Method for detecting defects in a diversion tunnel based on a three-dimensional model

Abstract

This invention relates to the field of tunnel detection and identification technology, specifically a method for detecting defects in water diversion tunnels based on a 3D model. The method includes: collecting laser point cloud data, panoramic image sequences, and ground-penetrating radar (GPR) profile data of the target water diversion tunnel section to form a multi-source detection data set. A 3D mesh model of the tunnel structure is reconstructed using the laser point cloud data. The panoramic image sequence is mapped onto the model to generate a textured 3D model. Abnormal signal segments from the GPR profile data are identified to obtain a set of radar anomaly depth markers, which are then projected onto the textured 3D model to construct a 3D detection model with defect markers. Candidate defect regions are extracted through region growing and segmentation. The opening width and extension length of each candidate defect region are quantified. Based on geometric features, the defect type is classified, and a defect detection report is output. This method achieves the fusion and correlation of multi-source detection data in 3D space, completing automatic defect region segmentation and geometric feature quantification.

Description

Technical Field

[0001] This invention relates to the field of tunnel detection and identification technology, and in particular to a method for detecting defects in water diversion tunnels based on three-dimensional models. Background Technology

[0002] Water diversion tunnels develop various structural defects during long-term operation. Current defect detection operations for water diversion tunnels mostly rely on single sensing devices for data acquisition, with different types of detection data processed and stored separately. A collaborative operation mode for multi-source heterogeneous data processing has not been established. The detection process only uses image acquisition equipment or ground-penetrating radar equipment, lacking the three-dimensional spatial structural foundation built from laser point clouds, and thus failing to reconstruct the complete relationship between the tunnel's overall spatial configuration and surface texture.

[0003] Current detection methods primarily rely on two-dimensional planar data for anomaly identification and analysis. However, the profile anomaly signals generated by ground-penetrating radar cannot establish a precise correspondence with the actual spatial location of the tunnel, and anomaly markers cannot be located at corresponding positions on the physical structure. Two-dimensional data systems struggle to accurately delineate the boundaries of defect areas, cannot quantitatively record the geometric parameters of the defects themselves, and rely on subjective identification based on human experience for the classification of defects.

[0004] Traditional inspection methods lack automated area segmentation and feature quantification processes, and there are no unified geometric parameter evaluation standards as a basis for defect classification. They involve a significant amount of manual intervention, and the inspection process lacks a standardized and unified system. This type of operation mode makes it difficult to achieve spatial integration and standardized classification of defect information, fails to generate structured inspection results, and cannot meet the actual operational needs of refined and three-dimensional defect inspection in water diversion tunnels. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for detecting defects in water diversion tunnels based on three-dimensional models.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for detecting defects in water diversion tunnels based on a three-dimensional model, comprising:

[0007] Acquire a multi-source detection data set for the target water diversion tunnel section, the multi-source detection data set including laser point cloud data, panoramic image sequence and ground-penetrating radar profile data;

[0008] The laser point cloud data is subjected to three-dimensional surface reconstruction processing to generate a three-dimensional mesh model of the tunnel structure;

[0009] The panoramic image sequence is mapped onto the three-dimensional mesh model of the tunnel structure to generate a texture-mapped three-dimensional model.

[0010] The ground-penetrating radar profile data is used to identify anomalous signal segments, resulting in a set of radar anomalous depth markers.

[0011] The radar anomaly depth marker set is projected onto the texture mapping 3D model to generate a 3D detection model with defect markers;

[0012] The three-dimensional detection model with defect markings is subjected to region growing and segmentation to extract multiple candidate defect regions;

[0013] Geometric feature quantization is performed on each candidate defect region to obtain the opening width and extension length of each candidate defect region;

[0014] The candidate defect areas are classified into different types based on the opening width and the extension length, and a defect detection report for the water diversion tunnel is output.

[0015] As a further aspect of the present invention, the step of performing three-dimensional surface reconstruction processing on the laser point cloud data to generate a three-dimensional mesh model of the tunnel structure specifically includes:

[0016] The laser point cloud data is subjected to voxel downsampling filtering to obtain homogenized point cloud data;

[0017] The normal vectors of the homogenized point cloud data are redirected so that all normal vectors point towards the direction of the tunnel's central axis.

[0018] An initial triangular mesh is constructed based on the homogenized point cloud data, and the manifold edges of the initial triangular mesh are repaired to obtain a manifold triangular mesh.

[0019] The manifold triangular mesh is subjected to Laplacian smoothing iteration to eliminate abrupt changes in normal between adjacent triangular faces;

[0020] The smoothed manifold triangular mesh is output as the three-dimensional mesh model of the tunnel structure.

[0021] As a further aspect of the present invention, the step of mapping the panoramic image sequence to the three-dimensional mesh model of the tunnel structure to generate a texture-mapped three-dimensional model specifically includes:

[0022] Radial distortion correction is performed on each image in the panoramic image sequence to obtain the corrected image sequence;

[0023] Extract the vertex coordinates of each triangular facet in the three-dimensional mesh model of the tunnel structure, and transform the vertex coordinates to the camera coordinate system corresponding to the corrected image sequence;

[0024] Based on the transformed vertex coordinates, the visibility occlusion relationship of each triangular facet in each corrected image is calculated, and a visibility determination matrix is ​​generated.

[0025] Based on the visibility determination matrix, the best matching image is selected for each triangular facet, and the texture pixels in the best matching image are assigned to the corresponding triangular facet.

[0026] The textures of all triangular facets are blended at the seams to generate the texture mapping 3D model.

[0027] As a further aspect of the present invention, the step of identifying anomalous signal segments in the ground-penetrating radar profile data to obtain a radar anomaly depth marker set specifically includes:

[0028] Background removal filtering is performed on the ground-penetrating radar profile data to obtain denoised radar profile data;

[0029] The signal amplitude is scanned channel by channel in the denoised radar profile data to locate the abnormal waveforms whose amplitude exceeds a preset background noise threshold.

[0030] Hyperbolic fitting is performed at each abnormal waveform location to determine the top and bottom depths of the abnormal body;

[0031] The interval between the top depth and the bottom depth is marked as an abnormal depth interval;

[0032] All abnormal depth intervals are organized according to the mileage position of the corresponding radar survey line to generate the radar abnormal depth marker set.

[0033] As a further aspect of the present invention, the step of projecting the radar anomaly depth marker set onto the texture mapping 3D model to generate a 3D detection model with defect markers specifically includes:

[0034] Extract the radar survey line mileage value corresponding to each abnormal depth interval in the radar abnormal depth marker set;

[0035] The radar survey line mileage values ​​are mapped onto the longitudinal section coordinates of the texture mapping 3D model to obtain an anomaly projection point sequence;

[0036] Using each abnormal projection point as a seed point, the depth range corresponding to the abnormal depth interval is extended into the interior of the model along the radial direction of the texture-mapped 3D model to generate a 3D abnormal volume bounding box.

[0037] In the texture-mapped 3D model, the voxels inside the bounding box of the 3D anomaly are assigned defect category labels;

[0038] The texture-mapped 3D model with defect category labels is output as the 3D detection model with defect labels.

[0039] As a further aspect of the present invention, the step of performing region growing and segmentation on the three-dimensional detection model with defect markings to extract multiple candidate defect regions specifically includes:

[0040] Traverse all voxels with defect category labels in the three-dimensional detection model with defect labels, and select unvisited voxels as initial seed voxels.

[0041] Starting from the initial seed voxel, search for neighboring voxels with the same defect category label as the current voxel within a 26-neighborhood range, and merge the searched neighboring voxels into the same growth region.

[0042] Repeat the adjacent voxel search and incorporation operation until there are no more adjacent voxels that meet the conditions in the neighborhood of all voxels in the current growth region.

[0043] Mark the current growth region as an independent defect region and continue traversing the next unvisited voxel with a defect category label;

[0044] All independent defect regions are output as candidate defect regions.

[0045] As a further aspect of the present invention, the step of searching for neighboring voxels with the same defect category label as the current voxel within a 26-neighborhood range, starting from the initial seed voxel, and merging the searched neighboring voxels into the same growth region, specifically includes:

[0046] Obtain the three-dimensional coordinates of the current voxel, and calculate the coordinate offset vector of each of its twenty-six neighboring voxels based on the three-dimensional coordinates.

[0047] For each neighboring voxel, check whether the neighboring voxel is marked as visited.

[0048] If the neighboring voxel is not marked as visited and its defect category label is consistent with the current voxel's defect category label, then the index of the neighboring voxel is added to the growth queue.

[0049] Take the first voxel in the growth queue as the new current voxel and repeat the neighborhood voxel check operation.

[0050] When the growth queue is empty, stop the neighborhood search and output the index set of all voxels in the current growth region.

[0051] As a further aspect of the present invention, the step of geometric feature quantization of each candidate defect region to obtain the opening width and extension length of each candidate defect region specifically includes:

[0052] Extract the external exposed surfaces of all voxels in each candidate defect region to obtain a set of defect exposed surface patches;

[0053] Principal component analysis was performed on the set of defect-exposed patches to obtain the first principal component direction and the second principal component direction;

[0054] Project the voxels in each candidate defect region onto the direction of the first principal component, and calculate the difference between the maximum and minimum values ​​of the projected coordinates as the extension length;

[0055] Project the voxels in each candidate defect region onto the direction of the second principal component, and calculate the difference between the maximum and minimum values ​​of the projected coordinates as the opening width.

[0056] As a further aspect of the present invention, the step of classifying the candidate defect area into different types based on the opening width and the extension length, and outputting a water diversion tunnel defect detection report, specifically includes:

[0057] The opening width of each candidate defect region is compared with a preset width grading threshold sequence to determine the width grade of each candidate defect region.

[0058] The extension length of each candidate defect region is compared with a preset length grading threshold sequence to determine the extension level of each candidate defect region.

[0059] Construct a disease type binary for each candidate defect region based on the width level and the extension level;

[0060] All candidate defect regions are classified and statistically analyzed according to the disease type binary group, and the number of defects in each category is statistically analyzed.

[0061] The statistical results are correlated with the position coordinates of each candidate defect region in the three-dimensional detection model with defect markings, and the result is output to generate the water diversion tunnel defect detection report.

[0062] As a further aspect of the present invention, the step of comparing the opening width of each candidate defect region with a preset width grading threshold sequence to determine the width grade of each candidate defect region specifically includes:

[0063] Obtain a preset width grading threshold sequence, wherein the width grading threshold sequence contains multiple width boundary values ​​arranged in ascending order;

[0064] The opening width of each candidate defect region is compared sequentially with the plurality of width boundary values;

[0065] When the opening width is less than the current width boundary value, the level corresponding to the current width boundary value is determined as the width level of the candidate defect area;

[0066] When the opening width is greater than or equal to all width boundary values, the highest level is determined as the width level of the candidate defect region;

[0067] Record the width level of each candidate defect area as the basis for subsequent disease type classification.

[0068] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0069] A multi-source detection data system is constructed by integrating laser point cloud data, panoramic image sequences, and ground-penetrating radar profile data. Laser point cloud data is used to reconstruct a 3D mesh model of the tunnel structure, and the panoramic image sequences are then mapped onto the 3D mesh model to form a texture-mapped 3D model. Multiple heterogeneous detection data are incorporated into the same 3D spatial architecture for unified processing, breaking the inherent pattern of independent and discrete processing of various detection data. Tunnel spatial structure information and surface texture information are synchronously integrated, and ground-penetrating radar detection data can be matched to spatial locations based on the 3D model, constructing a complete digital carrier of the tunnel.

[0070] After identifying anomalous signal segments in the ground-penetrating radar profile data, the radar anomaly depth marker set is projected onto a texture-mapped 3D model. Candidate defect regions are then automatically extracted using a region growing segmentation algorithm. Geometric feature quantification is performed on each candidate defect region to obtain opening width and extension length parameters, which are then used as the basis for disease type classification. This approach abandons reliance on manual experience for disease identification; defect region boundary delineation follows a fixed algorithmic logic, defect geometric attributes are presented in numerical form, and disease classification is performed according to a unified quantification standard.

[0071] All defect marking, region segmentation, feature quantification, and disease classification processes are completed in a closed loop using a 3D detection model. Data from each detection stage maintains spatial correlation, and the location of defects corresponds precisely to the overall tunnel structure. This reduces the involvement of subjective human judgment, ensures unified execution standards for each detection process, forms a standardized processing chain for defect detection, and guarantees traceability and standardization of detection results, thus meeting the needs of batch and standardized defect detection operations in water diversion tunnels. Attached Figure Description

[0072] Figure 1 This is a state diagram of the water diversion tunnel defect detection method based on a three-dimensional model as described in this invention;

[0073] Figure 2 Flowchart for generating a 3D mesh model of a tunnel structure;

[0074] Figure 3 This is a flowchart of the process for region growing and segmentation and candidate defect region extraction. Detailed Implementation

[0075] 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.

[0076] See Figure 1 The specific implementation method of a water diversion tunnel defect detection method based on a three-dimensional model is as follows.

[0077] A multi-source detection data set for the target water diversion tunnel section is acquired, including laser point cloud data, panoramic image sequences, and ground-penetrating radar profile data. The laser point cloud data undergoes 3D surface reconstruction processing to generate a 3D mesh model of the tunnel structure. The panoramic image sequences are mapped onto the 3D mesh model to generate a texture-mapped 3D model. Anomaly signal sections are identified from the ground-penetrating radar profile data to obtain a set of radar anomaly depth markers. This set is projected onto the texture-mapped 3D model to generate a 3D detection model with defect markers. Region growing and segmentation are performed on the 3D detection model with defect markers to extract multiple candidate defect regions. Geometric feature quantization is performed on each candidate defect region to obtain its opening width and extension length. Based on the opening width and extension length, the candidate defect regions are classified into different defect types, and a water diversion tunnel defect detection report is output.

[0078] In one embodiment of the present invention, see [reference] Figure 2 The laser point cloud data is voxel downsampling filtered to obtain homogenized point cloud data. Normal vectors are then redirected to point towards the tunnel's central axis. An initial triangular mesh is constructed based on the homogenized point cloud data, and manifold edge repair is performed to obtain a manifold triangular mesh. Laplacian smoothing iteration is then applied to the manifold triangular mesh to eliminate abrupt changes in normal vectors between adjacent triangular faces. The smoothed manifold triangular mesh is output as a 3D mesh model of the tunnel structure. Radial distortion correction is performed on each image in the panoramic image sequence to obtain a corrected image sequence. The vertex coordinates of each triangular facet in the 3D mesh model of the tunnel structure are extracted and transformed to the camera coordinate system corresponding to the corrected image sequence. Based on the transformed vertex coordinates, the visibility occlusion relationship of each triangular facet in each corrected image is calculated, generating a visibility determination matrix. Based on the visibility determination matrix, the best matching image is selected for each triangular facet, and the texture pixels in the best matching image are assigned to the corresponding triangular facet. The textures of all triangular facets are then color-blended at the seams to generate a texture-mapped 3D model.

[0079] In the specific implementation, voxel downsampling filtering is performed on the laser point cloud data to obtain homogenized point cloud data. Voxel downsampling filtering divides the three-dimensional space where the laser point cloud data is located into voxel grids of fixed size. Within each voxel grid, the laser point cloud data point closest to the center of the voxel grid is retained, and the remaining laser point cloud data points within the voxel grid are discarded, thus obtaining homogenized point cloud data with uniform spatial distribution. In the specific implementation, the normal vectors of the homogenized point cloud data are redirected so that all normal vectors point towards the direction of the tunnel's central axis. Normal vector redirection is achieved by calculating the initial normal vector direction of each homogenized point cloud data point through the local neighborhood covariance matrix, calculating the coordinates of the centroid of the homogenized point cloud data as a whole, and performing a dot product operation between the initial normal vector direction of each homogenized point cloud data point and the direction vector from that homogenized point cloud data point to the centroid. If the dot product result is negative, the normal vector direction of that homogenized point cloud data point is reversed, thus ensuring that all normal vectors point towards the direction of the tunnel's central axis.

[0080] In some embodiments, an initial triangular mesh is constructed based on homogenized point cloud data, and manifold edge repair is performed on the initial triangular mesh to obtain a manifold triangular mesh. The initial triangular mesh is constructed using the Poisson surface reconstruction method, taking homogenized point cloud data as input, and obtaining the initial triangular mesh by constructing an octree and solving the Poisson equation. Manifold edge repair detects each edge in the initial triangular mesh that is shared by more than two triangular faces. For non-manifold edges shared by three or more triangular faces, the redundant triangular faces sharing the edge are deleted and the vertices are reconnected to ensure that each edge is shared by at most two triangular faces, thus obtaining a manifold triangular mesh. In some embodiments, Laplacian smoothing iteration is performed on the manifold triangular mesh to eliminate abrupt changes in normal between adjacent triangular faces. The Laplacian smoothing iteration updates the position of each vertex according to the formula:

[0081]

[0082] in: Let represent the current coordinate vector of the i-th vertex in the triangular mesh of the manifold. This represents the updated coordinate vector of the i-th vertex in the triangular mesh of the manifold. Let represent the set of indices of all vertices adjacent to the i-th vertex. This represents the number of vertices adjacent to the i-th vertex. Let represent the coordinate vector of the j-th vertex adjacent to the i-th vertex. The smoothing factor is represented by a positive number. The above update process is iteratively executed until the maximum value of all vertex displacements between two adjacent iterations is less than the preset convergence threshold. The smoothed manifold triangular mesh is then output as a three-dimensional mesh model of the tunnel structure.

[0083] In the specific implementation, radial distortion correction is performed on each image in the panoramic image sequence to obtain the corrected image sequence. The radial distortion correction adopts the Brownian distortion model. For each pixel in each image in the panoramic image sequence, the radial distortion offset is calculated based on the radial distance from the pixel to the principal point of the image. The original coordinates of each pixel are subtracted from the radial distortion offset to obtain the corrected pixel coordinates. The pixel value corresponding to the corrected pixel coordinates is then assigned to the new image to obtain the corrected image sequence. In the specific implementation, the vertex coordinates of each triangular facet in the 3D mesh model of the tunnel structure are extracted and transformed to the camera coordinate system corresponding to the corrected image sequence. The transformation process is achieved by multiplying the vertex coordinates of each triangular facet by the transformation matrix from the world coordinate system of the 3D mesh model of the tunnel structure to the camera coordinate system corresponding to each corrected image. This transformation matrix includes rotation and translation components. The rotation component is composed of the camera attitude angle, and the translation component is composed of the camera's position coordinates in the world coordinate system.

[0084] Optionally, the visibility occlusion relationship of each triangular facet in each corrected image is calculated based on the converted vertex coordinates, generating a visibility determination matrix. When calculating the visibility occlusion relationship, for each triangular facet, the three vertices of the triangular facet are projected onto the pixel plane of the corrected image to obtain three projection points. The triangular region enclosed by the three projection points is defined as the projection region of the triangular facet in the corrected image. A ray is emitted from the optical center of the camera towards the center point of the triangular facet. It is determined whether the ray intersects with other triangular facets in the 3D mesh model of the tunnel structure except for the current triangular facet. If there is no intersection, the current triangular facet is marked as visible in the corrected image. If there is an intersection, it is marked as invisible. The visibility states of all triangular facets in all corrected images are organized into a two-dimensional matrix. The rows of the two-dimensional matrix correspond to the triangular facet index, the columns correspond to the corrected image index, and the matrix elements are visible or invisible markers. This two-dimensional matrix is ​​the visibility determination matrix.

[0085] Optionally, based on the visibility determination matrix, the best matching image is selected for each triangular facet, and the texture pixels in the best matching image are assigned to the corresponding triangular facet. When selecting the best matching image, for each triangular facet, the row corresponding to the triangular facet in the visibility determination matrix is ​​traversed, all corrected images marked as visible are collected, the ratio of the projected area of ​​the triangular facet in each visible corrected image to the original area of ​​the triangular facet is calculated, and the corrected image with the largest ratio is selected as the best matching image. All pixels contained in the projection area of ​​the triangular facet in the best matching image are backsampled according to the projection mapping relationship to obtain the texture pixel value corresponding to each pixel, and the texture pixel value is assigned to the corresponding point on the triangular facet.

[0086] It is understandable that the textures of all triangular facets are color-fused at the seams to generate a texture-mapped 3D model. The color fusion employs a multi-band fusion method, detecting pixel differences at the texture seams between adjacent triangular facets. For each seam, the color difference between texture pixels on both sides of the seam is calculated. Gaussian and Laplacian pyramids are constructed to decompose the color difference at multiple scales, and feathering fusion is performed at each scale. The fused multi-scale result is reconstructed into a complete texture image, which is then remapped onto the tunnel structure's 3D mesh model to obtain the texture-mapped 3D model. In practice, the steps of voxel downsampling filtering of laser point cloud data, normal vector redirection, initial triangular mesh construction, manifold edge repair, Laplacian smoothing iteration, radial distortion correction of panoramic image sequences, vertex coordinate transformation, visibility occlusion calculation, optimal matching image selection, texture pixel assignment, and seam color fusion are executed sequentially. The output texture-mapped 3D model retains the geometric accuracy of the tunnel structure's 3D mesh model while possessing a realistic texture appearance.

[0087] In one embodiment of the present invention, background filtering is performed on the ground-penetrating radar profile data to obtain denoised radar profile data. The signal amplitude is scanned channel by channel in the denoised radar profile data to locate abnormal waveforms whose amplitude exceeds a preset background noise threshold. Hyperbolic fitting is performed on each abnormal waveform location to determine the top and bottom depths of the anomaly. The interval between the top and bottom depths is marked as an abnormal depth interval. All abnormal depth intervals are organized according to the mileage positions of the corresponding radar survey lines to generate a radar abnormal depth marker set. The radar survey line mileage value corresponding to each abnormal depth interval in the radar abnormal depth marker set is extracted. The radar survey line mileage value is mapped onto the longitudinal section coordinates of the texture-mapped 3D model to obtain an abnormal projection point sequence. Using each abnormal projection point sequence as a seed point, the depth range corresponding to the abnormal depth interval is extended radially into the interior of the texture-mapped 3D model to generate a 3D anomaly bounding box. In the texture-mapped 3D model, the voxels inside the 3D anomaly bounding box are assigned defect category labels. The texture-mapped 3D model with defect category labels is output as a 3D detection model with defect labels.

[0088] In practice, background removal filtering is performed on the ground-penetrating radar profile data to obtain denoised radar profile data. Background removal filtering is achieved by calculating the average value of each radar signal in the ground-penetrating radar profile data as the background signal, and subtracting the corresponding background signal amplitude from the amplitude of each sampling point of each radar signal, thereby eliminating system noise and direct wave interference in the ground-penetrating radar profile data. In practice, the signal amplitude is scanned channel by channel in the denoised radar profile data to locate abnormal waveform positions where the amplitude exceeds a preset background noise threshold. The preset background noise threshold is set to three times the root mean square of the amplitude of all sampling points in the denoised radar profile data. During channel scanning, each sampling point of each radar signal is traversed. When the absolute value of the amplitude of a certain sampling point exceeds the preset background noise threshold, the center position of the time window where the sampling point is located is recorded as an abnormal waveform position.

[0089] In some embodiments, hyperbolic fitting is performed at each anomalous waveform location to determine the top and bottom depths of the anomalous body. The hyperbolic fitting uses the standard hyperbolic equation to fit the diffraction wave morphology around the anomalous waveform location.

[0090]

[0091] in: Indicates the radar's two-way travel time. This represents the minimum two-way travel time corresponding to the top of the anomaly. Indicates the horizontal position on the radar survey line. This indicates the horizontal position corresponding to the center of the anomalous body. This represents the length of the semi-axis of the hyperbola on the time axis. The parameters represent the semi-axis lengths of the hyperbola on the spatial axes, obtained by fitting using the least squares method. and Then, the top depth of the anomaly is calculated as the electromagnetic wave propagation speed in concrete multiplied by... Half of the bottom depth of the anomaly is calculated as the electromagnetic wave propagation speed in concrete multiplied by [the value]. Half of the depth; in some embodiments, the interval between the top depth and the bottom depth is marked as an abnormal depth interval, and all abnormal depth intervals are organized according to the mileage position of the corresponding radar survey line to generate a radar abnormal depth mark set. Each abnormal depth interval records its corresponding radar survey line mileage value, top depth value and bottom depth value.

[0092] Optionally, the radar survey line mileage value corresponding to each anomaly depth interval in the radar anomaly depth marker set is extracted, and the radar survey line mileage value is mapped to the longitudinal section coordinates of the texture-mapped 3D model to obtain an anomaly projection point sequence. During the mapping process, the radar survey line mileage value is used as the coordinate value in the longitudinal section direction of the texture-mapped 3D model, and the fixed offset of the radar survey line on the tunnel cross section is used as the lateral coordinate value of the longitudinal section, thereby generating a series of discrete anomaly projection points on the surface of the texture-mapped 3D model. Optionally, using each anomaly projection point sequence as a seed point, along the texture-mapped 3D model... The radial direction extends into the model to represent the depth range corresponding to the abnormal depth interval, generating a 3D anomaly bounding box. The radial direction is defined as the normal direction from the central axis of the tunnel to the tunnel wall. For each seed point, the texture map moves the 3D model surface inward along the radial direction by the distance corresponding to the depth range. A cuboid bounding box is constructed with the seed point and the endpoint after the movement as two diagonal vertices. The bottom surface of the cuboid bounding box is parallel to the tangent plane of the tunnel surface, and the top surface is located on a plane inside the model that is parallel to the bottom surface. This cuboid bounding box is the 3D anomaly bounding box.

[0093] It can be understood that in a texture-mapped 3D model, voxels inside the bounding box of a 3D anomaly are assigned defect category labels. The defect category labels are encoded with integers, and different types of geological defects correspond to different integer values. For each 3D anomaly bounding box, all voxels located inside the bounding box in the texture-mapped 3D model are traversed, and the attribute field of each voxel is set to the corresponding defect category label. It can also be understood that the texture-mapped 3D model with defect category labels is output as a 3D detection model with defect labels. The 3D detection model with defect labels contains the geometric information, texture information, and defect category label information of each voxel of the texture-mapped 3D model. When stored in a 3D mesh file format, the defect category label is written to the file as an additional attribute of each vertex or each triangle facet.

[0094] In one embodiment of the present invention, see [reference] Figure 3The process iterates through all voxels with defect category labels in the 3D detection model with defect markers. Unvisited voxels are selected as initial seed voxels. Starting from the initial seed voxel, a search is conducted within a 26-neighborhood for adjacent voxels with the same defect category label as the current voxel. The searched adjacent voxels are then merged into the same growth region. This process is repeated until there are no more adjacent voxels that meet the criteria in the neighborhood of all voxels in the current growth region. The current growth region is then marked as an independent defect region. The process continues to iterate through the next unvisited voxel with a defect category label and outputs all independent defect regions as candidate defect regions. Starting from the initial seed voxel, the system searches for neighboring voxels with the same defect category label as the current voxel within a 26-neighborhood and incorporates the found neighboring voxels into the same growth region. During this process, the 3D coordinates of the current voxel are obtained, and the coordinate offset vector of each neighboring voxel within its 26-neighborhood is calculated based on these coordinates. For each neighboring voxel, it checks whether the neighboring voxel is marked as visited. If the neighboring voxel is not marked as visited and its defect category label matches that of the current voxel, its index is added to the growth queue. The first voxel in the growth queue is then taken as the new current voxel, and the neighboring voxel checking operation is repeated. When the growth queue is empty, the neighborhood search stops, and the set of indices of all voxels within the current growth region is output.

[0095] In the specific implementation, all voxels with defect category labels in the defect-marked 3D detection model are traversed. Unvisited voxels are selected as initial seed voxels. The defect-marked 3D detection model consists of a voxel array. Each voxel contains a defect category label field and an access flag field. The traversal process checks the defect category label of each voxel in the order of storage in 3D space. If the defect category label is non-zero and the access flag field is in an unvisited state, the voxel is selected as the initial seed voxel. In the specific implementation, starting from the initial seed voxel, neighboring voxels with the same defect category label as the current voxel are searched within a 26-neighborhood range. The searched neighboring voxels are merged into the same growth region. The 26-neighborhood range includes all adjacent positions of the current voxel that are offset by one unit in each of the three directions of X-axis, Y-axis and Z-axis in 3D space, for a total of 26 neighboring voxels.

[0096] In some embodiments, the neighboring voxel search and incorporation operation is repeatedly performed until there are no more neighboring voxels that meet the conditions in the neighborhood of all voxels in the current growth region. The current growth region is marked as an independent defect region, and the next unvisited voxel with a defect category label is traversed. All independent defect regions are output as candidate defect regions. Each independent defect region corresponds to a connected component. The candidate defect region list records the set of all voxel indexes contained in each defect region. In some embodiments, in the process of searching for neighboring voxels with the same defect category label as the current voxel in the twenty-six neighborhoods starting from the initial seed voxel and incorporating the searched neighboring voxels into the same growth region, the three-dimensional coordinate value of the current voxel is obtained, and the coordinate offset vector of each neighboring voxel in its twenty-six neighborhoods is calculated based on the three-dimensional coordinate value. The coordinate offset vector consists of three components, each of which takes the value of negative one, zero or positive one, and the three components cannot be zero at the same time. A total of twenty-six coordinate offset vectors are generated.

[0097] Optionally, for each neighboring voxel, check whether the neighboring voxel is marked as visited. The visited status is obtained by reading the voxel's access flag field. If the access flag field is true, it means that the voxel has been visited; if it is false, it means that it has not been visited. Optionally, if the neighboring voxel is not marked as visited and the defect category label of the neighboring voxel is consistent with the defect category label of the current voxel, then add the index of the neighboring voxel to the growth queue. The growth queue adopts a first-in-first-out data structure. When adding the voxel to the queue, the access flag field of the neighboring voxel is set to visited status. It can be understood that the first voxel in the growth queue is taken as the new current voxel, and the neighborhood voxel check operation is repeatedly performed. Each time, a voxel index is taken from the head of the growth queue, and the corresponding 3D coordinate value is obtained according to the voxel index. The coordinate offset vector of its 26 neighbors is recalculated, and the access status and label consistency are checked for each neighbor voxel. Neighbor voxels that meet the conditions continue to be added to the tail of the growth queue. It can also be understood that when the growth queue is empty, the neighborhood search is stopped and the index set of all voxels in the current growth region is output. The condition for the growth queue to be empty is that there are no voxel indexes to be processed in the queue. At this time, all voxels connected to the initial seed voxel and having the same defect category label have been accessed and recorded. All voxel indexes that have been added to the growth queue during the growth process are summarized to form the index set of the current growth region.

[0098] In practice, the process of repeatedly performing the adjacent voxel search and incorporation operation is carried out according to the following steps: During the first execution, the initial seed voxel is used as the current voxel. All unvisited neighboring voxels with the same defect category label within its 26 neighborhoods are searched. These neighboring voxels are added to the growth queue and marked as visited. The initial seed voxel is also marked as visited. During the second execution, the first voxel is taken from the growth queue as the new current voxel. The 26 neighborhoods of the new current voxel are searched. Neighboring voxels that meet the conditions are added to the tail of the growth queue and marked as visited. Each time a new voxel is searched and added from a voxel, the voxel is removed from the growth queue. The operations of taking out, searching, and adding are repeated until there are no voxels remaining in the growth queue.

[0099] In one embodiment of the present invention, when geometric feature quantization is performed on each candidate defect region to obtain the opening width and extension length of each candidate defect region, the external exposed surfaces of all voxels in each candidate defect region are extracted to obtain a set of defect exposed surface patches. Principal component analysis is performed on the set of defect exposed surface patches to obtain the first principal component direction and the second principal component direction. The voxels in each candidate defect region are projected onto the first principal component direction, and the difference between the maximum and minimum values ​​of the projected coordinates is calculated as the extension length. The voxels in each candidate defect region are projected onto the second principal component direction, and the difference between the maximum and minimum values ​​of the projected coordinates is calculated as the opening width.

[0100] In practice, geometric feature quantization is performed on each candidate defect region to obtain its opening width and extension length. The external exposed surfaces of all voxels within each candidate defect region are extracted, resulting in a set of defect exposed surfaces. An external exposed surface is defined as the surface corresponding to a voxel in the candidate defect region whose at least one neighboring voxel does not belong to the current candidate defect region. Specifically, for each voxel in the candidate defect region, its six face directions are checked. If a neighboring voxel in a certain face direction does not exist or its defect category label differs from that of the current voxel, then that face is extracted as an external exposed surface. All candidate defect regions... All externally exposed surfaces constitute a set of defect-exposed surfaces. In a specific implementation, principal component analysis is performed on the set of defect-exposed surfaces to obtain the first principal component direction and the second principal component direction. Principal component analysis calculates the covariance matrix of the coordinates of all vertex coordinates of the defect-exposed surfaces in the set, solves for the eigenvalues ​​and eigenvectors of the covariance matrix, takes the eigenvector corresponding to the largest eigenvalue as the first principal component direction, and takes the eigenvector corresponding to the second largest eigenvalue as the second principal component direction. The first principal component direction represents the direction with the largest variance in the set of defect-exposed surfaces, and the second principal component direction is orthogonal to the first principal component direction and represents the direction with the largest remaining variance.

[0101] In some embodiments, voxels in each candidate defect region are projected onto the direction of the first principal component, and the difference between the maximum and minimum projected coordinates is calculated as the extension length. For each voxel in the candidate defect region, the projection process calculates the dot product of the voxel's center point coordinate vector and the first principal component direction vector to obtain the voxel's projected coordinate value in the first principal component direction. After traversing all voxels in the candidate defect region, a sequence of projected coordinate values ​​is obtained. The maximum and minimum values ​​in the sequence are found, and the difference between the maximum and minimum values ​​is calculated; this difference is the extension length. In some embodiments, voxels in each candidate defect region are projected onto the direction of the second principal component, and the difference between the maximum and minimum projected coordinates is calculated as the opening width. For each voxel in the candidate defect region, the projection process calculates the dot product of the voxel's center point coordinate vector and the second principal component direction vector to obtain the voxel's projected coordinate value in the second principal component direction. After traversing all voxels in the candidate defect region, a sequence of projected coordinate values ​​is obtained. The maximum and minimum values ​​in the sequence are found, and the difference between the maximum and minimum values ​​is calculated; this difference is the opening width.

[0102] Optionally, before calculating the projected coordinate values, the center point coordinates of all voxels in the candidate defect region are centered. The centering process converts the original center point coordinates of each voxel into translated coordinates according to the formula:

[0103]

[0104] in: This represents the original center point coordinate vector of the k-th voxel. This indicates the total number of voxels in the current candidate defect region. This represents the original center point coordinate vector of the i-th voxel. This represents the coordinate vector centered on the k-th voxel. The mean vector of the center point coordinates of all voxels is represented by a vector. After centering, the calculation of the projected coordinate values ​​is based on a coordinate distribution with zero mean. Optionally, after centering the center point coordinates of all voxels in each candidate defect region, the dot product of the centered coordinate vector and the first principal component direction vector is calculated to obtain the projection value of the extension length, and the dot product of the centered coordinate vector and the second principal component direction vector is calculated to obtain the projection value of the opening width.

[0105] It can be understood that the first principal component direction corresponds to the main extension direction of the defect region, and the second principal component direction corresponds to the width direction of the defect region perpendicular to the main extension direction. The extension length characterizes the scale of the defect on the tunnel axis or main orientation, and the opening width characterizes the degree of opening of the defect in the direction orthogonal to the extension direction. It can also be understood that for each candidate defect region, the above-mentioned external exposure surface extraction, principal component analysis and projection calculation operations are performed independently to obtain the opening width value and extension length value of each candidate defect region. These values ​​are stored in floating-point form and associated with the corresponding candidate defect region identifier.

[0106] In one embodiment of the present invention, when classifying the defect type of candidate defect areas according to the opening width and extension length, the opening width of each candidate defect area is compared with a preset width classification threshold sequence to determine the width level of each candidate defect area, and the extension length of each candidate defect area is compared with a preset length classification threshold sequence to determine the extension level of each candidate defect area. A defect type binary group is constructed for each candidate defect area based on the width level and extension level. All candidate defect areas are classified and statistically analyzed according to the defect type binary group to generate the quantity statistics of each category of defects. The quantity statistics are correlated with the position coordinates of each candidate defect area in the three-dimensional detection model with defect markers and output to generate a water diversion tunnel defect detection report. When determining the width level of each candidate defect region by comparing its opening width with a preset width grading threshold sequence, the preset width grading threshold sequence is obtained. This sequence contains multiple width boundary values ​​arranged in ascending order. The opening width of each candidate defect region is compared with these multiple width boundary values ​​sequentially. When the opening width is less than the current width boundary value, the level corresponding to the current width boundary value is determined as the width level of the candidate defect region. When the opening width is greater than or equal to all width boundary values, the highest level is determined as the width level of the candidate defect region. The width level of each candidate defect region is recorded as the basis for subsequent disease type grading.

[0107] In practice, candidate defect areas are classified into disease types based on their opening width and extension length. The opening width of each candidate defect area is compared with a preset width classification threshold sequence to determine the width level of each candidate defect area. The width classification threshold sequence contains multiple width boundary values ​​arranged in ascending order, and each width boundary value corresponds to a width level. In practice, the extension length of each candidate defect area is compared with a preset length classification threshold sequence to determine the extension level of each candidate defect area. The length classification threshold sequence contains multiple length boundary values ​​arranged in ascending order, and each length boundary value corresponds to an extension level. A disease type tuple is constructed for each candidate defect area based on the width level and extension level. The disease type tuple is represented in the form of (width level value, extension level value).

[0108] In some embodiments, when determining the width level of each candidate defect region by comparing its opening width with a preset width grading threshold sequence, a preset width grading threshold sequence is obtained. This sequence contains multiple width boundary values ​​arranged in ascending order. The opening width of each candidate defect region is compared sequentially with these multiple width boundary values. When the opening width is less than the current width boundary value, the level corresponding to the current width boundary value is determined as the width level of the candidate defect region. When the opening width is greater than or equal to all width boundary values, the highest level is determined as the width level of the candidate defect region. The width level of each candidate defect region is recorded as the basis for subsequent disease type grading. In some embodiments, a grading formula is used to calculate the width level value.

[0109]

[0110] in: Indicates the opening width of the candidate defect region. This indicates the preset width step size. Indicates the preset maximum width level. This indicates the rounding up operation. and These represent operations to retrieve the minimum and maximum values, respectively. This represents the calculated width level value.

[0111] Optionally, all candidate defect regions are classified and statistically analyzed according to the defect type tuple, generating statistical results of the number of defects in each category. The classification and statistical process traverses all candidate defect regions. For each candidate defect region, its defect type tuple is extracted. In the statistical dictionary, the tuple is used as the key value, and the corresponding count is incremented by one. After the traversal is completed, each key value in the statistical dictionary corresponds to a defect category, and the number associated with each key value is the number of defects in that category. Optionally, the statistical results are associated with the position coordinates of each candidate defect region in the 3D detection model with defect markings and output to generate a water diversion tunnel defect detection report. The position coordinates are represented by the 3D coordinates of the geometric center point of each candidate defect region in the 3D detection model with defect markings. The geometric center point is obtained by calculating the arithmetic mean of the coordinates of the center points of all voxels in the candidate defect region.

[0112] It is understandable that in the process of classifying defects, the width level and the extension level can be divided independently. The width level reflects the degree of opening of the defect in the opening direction, and the extension level reflects the length of the defect in the extension direction. The defect type tuple integrates the information of these two dimensions. It is also understandable that the generated water diversion tunnel defect detection report contains the following: a unique identifier for each candidate defect area, a defect type tuple, the opening width value, the extension length value, the coordinates of the geometric center point, and the total number of defects of each type after classification and statistics according to the defect type tuple. The water diversion tunnel defect detection report is output in a structured text format, with each line recording the information of a defect area, and a classification and statistical summary table attached.

[0113] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for detecting defects in water diversion tunnels based on three-dimensional models, characterized in that, The method includes: Acquire a multi-source detection data set for the target water diversion tunnel section, the multi-source detection data set including laser point cloud data, panoramic image sequence and ground-penetrating radar profile data; The laser point cloud data is subjected to three-dimensional surface reconstruction processing to generate a three-dimensional mesh model of the tunnel structure; The panoramic image sequence is mapped onto the three-dimensional mesh model of the tunnel structure to generate a texture-mapped three-dimensional model. The ground-penetrating radar profile data is used to identify anomalous signal segments, resulting in a set of radar anomalous depth markers. The radar anomaly depth marker set is projected onto the texture mapping 3D model to generate a 3D detection model with defect markers; The three-dimensional detection model with defect markings is subjected to region growing and segmentation to extract multiple candidate defect regions; Geometric feature quantization is performed on each candidate defect region to obtain the opening width and extension length of each candidate defect region; The candidate defect areas are classified into different types based on the opening width and the extension length, and a defect detection report for the water diversion tunnel is output.

2. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 1, characterized in that, The steps of performing three-dimensional surface reconstruction processing on the laser point cloud data to generate a three-dimensional mesh model of the tunnel structure specifically include: The laser point cloud data is subjected to voxel downsampling filtering to obtain homogenized point cloud data; The normal vectors of the homogenized point cloud data are redirected so that all normal vectors point towards the direction of the tunnel's central axis. An initial triangular mesh is constructed based on the homogenized point cloud data, and the manifold edges of the initial triangular mesh are repaired to obtain a manifold triangular mesh. The manifold triangular mesh is subjected to Laplacian smoothing iteration to eliminate abrupt changes in normal between adjacent triangular faces; The smoothed manifold triangular mesh is output as the three-dimensional mesh model of the tunnel structure.

3. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 1, characterized in that, The step of mapping the panoramic image sequence to the three-dimensional mesh model of the tunnel structure to generate a texture-mapped three-dimensional model specifically includes: Radial distortion correction is performed on each image in the panoramic image sequence to obtain the corrected image sequence; Extract the vertex coordinates of each triangular facet in the three-dimensional mesh model of the tunnel structure, and transform the vertex coordinates to the camera coordinate system corresponding to the corrected image sequence; Based on the transformed vertex coordinates, the visibility occlusion relationship of each triangular facet in each corrected image is calculated, and a visibility determination matrix is ​​generated. Based on the visibility determination matrix, the best matching image is selected for each triangular facet, and the texture pixels in the best matching image are assigned to the corresponding triangular facet. The textures of all triangular facets are blended at the seams to generate the texture mapping 3D model.

4. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 1, characterized in that, The steps of identifying anomalous signal segments in the ground-penetrating radar profile data to obtain a set of radar anomaly depth markers specifically include: Background removal filtering is performed on the ground-penetrating radar profile data to obtain denoised radar profile data; The signal amplitude is scanned channel by channel in the denoised radar profile data to locate the abnormal waveforms whose amplitude exceeds a preset background noise threshold. Hyperbolic fitting is performed at each abnormal waveform location to determine the top and bottom depths of the abnormal body; The interval between the top depth and the bottom depth is marked as an abnormal depth interval; All abnormal depth intervals are organized according to the mileage position of the corresponding radar survey line to generate the radar abnormal depth marker set.

5. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 1, characterized in that, The step of projecting the radar anomaly depth marker set onto the texture mapping 3D model to generate a 3D detection model with defect markers specifically includes: Extract the radar survey line mileage value corresponding to each abnormal depth interval in the radar abnormal depth marker set; The radar survey line mileage values ​​are mapped onto the longitudinal section coordinates of the texture mapping 3D model to obtain an anomaly projection point sequence; Using each abnormal projection point as a seed point, the depth range corresponding to the abnormal depth interval is extended into the interior of the model along the radial direction of the texture-mapped 3D model to generate a 3D abnormal volume bounding box. In the texture-mapped 3D model, the voxels inside the bounding box of the 3D anomaly are assigned defect category labels; The texture-mapped 3D model with defect category labels is output as the 3D detection model with defect labels.

6. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 1, characterized in that, The steps of performing region growing and segmentation on the 3D detection model with defect markers to extract multiple candidate defect regions specifically include: Traverse all voxels with defect category labels in the three-dimensional detection model with defect labels, and select unvisited voxels as initial seed voxels. Starting from the initial seed voxel, search for neighboring voxels with the same defect category label as the current voxel within a 26-neighborhood range, and merge the searched neighboring voxels into the same growth region. Repeat the adjacent voxel search and incorporation operation until there are no more adjacent voxels that meet the conditions in the neighborhood of all voxels in the current growth region. Mark the current growth region as an independent defect region and continue traversing the next unvisited voxel with a defect category label; All independent defect regions are output as candidate defect regions.

7. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 6, characterized in that, Starting with the initial seed voxel, the step of searching for neighboring voxels with the same defect category label as the current voxel within a 26-neighborhood range, and merging the searched neighboring voxels into the same growth region, specifically includes: Obtain the three-dimensional coordinates of the current voxel, and calculate the coordinate offset vector of each of its twenty-six neighboring voxels based on the three-dimensional coordinates. For each neighboring voxel, check whether the neighboring voxel is marked as visited. If the neighboring voxel is not marked as visited and its defect category label is consistent with the current voxel's defect category label, then add the index of the neighboring voxel to the growth queue. Take the first voxel in the growth queue as the new current voxel and repeat the neighborhood voxel check operation. When the growth queue is empty, stop the neighborhood search and output the index set of all voxels in the current growth region.

8. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 1, characterized in that, The step of performing geometric feature quantization on each candidate defect region to obtain the opening width and extension length of each candidate defect region specifically includes: Extract the external exposed surfaces of all voxels in each candidate defect region to obtain a set of defect exposed surface patches; Principal component analysis was performed on the set of defect-exposed patches to obtain the first principal component direction and the second principal component direction; Project the voxels in each candidate defect region onto the direction of the first principal component, and calculate the difference between the maximum and minimum values ​​of the projected coordinates as the extension length; Project the voxels in each candidate defect region onto the direction of the second principal component, and calculate the difference between the maximum and minimum values ​​of the projected coordinates as the opening width.

9. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 1, characterized in that, The steps of classifying the candidate defect areas according to the opening width and the extension length, and outputting a water diversion tunnel defect detection report, specifically include: The opening width of each candidate defect region is compared with a preset width grading threshold sequence to determine the width grade of each candidate defect region. The extension length of each candidate defect region is compared with a preset length grading threshold sequence to determine the extension level of each candidate defect region. Construct a disease type binary for each candidate defect region based on the width level and the extension level; All candidate defect regions are classified and statistically analyzed according to the disease type binary group, and the number of defects in each category is statistically analyzed. The statistical results are correlated with the position coordinates of each candidate defect region in the three-dimensional detection model with defect markings, and the result is output to generate the water diversion tunnel defect detection report.

10. The method for detecting defects in water diversion tunnels based on a three-dimensional model according to claim 9, characterized in that, The step of comparing the opening width of each candidate defect region with a preset width grading threshold sequence to determine the width level of each candidate defect region specifically includes: Obtain a preset width grading threshold sequence, wherein the width grading threshold sequence contains multiple width boundary values ​​arranged in ascending order; The opening width of each candidate defect region is compared sequentially with the plurality of width boundary values; When the opening width is less than the current width boundary value, the level corresponding to the current width boundary value is determined as the width level of the candidate defect area; When the opening width is greater than or equal to all width boundary values, the highest level is determined as the width level of the candidate defect region; Record the width level of each candidate defect area as the basis for subsequent disease type classification.

Images