting term, used to control the degree to which the deformed point cloud is integrated into the target point cloud; These are preset weight coefficients for the regularization smoothing term, used to control the rigidity and continuity of the control graph topology. Step 5.4: Construct data fitting constraints in the image point cloud. Searching for laser points Corresponding point The deformed laser point cloud is constrained to converge with the image point cloud to correct geometric drift. , in, To establish the total number of valid matching point pairs, For laser points In image point cloud The target matching point coordinates are shown in the figure; assuming each control node... Each has an independent local affine transformation, which includes a rotation matrix. and a translation vector Laser point cloud after rigid registration any point in Coordinates after deformation It is determined by the transformations of its neighboring nodes: , in, Point In the control chart The set of nearest neighbor control nodes; The skin weights are calculated in step 5.2. The physical meaning of this formula is: through an interpolation algorithm, the local rotations and translations of sparse control nodes are smoothly transferred to the dense laser point cloud, thereby simulating a non-rigid bending correction effect.

[0029] Step 5.5: Construct regularized smoothing constraints to ensure consistent transformations between adjacent control nodes, maintain the continuity of the point cloud topology, and prevent overfitting or tearing. , in, This represents the total number of control nodes selected in step 5.2; To control the topology of the graph and the nodes Adjacent control node sets; and Adjacent control nodes Spatial coordinates and translation vector parameters.

[0030] Step 5.6: Minimize using the Gauss-Newton method Solve for the optimal local affine transformation parameters of all control nodes (i.e., the optimal rotation matrix of each control node). With translation vector Using a linear interpolation algorithm and the skin weights calculated in step 5.2, the optimal transformation parameters are applied to the rigidly registered laser point cloud. Each point in the sparse control node is processed to smoothly propagate the deformation of the sparse control node to the global point cloud, ultimately outputting a fused point cloud after non-rigid drift correction. .

[0031] Step 6: Based on the design drawings and parameters of the prefabricated components, establish a three-dimensional parametric model, and discretize and sample it to generate a standard virtual point cloud. The generated standard virtual point cloud serves as the detection benchmark. like Figure 7 As shown.

[0032] Step 6 specifically includes the following steps: Step 6.1: Define the parametric geometric features of the precast bridge components, obtain the key design structural parameters of the precast components, and define the set of geometric parameters used to describe the cross-sectional shape as... ; Step 6.2: Based on the key design structural parameters, reconstruct the standard 3D solid model using a parametric modeling algorithm. The specific process is as follows: Establish a local modeling coordinate system based on the cross-section parameter set. Calculate the coordinates of each corner point of the cross section in a two-dimensional plane. And connect the corner points in topological order to form a closed two-dimensional cross-sectional profile. ; Define the central axis vector of the component Its modulus is equal to the total length of the input component. ; Two-dimensional cross-sectional profile Vector along the central axis A spatial stretching operation is performed to generate a closed 3D geometric solid. The edges of the solid are then processed according to preset chamfer radius parameters to obtain the final standard 3D solid model. ; Step 6.3: For the standard 3D solid model Triangular mesh discretization is performed to generate a closed triangular mesh model. The Poisson disk sampling algorithm is then used to uniformly resample the surface of the triangular mesh model, generating a standard virtual point cloud. .

[0033] Step 7: Based on the geometric features of the point cloud, fuse the point cloud. Semantic segmentation is performed to divide it into point cloud sets of sub-components such as top plate, bottom plate, web plate, and flange plate.

[0034] Step 7 specifically includes the following steps: Step 7.1: For the fused point cloud Calculate the local normal vector at each point and set the RANSAC distance threshold. Normal consistency threshold and minimum number of cluster points ; Step 7.2: Employ the Sequential Random Sample Consistency (RANSAC) strategy in the fused point cloud. In the intermediate loop iteration, three points are randomly sampled each time to fit a planar model, and the points that satisfy the distance threshold are extracted. and normal consistency threshold Intra-points as independent point cloud clusters And then peel it off from the origin cloud until the number of remaining points is less than ; Step 7.3: Process the extracted independent point cloud clusters Perform plane fitting to obtain its normal vector and plane intercept Calculate the normal vector Angle with the global coordinate system Z-axis Set verticality threshold and levelness threshold ,like Then it is determined to be a horizontal component. If so, it is classified as a vertical component; Step 7.4: For horizontal components, based on the intercept parameter in the plane equation... Sort the height information of the representatives, and then... The largest one is marked as the top plate. The smallest component is labeled as the base plate, and the rest as flange plates. For vertical components, they are labeled as web plates, thus completing the semantic segmentation of each part of the component and obtaining the point cloud set of each sub-component.

[0035] Step 8: For the point cloud sets of each sub-component obtained in Step 7, fit the optimal geometric plane or quadratic surface using the global least squares method, and extract key point, line, and surface features, such as... Figure 8 and Figure 9 As shown.

[0036] Step 8 specifically includes the following steps: Step 8.1: For each semantically segmented sub-component point cloud cluster output from Step 7 Establish the general form of the plane equation The overall least squares method is used for parameter estimation, and an objective function is constructed to minimize the sum of squared perpendicular distances from all interior points to the fitting plane. : , in, For point cloud clusters The number of points in the middle, For point cloud clusters The Middle The coordinates of each point are obtained; the optimal plane parameters are obtained by solving the eigenvalue or singular value decomposition of the objective function. , , , ; Step 8.2: Based on the planar parameters of each sub-component obtained by fitting in Step 8.1, identify two adjacent but non-parallel sub-components, solve the planar equations of the two sub-components, and solve the equation of the intersection of the two planes as the measured edge feature lines of the precast component. Step 8.3: Identify three adjacent but non-coplanar sub-components, establish the plane equations of the three sub-components, and solve for the coordinates of the intersection point of the three surfaces, which will be used as the measured key corner points of the precast component.

[0037] Step 9: Based on the key point, line, and surface features extracted in Step 8, calculate the key geometric dimensions of the prefabricated components and compare them with the standard virtual point cloud model in Step 6. Compare the measurements to determine if the dimensional deviations meet the preset deviations. .

[0038] Step 9 specifically includes the following steps: Step 9.1: Extract the sub-component planar parameters and key corner coordinates calculated in Step 8. For the component height and web thickness, determine them by the distance between the corresponding approximate parallel planes on both sides of the component. , in, Reference plane The normal vector parameters; For plane The intercept parameter; Opposite planes constrained by a plane Equivalent intercept parameters under the same normal vector direction; for dimensions such as component length, top plate width, and bottom plate width, determined by key corner points at the component ends or edges, extracting the coordinates of the key corner points defining the dimensions. and Calculate the Euclidean distance between two points: , Step 9.2: Convert the standard virtual point cloud from Step 6 into a virtual point cloud. Design values ​​of key structural parameters Compared with the measured geometric dimensions The comparison is performed, and the allowable deviation thresholds for each dimension are set according to the preset quality acceptance specifications. The dimensional deviations are then calculated. , if the deviation If the dimension is within the allowable deviation threshold range, it is considered qualified; otherwise, it is considered unqualified.

[0039] Step 10: Using the best-fit geometric plane or quadratic surface from Step 8 as a reference, calculate the vertical distance residuals from each point in the point cloud of each sub-component to the corresponding reference plane. Evaluate the surface smoothness quality of the components through residual statistical analysis. Figure 10 As shown.

[0040] Step 10 specifically includes the following steps: Step 10.1: Traverse each planar sub-component point cloud cluster extracted in Step 7. Obtain the optimal plane equation parameters for fitting the sub-component in step 8. , , and ; Step 10.2: Calculate the point cloud clusters of sub-components Each point in Vertical distance residual to the fitting plane of the sub-component The calculation is as follows: , Based on the calculated residuals The root mean square error (RMSE) of the surface of this sub-component is statistically analyzed to quantitatively evaluate the surface flatness quality of this part. The calculation is as follows: , in Point cloud cluster for this sub-component The total number of point clouds in the cloud; Step 10.3: Based on the precast component quality acceptance specifications, set the allowable deviation threshold for the surface flatness of the components. First, the entire sub-component. With threshold Perform macroscopic comparison and judgment; if If so, the overall flatness of the sub-component is deemed acceptable; if If the flatness of the sub-component exceeds the limit, then the overall flatness of the sub-component is determined to be excessive; secondly, the residuals at each point are... With threshold Compare: If The surface flatness at that point is deemed acceptable; if The surface flatness at that point is determined to be out of tolerance. Finally, the flatness analysis results of all sub-components are mapped back to the three-dimensional model to generate a three-dimensional full-field deviation chromatogram covering the entire surface of the component, which intuitively represents the unevenness mass distribution of each part of the beam surface.

Claims

1. A handheld image point cloud and laser point cloud fusion detection method for precast bridge components, characterized in that, Includes the following steps: Step 1: Plan the scanning path according to the geometric dimensions of the precast component, and use a handheld 3D laser scanner with a built-in camera to perform a full-coverage scan of the precast component, simultaneously acquiring the laser point cloud of the precast component. and image sequences ; Step 2: Use the structure-of-motion reconstruction algorithm and multi-view stereo vision algorithm to process the image sequence. Reconstruction is performed to obtain an image point cloud with RGB color information. ; Step 3: Process the laser point cloud and image point cloud Preprocessing was performed separately, and laser point clouds were used. Based on the geometric feature distribution, the scale scaling factor is calculated to transform the image point cloud. Normalized to the true physical scale, laser point clouds with consistent physical scale are obtained. and image point cloud ; Step 4: Based on the Fast Point Feature Histogram (FPFH) algorithm, complete the initial pose alignment of the two sets of point clouds to obtain the coarsely registered point cloud; Step 5: Based on the coarsely registered point cloud obtained in Step 4, rigid fine registration is first performed using the point-to-surface iterative nearest point algorithm. Then, a non-rigid drift correction model based on the embedded deformation map is constructed. The deformation field is solved using the global constraints of the image point cloud to correct the cumulative drift of the laser point cloud, resulting in the fused point cloud. ; Step 6: Based on the design drawings and parameters of the prefabricated components, establish a 3D parametric model and generate a standard virtual point cloud. ; Step 7: Based on the geometric features of the point cloud, fuse the point cloud. Semantic segmentation is performed to divide the data into point cloud sets for each sub-component: top plate, bottom plate, web plate, and flange plate. The point cloud clusters for each sub-component are then output. ; Step 8: For the point cloud sets of each sub-component obtained in Step 7, fit the best geometric plane or quadratic surface using the global least squares method, and extract key point, line and surface features. Step 9: Based on the key point, line, and surface features extracted in Step 8, calculate the key geometric dimensions of the prefabricated components and compare them with the standard virtual point cloud model in Step 6. Compare the measurements to determine if the dimensional deviations meet the preset deviations. ; Step 10: Using the best fitted geometric plane or quadratic surface in Step 8 as a reference, calculate the vertical distance residual from each point in the point cloud of each sub-component to the corresponding reference plane. Evaluate the flatness quality of the component surface through residual statistical analysis, map the flatness analysis results of all sub-components back to the three-dimensional parametric model, and generate a three-dimensional full-field deviation chromatogram covering the entire surface of the component.

2. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 2 specifically includes the following steps: Step 2.1: Process the acquired image sequence Feature extraction and matching are performed. The scale-invariant feature transformation algorithm is used to detect key feature points in the image, and corresponding feature point pairs are established between adjacent images by calculating the Euclidean distance of the feature descriptors. Step 2.2: Sparse reconstruction is performed based on the Structure for Motion Reconstruction (SMR) algorithm. The intrinsic and extrinsic parameters of the camera are recalculated and globally optimized using pairs of identical feature points. The camera pose parameters and 3D feature point coordinates are then jointly optimized using the bundle adjustment algorithm. The bundle adjustment algorithm aims to minimize the reprojection error by iteratively solving the objective function. The minimum value is found to obtain the optimal camera parameters. and the optimal 3D feature point coordinates The globally optimized set of 3D feature points constitutes a high-precision sparse point cloud. ， In the formula, For the number of images, The number of feature points, For the first The feature point at the th ... Visibility weights in an image For the first The feature point at the th ... The coordinates of the observed pixels on the image. For camera projection model, For the first Camera pose parameters For the first The coordinates of the feature points; Step 2.3: Perform dense reconstruction based on multi-view stereo vision algorithm. Using the camera pose parameters and sparse point cloud determined in Step 2.2, calculate the depth map for each image. Then, eliminate mismatches by checking depth consistency and epipolar constraints. Backproject and fuse all depth maps to generate dense image point cloud data with RGB color information. .

3. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 3 specifically includes the following steps: Step 3.1: Traverse the laser point cloud data separately and image point cloud data Each point in Searching points based on kd-tree data structure of nearest neighbor Calculation points arrive The average distance between the nearest neighbors : ， in, This represents the Euclidean distance between two points; This indicates the number of nearest neighbors; Step 3.2: Calculate the average distance of all points in both the laser point cloud and the image point cloud. Given the distribution of values, calculate the corresponding mean values ​​for each. and standard deviation : ， in The number of points in a point element; Step 3.3: Set the standard deviation multiple threshold For each point Make a judgment if the conditions are met. If the point is not an outlier, it will be marked as an outlier and removed, while the remaining valid point cloud will be retained. Step 3.4: Construct a 3D voxel mesh from the denoised effective point cloud and set the voxel side length parameters. Calculate the 3D bounding box of the point cloud and divide the bounding box space into multiple voxel units. All points in the laser point cloud data and image point cloud data are mapped to the corresponding voxel units respectively; Step 3.5: For each non-empty voxel cell containing point cloud data, calculate the centroid coordinates of all points within the voxel cell. The centroid is then used as the sole representative point of the voxel unit, replacing all original points within the voxel, to obtain the preprocessed laser point cloud. and image point cloud Centroid coordinates The calculation formula is: ， in, This represents the number of points contained within the current voxel unit. The first voxel The coordinates of the points; Step 3.6: Assemble the preprocessed laser point cloud With image point cloud The centroids of all voxel units were calculated, and the preprocessed laser point cloud was calculated. With image point cloud The ratio of the 3D bounding box or principal component eigenvalues ​​is used to obtain the scale scaling factor. Isotropic scaling of the image point cloud using a scaling factor. To make the image point cloud Capable of laser point cloud Consistent physical scale, outputting laser point clouds with consistent physical scale. and image point cloud .

4. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 4 specifically includes the following steps: Step 4.1: Process laser point clouds of consistent physical scales separately. and image point cloud For each point in the point cloud, its normal vector is calculated using principal component analysis. and points in its neighborhood Establish the Dabu marker frame and calculate the three angular characteristic elements between them. ,in The normal vector of the neighboring point is in Projected components on the axis, The normal vector of the line connecting two points is the vector at the center point. The projection on The yaw angle of the normal vector of a neighboring point on the plane; Step 4.2: Based on the three angular feature elements calculated in Step 4.1 Each of their respective value ranges is divided into a preset number of feature intervals; Statistical points The frequency distribution of point pairs formed by the point and all its neighboring points within each feature interval is calculated, and the frequency distributions from the three angles are concatenated to generate a point. Simplified point feature histogram And by using weighted neighborhood features, the points are calculated Fast point feature histogram : ， in, For point The number of neighborhood points used for feature calculation For the first in this neighborhood One point, For point Its first neighborhood points Euclidean distance weights between them; For the corresponding neighborhood points Simplified point feature histogram; Step 4.3: Calculate laser point clouds at a consistent physical scale and image point cloud The Euclidean distance between the FPFH feature descriptors of each point is used to construct the feature correspondence. The sampling consistency algorithm is used for iterative filtering. During the iteration process, the geometric consistency between corresponding points is constrained, and incorrect corresponding point pairs are eliminated to obtain the correct corresponding point pairs after filtering. Step 4.4: Based on the selected correct corresponding point pairs, solve for the rigid body transformation matrix using singular value decomposition. laser point cloud Transform to image point cloud In the coordinate system, the initial pose alignment of the two sets of data is completed to obtain the point cloud after coarse registration.

5. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 5 specifically includes the following steps: Step 5.1: Using image point clouds To achieve this, the point-to-surface ICP algorithm is used to analyze laser point clouds. Iterative optimization is performed to eliminate the rigid pose deviation between the two, resulting in a rigidly registered laser point cloud. The least squares objective function based on the point-to-surface distance is... The expression is: ， in, For a three-dimensional rigid body rotation matrix, It is a translation vector; This represents the number of valid correspondence pairs established during the iteration process. laser point cloud The first in One source point, For image point clouds Zhongyu Matching target points; For target point The unit normal vector at that location, Represents the dot product of vectors; Step 5.2: Construct an embedded deformation map for the rigidly registered laser point cloud. Perform spatial uniform downsampling, select A set of control nodes For rigidly registered laser point clouds Each point in In the set Search for it in the middle Find the nearest neighbor control nodes and calculate the point. Received its first Nearest neighbor control nodes Influence on skin weight ,in, The set number of nearest neighbor control nodes. For point To its The maximum Euclidean distance among the nearest neighbor control nodes is used to ensure a smooth transition of weights in the deformation space. Step 5.3: Establish the non-rigid drift correction energy function and define a data fitting term. and regularization smoothing term Total energy objective function This is used to solve for the local affine transformation parameters of each control node: ， in, These are preset weighting coefficients for the data fitting term, used to control the degree to which the deformed point cloud is integrated into the target point cloud; These are preset weight coefficients for the regularization smoothing term, used to control the rigidity and continuity of the control graph topology. Step 5.4: Construct data fitting constraints in the image point cloud. Searching for laser points Corresponding point The constrained laser point cloud is brought closer to the image point cloud to correct geometric drift; data fitting term The expression is: ， in, To establish the total number of valid matching point pairs, For laser points In image point cloud The target matching point coordinates are shown in the figure; assuming each control node... Each has an independent local affine transformation, which includes a rotation matrix. and a translation vector Laser point cloud after rigid registration any point in Coordinates after deformation It is determined by the transformations of its neighboring nodes: ， in, Point In the control chart The set of nearest neighbor control nodes; The skin weights calculated in step 5.2; Step 5.5: Construct regularization smoothing constraints to ensure consistent transformations between adjacent control nodes, maintain the continuity of the point cloud topology, and prevent overfitting or tearing. The regularization smoothing term... The expression is: ， in, This represents the total number of control nodes selected in step 5.2; To control the topology of the graph and the nodes Adjacent control node sets; and Adjacent control nodes Spatial coordinates and translation vector parameters; Step 5.6: Minimize using the Gauss-Newton method Solve for the optimal local affine transformation parameters of all control nodes, i.e., the optimal rotation matrix of each control node. With translation vector Using the linear interpolation algorithm and the skin weights calculated in step 5.2, the solved optimal transformation parameters are applied to the rigidly registered laser point cloud. Each point in the sparse control node is processed to smoothly propagate the deformation of the sparse control node to the global point cloud, ultimately outputting a fused point cloud after non-rigid drift correction. .

6. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 6 specifically includes the following steps: Step 6.1: Define the parametric geometric features of the precast bridge components, obtain the key design structural parameters of the precast components, and define the set of geometric parameters used to describe the cross-sectional shape as... ; Step 6.2: Based on key design structural parameters, reconstruct the standard 3D solid model using parametric modeling algorithms. First, establish a local modeling coordinate system based on the cross-sectional parameter set. Calculate the coordinates of each corner point of the cross section in a two-dimensional plane. And connect the corner points in topological order to form a closed two-dimensional cross-sectional profile. ; Redefine the central axis vector of the component Its modulus is equal to the total length of the input component. Finally, the two-dimensional cross-sectional profile was... Vector along the central axis A spatial stretching operation is performed to generate a closed 3D geometric solid. The edges of the solid are then processed according to preset chamfer radius parameters to obtain the final standard 3D solid model. ; Step 6.3: For the standard 3D solid model Triangular mesh discretization is performed to generate a closed triangular mesh model; a Poisson disk sampling algorithm is used to uniformly resample the surface of the triangular mesh model to generate a standard virtual point cloud. .

7. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 7 specifically includes the following steps: Step 7.1: For the fused point cloud Calculate the local normal vector at each point and set the RANSAC distance threshold. Normal consistency threshold and minimum number of cluster points ; Step 7.2: Employ the serialized random sampling consistency algorithm in the fused point cloud. In the intermediate loop iteration, three points are randomly sampled each time to fit a planar model, and the points that satisfy the distance threshold are extracted. and normal consistency threshold Intra-points as independent point cloud clusters And then peel it off from the origin cloud until the number of remaining points is less than ; Step 7.3: Process the extracted independent point cloud clusters Perform plane fitting to obtain its normal vector and plane intercept Calculate the normal vector Angle with the global coordinate system Z-axis Set verticality threshold and levelness threshold ,like Then it is determined to be a horizontal component. If so, it is classified as a vertical component; Step 7.4: For horizontal components, based on the intercept parameter in the plane equation... Sort the height information of the representatives, and then... The largest one is marked as the top plate. The smallest component is labeled as the base plate, and the rest as flange plates; for vertical components, they are labeled as web plates, thus completing the semantic segmentation of each part of the component, obtaining the point cloud set of each sub-component, and outputting the point cloud cluster of each sub-component. .

8. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 8 specifically includes the following steps: Step 8.1: For each semantically segmented sub-component point cloud cluster output from Step 7 Establish the general form of the plane equation The overall least squares method is used for parameter estimation, and an objective function is constructed to minimize the sum of squared perpendicular distances from all interior points to the fitting plane. : ， in, For point cloud clusters The number of points in the middle, For point cloud clusters The Middle The coordinates of each point are obtained; the optimal plane parameters are obtained by solving the eigenvalue or singular value decomposition of the objective function. , , , ; Step 8.2: Based on the planar parameters of each sub-component obtained by fitting in Step 8.1, identify two adjacent but non-parallel sub-components, solve the planar equations of the two sub-components, and solve the equation of the intersection of the two planes as the measured edge feature lines of the precast component. Step 8.3: Identify three adjacent but non-coplanar sub-components, establish the plane equations of the three sub-components, and solve for the coordinates of the intersection point of the three surfaces, which will be used as the measured key corner points of the precast component.

9. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 9 specifically includes the following steps: Step 9.1: Extract the sub-component planar parameters and key corner coordinates calculated in Step 8. For the component height and web thickness, determine them by the distance between the corresponding approximate parallel planes on both sides of the component. ， in, Reference plane The normal vector parameters; For plane The intercept parameter; Opposite planes constrained by a plane Equivalent intercept parameters under the same normal vector direction; for each dimension of component length, top plate width, and bottom plate width, determined by key corner points at the end or edge of the component, and the coordinates of the key corner points defining the two ends of this dimension are extracted. and Calculate the Euclidean distance between two points: ， Step 9.2: Convert the standard virtual point cloud from Step 6 into a virtual point cloud. Design values ​​of key structural parameters Compared with the measured geometric dimensions The comparison is performed, and the allowable deviation thresholds for each dimension are set according to the preset quality acceptance specifications. The dimensional deviations are then calculated. , if the deviation If the dimension is within the allowable deviation threshold range, it is considered qualified; otherwise, it is considered unqualified.

10. The handheld image point cloud and laser point cloud fusion detection method for precast bridge components according to claim 1, characterized in that, Step 10 specifically includes the following steps: Step 10.1: Traverse each planar sub-component point cloud cluster extracted in Step 7. Obtain the optimal plane equation parameters for fitting the sub-component in step 8. , , and ; Step 10.2: Calculate the point cloud clusters of sub-components Each point in Vertical distance residual to the fitting plane of the sub-component The calculation formula is: ， Based on the calculated residuals The root mean square error (RMSE) of the sub-component surface is statistically analyzed to quantitatively evaluate the surface flatness quality of that area. The calculation formula is as follows: ， in Point cloud cluster for this sub-component The total number of point clouds in the cloud; Step 10.3: Based on the precast component quality acceptance specifications, set the allowable deviation threshold for the surface flatness of the components. First, the entire sub-component... With threshold Perform macroscopic comparison and judgment: if If so, the overall flatness of the sub-component is deemed acceptable; if If the flatness of the sub-component exceeds the limit, then the overall flatness of the sub-component is determined to be excessive; secondly, the residuals at each point are... With threshold Compare: If The surface flatness at that point is deemed acceptable; if The surface flatness at that point is determined to be out of tolerance. Finally, the flatness analysis results of all sub-components are mapped back to the three-dimensional model to generate a three-dimensional full-field deviation chromatogram covering the entire surface of the component, which intuitively represents the unevenness mass distribution of each part of the beam surface.

Images