Bridge defect automatic detection and quantitative analysis method based on point cloud data processing

Abstract

The invention belongs to the technical field of bridge detection, and discloses a point cloud data processing-based bridge defect automatic detection and quantitative analysis method, which comprises the following steps of: firstly, acquiring bridge point cloud data; secondly, preprocessing the acquired three-dimensional point cloud data; then, bridge component segmentation is carried out; secondly, performing defect detection on each component unit by using a region growing clustering segmentation algorithm; then, a Kriging interpolation method is adopted to carry out intelligent repair and reconstruction on the identified defect part, and missing information is filled; and finally, carrying out defect quantitative analysis. According to the method, automatic and refined processing is achieved in the whole process from data acquisition and preprocessing to defect recognition and reconstruction and quantitative analysis, the efficiency and accuracy of bridge detection are remarkably improved, and multiple key technical steps from data acquisition to quantitative analysis are effectively integrated; the bridge detection process is optimized, and the automation level of the work and the accuracy of the result are improved.

Description

Technical Field

[0002] The invention relates to the technical field of bridge detection, and specifically discloses an automatic detection and quantitative analysis method for bridge defects based on point cloud data processing. Background Art

[0003] Bridges are a core component of the transportation system, bearing the growing pressure of traffic flow and freight. As time goes by and under the influence of the natural environment, bridge structures face a variety of safety hazards, including deformation, cracks, spalling and corrosion, which pose a direct threat to the structural safety and service life of bridges.

[0004] Traditional bridge defect detection methods mainly rely on manual inspections, and professional tools such as bridge inspection vehicles and crack width gauges are also often used. During the inspection process, special vehicles equipped with sensors move along the surface of the bridge to collect structural data in real time, while staff use width gauges to read the specific information of the cracks and manually record their location, width, length and area. This method often requires traffic closures, which is not only time-consuming and costly, but also the inspection results are affected by the technical level and experience of the personnel, are highly subjective and have certain blind spots, and are prone to omissions or misjudgments of potential problems.

[0005] Chinese Patent Name: Intelligent Detection Method for Bridge Construction Based on Image Processing, Authorization Announcement Number: CN116258722A, Authorization Announcement Date: June 13, 2023. The present invention is based on the characteristic area of ​​each target window to obtain the initial contour of the level set segmentation, so as to accurately obtain the complete area of ​​the defect, and then correct the target window of the target detection, so as to more accurately locate and detect the defect area. It is based on the measurement and analysis of two-dimensional images, and only performs defect detection. This traditional measurement method is often difficult to efficiently and comprehensively identify all potential defects, especially those small but may cause serious consequences. These limitations not only affect the comprehensive assessment of the health of the bridge, but also bring great challenges to the maintenance and repair of the bridge. Therefore, it is necessary to design a bridge defect automatic detection and quantitative analysis method based on point cloud data processing to solve the above problems. Summary of the invention

[0006] The present invention aims to propose an automatic detection and quantitative analysis method for bridge defects based on point cloud data processing. This method effectively solves the problems of low efficiency, high cost and easy omission of defects faced by traditional bridge defect detection methods that rely on manual inspections through comprehensive, efficient, high-precision and automated detection means.

[0007] To achieve the above-mentioned purpose, the technical solution adopted by the present invention is: a method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing, comprising the following steps: Step 1: Point cloud data collection: conduct on-site surveys of the target bridge and deploy laser scanners to perform all-round, high-precision 3D scanning to systematically obtain detailed point cloud data of the bridge; Step 2: Data preprocessing: preprocessing the collected 3D point cloud data, including filtering to remove noise, downsampling to improve computing efficiency, normal estimation to clarify the surface direction of the point cloud, and registration to integrate multi-view scanning data to ensure the integrity and accuracy of the point cloud data; Step 3: Bridge component segmentation: segment the pre-processed complete bridge point cloud into multiple independent and simpler component units according to the various components of the bridge, thereby reducing the computational complexity of the subsequent algorithm and improving the processing speed; Step 4: Defect detection: Use the region growing clustering segmentation algorithm to detect defects in each component unit to achieve automatic identification of bridge defects; Step 5: Defect reconstruction: Use Kriging interpolation method to automatically reconstruct the identified defective parts, fill in the missing information, and restore their complete geometric shape so that they can be better used for quantitative analysis of bridge defects.

[0008] Step 6: Quantitative defect analysis: Calculate the volume and surface area of ​​the reconstructed defect holes based on the Delaunay method to provide data support for subsequent evaluation of the overall safety of the bridge and making bridge maintenance decisions.

[0009] The present invention is further improved in that: the step 1 specifically includes: step 1.1: field investigation and preparation, a detailed field investigation of the target bridge is carried out to fully understand the structural characteristics of the bridge, the existing damage status and its surrounding environment. The purpose of this stage is to clarify the specific scope of the scan and the required accuracy. According to the specific conditions of the bridge, a suitable laser scanner platform is built to ensure that the scanning equipment can stably and accurately capture the three-dimensional information of the bridge under different environmental conditions; step 1.2: equipment configuration and inspection, after the laser scanner is built, the necessary parameter configuration and system inspection are carried out, and the scanning speed, resolution and scanning of the scanner are adjusted according to the size of the bridge and the detailed scanning requirements. Angle to ensure that detailed and accurate point cloud data can be collected. The adjustment of these parameters must take into account the subsequent processing requirements and intended use of the data to ensure the practicality and operability of the data; Step 1.3: Implement 3D scanning, start the laser scanner, and perform a full-scale 3D scan of the target bridge. During the scanning process, pay attention to avoid possible obstructions and interference, and ensure that the lighting conditions in the scanning area are sufficient and uniform. This is the key to obtaining high-quality point cloud data; for the specific structural features of the bridge, a multi-angle and multi-station scanning strategy is adopted, which can not only enhance the integrity of the data, but also capture detailed information of the bridge from different perspectives, providing comprehensive data support for subsequent defect detection and analysis.

[0010] The further improvement of the present invention is that the step 2 specifically includes: step 2.1: statistical filtering, first, a statistical filtering algorithm is used to filter out noise points generated by scanning equipment and environmental interference. The method is based on the principle of statistical distribution, and the average distance between each point and its nearest K neighboring points is calculated, and it is assumed that these distances follow a Gaussian distribution; by setting a distance threshold calculated based on the mean and variance of these distances, points exceeding this threshold are regarded as abnormal outliers and removed, so as to avoid the impact of excessive data on subsequent processing efficiency and accuracy; step 2.2: voxel grid downsampling, and then the voxel grid downsampling algorithm is used to simplify the redundant points in the point cloud data, while retaining the spatial geometry of the point cloud model. Features, by dividing the filtered point cloud data into cubic units of uniform size (i.e., voxels), selecting a point inside each voxel as the representative of all the points inside it, and the representative point is an existing point in the voxel, the center point or the point closest to the center point. After voxel grid downsampling, the data density can be significantly reduced while retaining the morphological characteristics of the original surface; Step 2.3: Normal estimation, using the K-nearest-neighbors (KNN) search algorithm, based on the spatial position of a single point in the point cloud and the distribution characteristics of its neighboring points, identify the K nearest neighbors of each point, and use the spatial distribution of these neighboring points to estimate the normal direction; Step 2.4: Point cloud registration uses the fast point feature histogram algorithm combined with the random sampling consensus algorithm for global registration. On the basis of global registration, the iterative closest point algorithm is used for local registration to accurately align the point cloud data collected from different perspectives and at different times to build a complete three-dimensional model; Step 2.4.1: Use the fast point feature histogram (FPFH) algorithm to capture the geometric features of the point cloud data. FPFH is a method for describing the local geometric features of point cloud data. It analyzes the neighborhood information of each point and calculates its histogram descriptor to capture the detailed geometric features of the point cloud data; Step 2.4.2: Use the random sampling consensus (RANSAC) algorithm for global registration. The RANSAC algorithm randomly extracts sample points from the point cloud to preliminarily estimate the transformation matrix, and then determines the optimal global alignment parameters through an iterative optimization and verification process.In the global registration stage, the FPFH feature not only establishes the correspondence between point clouds and enhances the robustness of the registration, but also significantly improves the efficiency of the RANSAC algorithm, thereby ensuring the stability and efficiency of the global registration process; Step 2.4.3: On the basis of global registration, the iterative closest point (ICP) algorithm optimized by the Gauss-Newton method is used for refined local registration; the ICP algorithm continuously optimizes the correspondence between point clouds in an iterative manner and solves the rotation and translation parameters so that the overlapping parts of the two groups of point clouds in different coordinate systems are completely consistent after rotation and translation in the selected global coordinate system; in the local registration stage, the FPFH feature still plays an important role, helping the ICP algorithm to find the corresponding points more accurately to achieve more precise alignment. After the registration is completed, there are still noise, outliers and other abnormal values ​​in the point cloud data. At this time, the filtering operation of the aforementioned step 2.1 is performed again to further eliminate such noise and abnormal values, so as to make the point cloud data smoother and more accurate.

[0011] A further improvement of the present invention is that the step three specifically includes: step 3.1: bridge component segmentation, using the RANSAC algorithm to perform component segmentation on the registered complete bridge point cloud, and segmenting the various components of the bridge, such as the bridge deck, bridge tower, bridge piers, guardrails, etc., to reduce the computational complexity of the subsequent algorithm and improve the processing speed; although the entire bridge structure is very complex, the various bridge components have clear geometric shapes (such as columns, planes, curved surfaces, etc.), which can be accurately described by parameterized mathematical models and the size differences between different components are obvious (such as the height of the pier is much larger than the thickness of the railing), which helps to distinguish different components in the point cloud data. The RANSAC algorithm makes full use of the clear geometric shapes, local structural regularity, scale difference information of the bridge components and the inherent anti-noise ability of the algorithm, and iteratively searches for the model parameters that best represent the characteristics of each component to accurately segment each independent component unit in the complex bridge point cloud data.

[0012] The further improvement of the present invention is that the step 4 specifically includes: step 4.1: in view of the fact that the amount of point cloud data is usually very large, the efficiency of directly using the traversal method to perform neighborhood search is low, so the strategy of constructing a KD tree is adopted in this method to achieve efficient neighborhood search. KD tree is a data structure that can effectively organize data in multidimensional space to accelerate query and update operations. By using KD tree, the efficiency of geometric feature calculation of related points can be significantly improved; step 4.2: in this step, key features including curvature, neighborhood density, reflection energy value and neighborhood variance are extracted for each point of the point cloud data. The curvature is obtained by analyzing the change of the normal of the point, the neighborhood density is determined by counting the number of neighboring points within a certain radius, and the reflection energy and neighborhood variance are calculated based on the reflection energy value of the neighboring points. These features are crucial for distinguishing defective parts from non-defective parts; step 4.3: in the last step of defect detection, the regional growing clustering segmentation algorithm is used to perform a detailed analysis of each point cloud bridge component unit. The process starts from the point with the minimum curvature, and is classified based on key features such as curvature, normal vector, neighborhood density, reflection energy value, and domain variance, and gradually expands to the surrounding area, continuously increasing the target area for segmentation. This process continues until all point clouds are covered, and similar areas are classified into the same cluster area according to the merging conditions. After the classification is completed, the defect distribution of the bridge is intuitively displayed by marking defect points and non-defect points with different colors, thereby completing the defect detection process of point cloud data.

[0013] A further improvement of the present invention is that the step five specifically includes: Step 5.1: Using the Kriging interpolation algorithm, according to the spatial distribution and attribute characteristics of the existing points in the point cloud data, calculate the statistics such as the weight and variance of each point, accurately estimate the point cloud data of the defective part, and generate a continuous and smooth interpolation result. Through Kriging interpolation, the defective area can be automatically repaired and reconstructed, and the missing data can be filled, so that the repaired point cloud data is highly consistent with the actual situation. This process provides a reliable data basis for the subsequent quantitative analysis of bridge defects, helps to more accurately evaluate the safety of the bridge structure, and provides strong data support for the formulation of targeted maintenance strategies.

[0014] A further improvement of the present invention is that the step six specifically includes: step 6.1: after the point cloud defect reconstruction is completed, the original defective part and the reconstructed part form a closed hole, and the surface area and volume of the closed hole are calculated using the Delaunay triangulation and Delaunay tetrahedron partitioning algorithms respectively; the Delaunay triangulation algorithm constructs these discrete three-dimensional space points into a continuous, seamless and non-overlapping triangular grid system based on the spatial distribution characteristics of the reconstructed point cloud; the grid system closely fits the original data form, and its constituent unit, the triangle, tends to be equilateral as much as possible while maintaining structural stability, so as to minimize errors and improve computing efficiency through the optimized triangulation strategy, and delicately reproduce the fine features of the defective area; by accumulating each The area of ​​the non-overlapping triangular mesh can accurately calculate the surface area of ​​the closed hole; the Delaunay tetrahedron partitioning algorithm converts the point cloud data of the closed hole into a grid structure composed of several tetrahedrons, and all tetrahedrons together constitute a non-overlapping tetrahedral mesh covering the point cloud; the nearest neighbor of any point in each tetrahedron is still one of the vertices of the tetrahedron and for any tetrahedron, its circumscribed sphere (the minimum enclosing sphere with the four vertices of the tetrahedron as the center) does not contain any data points; it is convenient to calculate the volume of each tetrahedron based on the constructed Delaunay tetrahedral mesh, and then add up the volumes of all tetrahedrons to obtain the volume of the closed hole; the volume and surface area of ​​the reconstructed defective hole calculated based on the Delaunay method provide data support for the subsequent evaluation of the overall safety of the bridge and the formulation of bridge maintenance decisions.

[0015] (1) This method can quickly obtain point cloud data of bridges through data collection using a 3D laser scanner. This greatly reduces the risk of inspectors working in high-risk environments, reduces interference from human factors, and improves work efficiency. Remote operation reduces the reliance on the personal experience and technical level of inspectors, making the data collection process more objective and reliable.

[0016] (2) The point cloud data-based processing method adopted in the present invention can comprehensively and meticulously detect bridge defects and realize the automation of defect detection. This automation process not only improves the reliability and accuracy of detection, but also ensures comprehensive coverage of the detection area, reducing the risk of missing or misjudging potential problems. This method can not only identify the defective parts of the bridge, but also perform quantitative analysis on these defects. This analysis method can more intuitively and accurately evaluate the degree of harm of defects to the safety of the bridge structure, providing strong data support for bridge maintenance and reinforcement decisions.

[0017] (3) Compared with traditional bridge inspection methods, the present invention significantly reduces the input of manpower and material resources, reduces the overall inspection cost, and reduces the waste of resources and duplication of work caused by improper human operation through automated and precise inspection processes, thereby improving the efficiency of resource utilization. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a flow chart of the automatic detection and quantitative analysis method of bridge defects based on point cloud data processing of the present invention. DETAILED DESCRIPTION

[0019] In this embodiment, see Figure 1 As shown, the present invention proposes an automatic detection and quantitative analysis method for bridge defects based on point cloud data processing, including point cloud data acquisition, data preprocessing, bridge component segmentation, defect detection, defect reconstruction, and defect quantitative analysis.

[0020] Step 1: Point cloud data collection: conduct on-site surveys of the target bridge and deploy laser scanners to perform all-round, high-precision 3D scanning to systematically obtain detailed point cloud data of the bridge.

[0021] Step 2: Data preprocessing: preprocessing the collected 3D point cloud data, including filtering to remove noise, downsampling to improve computing efficiency, normal estimation to clarify the surface direction of the point cloud, and registration to integrate multi-view scanning data to ensure the integrity and accuracy of the point cloud data.

[0022] Step 3: Bridge component segmentation: segment the preprocessed complete bridge point cloud into multiple independent and simpler component units according to the various components of the bridge, thereby reducing the computational complexity of subsequent algorithms and improving processing speed.

[0023] Step 4: Defect detection: Use the region growing clustering segmentation algorithm to perform defect detection on each component unit to achieve automatic identification of bridge defects.

[0024] Step 5: Defect reconstruction: Use Kriging interpolation method to automatically reconstruct the identified defective parts, fill in the missing information, and restore their complete geometric shape so that they can be better used for quantitative analysis of bridge defects.

[0025] Step 6: Quantitative defect analysis: Calculate the volume and surface area of ​​the reconstructed defect holes based on the Delaunay method to provide data support for subsequent evaluation of the overall safety of the bridge and making bridge maintenance decisions.

[0026] The step 1 specifically includes: Step 1.1: Field investigation and preparation, a detailed field investigation of the target bridge is conducted to fully understand the structural characteristics of the bridge, the existing damage status and the surrounding environment. The purpose of this stage is to clarify the specific scope and required accuracy of the scan, and to build a suitable laser scanner platform according to the specific conditions of the bridge to ensure that the scanning equipment can stably and accurately capture the three-dimensional information of the bridge under different environmental conditions; Step 1.2: Equipment configuration and inspection. After the laser scanner is set up, perform necessary parameter configuration and system inspection. Adjust the scanning speed, resolution and scanning angle of the scanner according to the size of the bridge and the detailed scanning requirements to ensure that detailed and accurate point cloud data can be collected. The adjustment of these parameters must take into account the subsequent processing requirements and the intended use of the data to ensure the practicality and operability of the data. Step 1.3: Implement 3D scanning, start the laser scanner, and perform a full-scale 3D scan of the target bridge. During the scanning process, pay attention to avoid possible obstructions and interference, and ensure that the lighting conditions in the scanning area are sufficient and uniform. This is the key to obtaining high-quality point cloud data. In view of the specific structural characteristics of the bridge, a multi-angle and multi-station scanning strategy is adopted, which can not only enhance the integrity of the data, but also capture detailed information of the bridge from different perspectives, providing comprehensive data support for subsequent defect detection and analysis. In order to achieve efficient and accurate point cloud data acquisition, the fastest and most accurate FARO Focus Premium 3D laser scanner on the market is selected in this embodiment. The device can capture up to 2 million data points per second, ensuring large-scale and high-density bridge surface scanning in a short time. The ranging error is controlled within ±1 mm, and the angle measurement accuracy reaches ±19 arc seconds, ensuring that the collected data is highly accurate and avoiding problems such as point cloud distortion caused by measurement errors.

[0027] The step 2 specifically includes: step 2.1: statistical filtering. First, a statistical filtering algorithm is used to filter out noise points generated by scanning equipment and environmental interference. This method is based on the principle of statistical distribution. The average distance between each point and its nearest K neighboring points is calculated, and it is assumed that these distances follow a Gaussian distribution. By setting a distance threshold calculated based on the mean and variance of these distances, points exceeding this threshold are regarded as abnormal outliers and removed to avoid excessive data affecting subsequent processing efficiency and accuracy. The distance threshold is obtained by the following formula:

[0028] is the average distance from a point to all points, that is:

[0029] Represents the standard deviation of the distance, that is:

[0030] Represents a custom standard deviation coefficient, which is used to control the impact of the distance standard deviation on the distance threshold; Step 2.2: Voxel grid downsampling. Then, the voxel grid downsampling algorithm is used to simplify the redundant points in the point cloud data while retaining the spatial geometric features of the point cloud model. The filtered point cloud data is divided into cubic units (i.e., voxels) of uniform size. A point is selected inside each voxel as the representative of all points inside it, and the representative point is an existing point in the voxel, the center point, or the point closest to the center point. After voxel grid downsampling, the data density can be significantly reduced while retaining the morphological features of the original surface. In each voxel, an existing point is preferentially selected as the representative point rather than a virtual point obtained by calculation. When the center point is an existing point, the center point is selected as the representative point. When the center point is not an existing point but is calculated by averaging the voxel center coordinates, the existing point closest to the center point in the voxel is selected as the representative point. To calculate the nearest point, the point with the smallest distance from the center point among all the points in the voxel must be found. is the desired representative point, is the set of all known points in the current voxel, c is the coordinate of the center point of the voxel, and the specific formula is: =

[0031] Step 2.3: Normal estimation. The K-Nearest-Neighbors (KNN) search algorithm is used to identify the K nearest neighbors of each point based on the spatial position of a single point in the point cloud and the distribution characteristics of its neighboring points, and the normal direction is estimated using the spatial distribution of these neighboring points. The KD tree method is used for KNN search to improve the efficiency of geometric feature calculation of related points to find the K nearest neighbor points of each point. Then the covariance matrix of each point is calculated to obtain the eigenvalue and eigenvector. The eigenvector corresponding to the minimum eigenvalue of the eigenvector is the normal direction of the point. The point set N contains the K nearest neighbors of a point p. The average position μ and covariance matrix Σ of these points are calculated:

[0032] K represents the number of nearest neighbor points, that is, the number of neighbor points used to estimate the normal; N represents the set of K nearest neighbor points of a point p; x i Represents the coordinate vector of the i-th neighbor point in the set N; Step 2.4: Use the fast point feature histogram algorithm combined with the random sampling consensus algorithm for global registration. On the basis of global registration, use the iterative closest point algorithm for local registration to accurately register the point cloud data collected from different perspectives and at different times to build a complete 3D model; Step 2.4.1: First, use the fast point feature histogram (FPFH, Fast Point Feature Histograms) algorithm captures the geometric features of point cloud data. FPFH is a method for describing the local geometric features of point cloud data. By analyzing the neighborhood information of each point, its histogram descriptor is calculated to capture the detailed geometric features of point cloud data. The calculation process of FPFH descriptor is as follows: First, the FPFH descriptor is calculated based on the normal direction of each point in the point cloud estimated in step 2.4. For each point, the relative position (distance and angle) of each point in its neighborhood is calculated, and this information is combined into a histogram. Finally, in order to make the descriptor invariant and comparable, the FPFH descriptor is normalized to have unit length. Among them, FPFH is an algorithm improved by point feature histogram (PFH, Point feature Histograms), which retains the main geometric characteristics of point description in PFH and reduces the computational complexity from O(nk 2 ) is reduced to O(nk), where n is the number of points in the point cloud data, and k is the number of points contained in the neighborhood of each point; Step 2.4.2: Use the Random Sample Consensus (RANSAC) algorithm for global alignment. The RANSAC algorithm randomly extracts sample points from the point cloud to preliminarily estimate the transformation matrix, and then determines the optimal global alignment parameters through an iterative optimization and verification process; In the global alignment stage, the FPFH feature not only establishes the correspondence between point clouds and enhances the robustness of the alignment, but also significantly improves the efficiency of the RANSAC algorithm, thereby ensuring the stability and efficiency of the global alignment process.

[0033] First, a group of data points (called inliers) are randomly selected from the dataset, and the model parameters are estimated based on the group of points; the selected data points are used to fit the model to the transformation matrix representing rotation and translation in point cloud registration; all data points are compared with the fitted model, and points whose fitting error with the model is less than a threshold are marked as inliers; metrics of model adaptation to the data, such as the number of inliers or fitting error, are calculated to evaluate the model; the above steps are iterated until the stopping condition is met; finally, the model with the largest number of inliers or the smallest fitting error is selected as the optimal model to achieve the best global registration.

[0034] Step 2.4.3: Based on the global registration, the iterative closest point (ICP) algorithm optimized by the Gauss-Newton method is used for refined local registration. The ICP algorithm continuously optimizes the correspondence between point clouds in an iterative manner and solves the rotation and translation parameters so that the overlapping parts of the two groups of point clouds in different coordinate systems are completely consistent after rotation and translation in the selected global coordinate system. In the local registration stage, the FPFH feature still plays an important role, helping the ICP algorithm to find the corresponding points more accurately to achieve more precise alignment.

[0035] Among them, two groups of point clouds are selected in the overlapping area of ​​the two groups of point cloud data after rough registration, P s and P t ; The points inside the two correspond one to one, and there are N p For the corresponding points, when there is no error, the rotation R and translation T are calculated, and P s Convert to P t , the conversion formula is:

[0036] However, due to the existence of noise, the above formula may not be valid for all points, so there are errors. The core of the ICP algorithm is to minimize an objective function, namely the error.

[0037] The source point cloud is transformed with the initial rotation and translation matrix to obtain a transformed point cloud, and then the transformed point cloud is compared with the target point cloud to determine whether it is the nearest point pair; after determining the corresponding points, the corresponding point pairs are used to estimate the rotation R and translation T to solve the optimal rotation and translation matrix; the two steps of finding corresponding points and optimizing R and T are iterated continuously until the neighboring point pairs no longer change; after completing the alignment, there are still noise, outliers and other abnormal values ​​in the point cloud data. At this time, the filtering operation of the aforementioned step 2.1 is performed again to further eliminate such noise and abnormal values, thereby making the point cloud data smoother and more accurate.

[0038] The step three specifically includes: step 3.1: bridge component segmentation, using the RANSAC algorithm to segment the complete bridge point cloud after registration, and segmenting the various components of the bridge, such as the bridge deck, bridge towers, bridge piers, guardrails, etc., to reduce the computational complexity of the subsequent algorithm and improve the processing speed; although the entire bridge structure is very complex, the various bridge components have clear geometric shapes (such as columns, planes, curved surfaces, etc.), which can be accurately described by parameterized mathematical models, and the size differences between different components are obvious (such as the height of the pier is much greater than the thickness of the railing), which helps to distinguish different components in the point cloud data; the RANSAC algorithm makes full use of the clear geometric shapes, local structural regularity, scale difference information of the bridge components and the inherent anti-noise ability of the algorithm, and iteratively searches for the model parameters that best represent the characteristics of each component to accurately segment each independent component unit in the complex bridge point cloud data.

[0039] Among them, a part of points are randomly selected from the aligned point cloud data as sample points, and various geometric models are used for fitting based on the selected sample points, including plane, cuboid, cylinder, etc., to estimate the various parameters of the model, such as the normal vector and intercept of the plane, the length, width and height of the cuboid, the axis, radius and height of the cylinder, etc.; then, the remaining point cloud data that has not been selected as samples is used to verify the accuracy of the obtained model, and by calculating the deviation and distance between the sample points and the model, it is determined which points belong to the model's internal points (i.e., points that match the model) and which points belong to the external points (i.e., points that do not match the model). This process will continue to iterate, and each iteration will update the current best model parameters until the preset termination condition is reached (the maximum number of iterations or the internal point ratio threshold is reached), and the model parameters with the largest number of internal points are considered to be optimal.

[0040] The specific steps of the RANSAC algorithm are as follows: (1) Random sampling and model estimation: For a data set S, randomly sample s data points to obtain a sample data set, and use these s data points to determine the model M. i , where s is the minimum number of points required to determine a model; (2) Inlier point identification and consistent data set construction: For all other points p in the data set S except the sample point, calculate the distance from the point to the model M i If the distance or residual is less than the threshold t, point p is added to the set S. i In the model, marked as model M i The interior points of the set S i Model M i A consistent dataset; (3) Model evaluation and iterative update: For the set Si, the number of points in the set is ,Will Compare with one of the following termination conditions, including: Reach the predetermined minimum internal point number threshold T min , the preset maximum number of iterations N has been executed max , the number of inliers in the current model exceeds the maximum number of inliers in all previous models, and the proportion of inliers reaches the threshold; if the termination condition is met, use S i Re-estimate the model parameters for all the internal points in to obtain a more accurate model , and output the result. If the termination condition is not met, return to step (1) and continue to the next round of iteration.

[0041] (4) Optimal model selection: After completing the preset N iterations (the actual number of iterations may be less than N because the termination condition may be reached early), compare the various consistent data sets Si generated during all iterations; select the model corresponding to the data set with the largest number of inliers or the highest proportion of inliers as the optimal model; use all the points in the consistent data set corresponding to the optimal model to re-estimate the model parameters and obtain the final model.

[0042] The step four specifically includes: Step 4.1: In view of the fact that the amount of point cloud data is usually very large, the efficiency of directly using traversal to perform neighborhood search is low. Therefore, the method adopts the strategy of constructing a KD tree to achieve efficient neighborhood search. The KD tree is a data structure that can effectively organize data in a multi-dimensional space to accelerate query and update operations. By using the KD tree, the efficiency of calculating the geometric features of related points can be significantly improved; Step 4.2: In this step, key features including curvature, neighborhood density, reflection energy value and neighborhood variance are extracted for each point of the point cloud data; the curvature is obtained by analyzing the change of the normal of the point, the neighborhood density is determined by counting the number of neighboring points within a certain radius, and the reflection energy and neighborhood variance are calculated based on the reflection energy values ​​of the neighboring points. These features are crucial for distinguishing defective parts from non-defective parts; Step 4.3: In the final step of defect detection, the regional growing clustering segmentation algorithm is used to perform a detailed analysis of each point cloud bridge component unit; the process starts from the point with the smallest curvature, and is classified based on key features such as curvature, normal vector, neighborhood density, reflection energy value and domain variance, and gradually expands to the surrounding area, continuously increasing the target area for segmentation. This process continues until all point clouds are covered. According to the merging conditions, similar areas are classified into the same clustering area. After the classification is completed, the defect points and non-defect points are marked with different colors to intuitively display the defect distribution of the bridge, thereby completing the defect detection process of the point cloud data.

[0043] Among them, the specific process of the region growing clustering segmentation algorithm is as follows: Set initial conditions: define an empty clustering region M and an empty seed point sequence N.

[0044] Select the initial seed point: Select the point C with the smallest curvature from the sorted point cloud as the initial seed point and add it to the seed point sequence N.

[0045] Set parameters: Set the normal vector angle threshold θ and the curvature threshold K. These two thresholds are used to determine whether a point should be included in the current clustering area.

[0046] Region growing: For each neighboring point P of the seed point C, calculate the angle between its normal vector and the normal vector of C. If the angle is less than the set threshold θ, it indicates that P and C have a high consistency in direction, and point P is included in the clustering region M. In order to find points that may belong to the same region but have not yet been included, traverse the neighboring points of the current clustering region M, and for points whose curvature is less than the set threshold K, add them to the seed point sequence N as potential new seed points.

[0047] Iterative process: After completing a round of growth, remove the processed seed point C from the seed point sequence N. If there are still remaining seed points in N, select the next seed point and repeat step (4) for region growing. Otherwise, proceed to the next step.

[0048] Termination condition: When the seed point sequence N is empty, that is, all available seed points have been processed and no new points are added to N, it means that the clustering process is completed and all points with similar features, such as similar curvature and normal vector direction, have been classified into the corresponding clustering area M.

[0049] Visualization of defect detection results: The detected defect points and non-defect points are marked with different colors to intuitively present the defect distribution of the point cloud and complete the point cloud defect detection process.

[0050] The step five specifically includes: Step 5.1: Use the Kriging interpolation algorithm to calculate the weight and variance of each point based on the spatial distribution and attribute characteristics of the existing points in the point cloud data, accurately estimate the point cloud data of the defective part, and generate a continuous and smooth interpolation result; Kriging interpolation can automatically repair and reconstruct the defective area, fill in the missing data, and make the repaired point cloud data highly consistent with the actual situation; this process provides a reliable data basis for the subsequent quantitative analysis of bridge defects, helps to more accurately evaluate the safety of bridge structures, and provides strong data support for the formulation of targeted maintenance strategies; the core of Kriging interpolation is to estimate the value of unknown points through the data of stored points, so as to realize the reconstruction of defective areas; first, calculate the semivariogram between known points to describe the spatial correlation of data points; the semivariogram includes the distance interval h and the semivariogram value Two parameters are used to quantify the degree of change of attribute values ​​at different distances. The calculation formula of the semivariogram function is as follows:

[0051] represents the semivariance value when the distance interval is h, N(h) represents the number of point pairs with a distance interval of h, and Z(x i ) and Z(x i +h) respectively represent the position x i and x i + The attribute value at h.

[0052] Based on the semivariogram, the weight of each known point in predicting the attribute value of the unknown point is calculated; the size of the weight depends on the distance between the known point and the unknown point and the spatial correlation between them. The weight calculation formula is as follows:

[0053] λ i represents the weight of the i-th known point in predicting the unknown point, γ(h ij ) represents the semivariance value corresponding to the distance interval between the i-th known point and the unknown point j, Represents the sum of the semi-variance values ​​corresponding to the distance intervals between all known points and the current unknown point j.

[0054] Using the attribute values ​​of known points and the corresponding weights, the attribute values ​​of unknown points are calculated by weighted average. The interpolation calculation formula is as follows:

[0055] Z 0 represents the predicted attribute value of unknown point j, λ i represents the weight of the i-th known point, Z i It represents the attribute value of the i-th known point, and outputs the interpolation result as new point cloud data to realize the reconstruction of the defect area.

[0056] The step six specifically includes: Step 6.1: After the point cloud defect reconstruction is completed, the original defective part and the reconstructed part form a closed hole, and the surface area and volume of the closed hole are calculated using the Delaunay triangulation and Delaunay tetrahedron partitioning algorithms respectively; the Delaunay triangulation algorithm constructs these discrete three-dimensional space points into a continuous, seamless and non-overlapping triangular grid system based on the spatial distribution characteristics of the reconstructed point cloud; the grid system closely fits the original data form, and its constituent unit, the triangle, tends to be equilateral as much as possible while maintaining structural stability, so as to minimize errors and improve computing efficiency through the optimized triangulation strategy, and delicately reproduce the fine features of the defective area. By accumulating the area of ​​each non-overlapping triangular grid, the surface area of ​​the closed hole can be accurately calculated. The Delaunay tetrahedral meshing algorithm converts the point cloud data of the closed hole into a grid structure composed of several tetrahedrons. All tetrahedrons together form a non-overlapping tetrahedral mesh covering the point cloud. The nearest neighbor of any point in each tetrahedron is still one of the vertices of the tetrahedron, and for any tetrahedron, its circumscribed sphere (the minimum enclosing sphere with the four vertices of the tetrahedron as the center) does not contain any data points. For the constructed Delaunay tetrahedral mesh, the volume of each tetrahedron is calculated, and then the volumes of all tetrahedrons are accumulated to obtain the volume of the closed hole. The Delaunay-based method calculates the volume and surface area of ​​the reconstructed defective hole to provide data support for the subsequent evaluation of the overall safety of the bridge and the formulation of bridge maintenance decisions.

[0057] Among them, the key to the Delaunay triangulation algorithm is to satisfy the empty circle property and the maximum and minimum angle principle. For any triangle in the mesh, its circumscribed circle does not contain any other points in the point cloud set, ensuring good separation between triangles, avoiding overlap, and ensuring the connectivity of the mesh. Under the premise of satisfying the empty circle property, Delaunay triangulation strives to make the minimum internal angle of all triangles as large as possible, avoiding the generation of overly narrow and long triangles, and making the shape of each triangle as uniform as possible and close to an equilateral triangle. The generated triangular mesh has uniformity and good geometric properties. If the triangle is defined by vertices A, B, and C, the formulas for calculating the area of ​​a single triangle and the total area are as follows:

[0058] “ " represents the cross product of the vectors, "||" represents the norm (modulus) of the vectors, N is the number of triangles in the triangular mesh, is the area of ​​the i-th triangle.

[0059] The key to the Delaunay tetrahedron partitioning algorithm is that it can uniquely determine a Delaunay tetrahedron partitioning structure for any given point cloud data set, ensuring that the tetrahedron mesh generated by the algorithm is not only adaptive to the point cloud distribution, but also has extremely high stability, which is crucial for volume calculations that rely on accurate and consistent mesh structures. If the tetrahedron is defined by four vertices A, B, C, and D, the formulas for calculating the volume of a single tetrahedron and the total volume are as follows: =

[0060]

[0061] "·" represents the vector dot product, "×" represents the vector cross product, "||" represents the norm (modulus) of the vector, N is the number of tetrahedrons in the tetrahedral mesh, is the volume of the i-th tetrahedron.

[0062] Based on the above content, we can subsequently combine multi-dimensional information such as the specific location and category of the defects to conduct a comprehensive and in-depth quantitative assessment of potential risks such as safety hazards, service performance degradation, and maintenance needs that may be caused by these defects in the bridge.

[0063] Those skilled in the art should understand that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, and these changes and improvements fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing, characterized in that: The following steps are involved: Step 1: collect point cloud data. For the target bridge, conduct on-site surveys and deploy 3D laser scanners to complete the 3D scanning of the bridge in an all-round and high-precision manner. Step 2: data preprocessing: preprocess the collected 3D point cloud data. This stage includes filtering to remove noise, downsampling to improve computing efficiency, normal estimation to clarify the surface direction of the point cloud, and alignment and integration of multi-view scanning data. Step 3: bridge component segmentation: divide the integrated bridge point cloud data into independent component units. Step 4: defect detection: apply the regional growing clustering segmentation algorithm to the segmented component units to divide them into defective parts and non-defective parts. Step 5: defect reconstruction: use the Kriging interpolation method to intelligently repair and reconstruct the identified defective parts to fill in the missing information. Step 6: defect quantitative analysis: calculate the volume and surface area of ​​the defect holes based on the Delaunay method.

2. The method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing according to claim 1 is characterized by: Step 1 is specifically divided into the following steps: Step 1.1: Field investigation and preparation. Conduct a detailed field investigation of the target bridge to fully understand the structural characteristics of the bridge, the existing damage status and its surrounding environment; Step 1.2: Equipment configuration and inspection. After the laser scanner is set up, perform parameter configuration and system inspection; adjust the scanner's scanning speed, resolution and scanning angle according to the size of the bridge and detailed scanning requirements; Step 1.3: Implement 3D scanning. Start the laser scanner to perform a full-scale 3D scan of the target bridge.

3. The method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing according to claim 1 is characterized by: Step 2 is specifically divided into the following steps: Step 2.1: Statistical filtering, using a statistical filtering algorithm to filter out noise points generated by scanning equipment and environmental interference; Step 2.2: Voxel grid downsampling, then using a voxel grid downsampling algorithm to streamline redundant points in the point cloud data while retaining the spatial geometric features of the point cloud model; Step 2.3: Normal estimation, using the K nearest neighbor search algorithm, based on the spatial position of a single point in the point cloud and the distribution characteristics of its neighboring points, identify the K nearest neighbors of each point, and use the spatial distribution of these neighboring points to estimate the normal direction; Step 2.4: Point cloud registration, using the fast point feature histogram algorithm combined with the random sampling consensus algorithm for global registration, and using the iterative closest point algorithm for local registration on the basis of global registration, accurately align the point cloud data collected from different perspectives and at different times to build a complete three-dimensional model; Step 2.4.1: Use the fast point feature histogram algorithm to capture the geometric features of the point cloud data; Step 2.4.2: Use the random sampling consensus algorithm for global registration; Step 2.4.3: On the basis of global registration, use the iterative closest point algorithm optimized by the Gauss-Newton method for refined local registration; Perform the filtering operation in step 2.1 above again to further eliminate such noise and outliers.

4. The method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing according to claim 1 is characterized by: Step three is specifically divided into the following steps: Step 3.1: Bridge component segmentation, using the RANSAC algorithm to segment the complete bridge point cloud after alignment, and segment the various components of the bridge.

5. The method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing according to claim 1 is characterized by: The step four specifically includes: step 4.1: adopting the strategy of constructing a KD tree to achieve efficient neighborhood search; step 4.2: in this step, extracting key features including curvature, neighborhood density, reflection energy value and neighborhood variance for each point of the point cloud data; step 4.3: in the last step of defect detection, using the regional growing clustering segmentation algorithm to perform a detailed analysis of each point cloud bridge component unit.

6. The method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing according to claim 1 is characterized by: The step five specifically includes: Step 5.1: Using the Kriging interpolation algorithm, according to the spatial distribution and attribute characteristics of the existing points in the point cloud data, calculate the statistics such as the weight and variance of each point, accurately estimate the point cloud data of the defective part, and generate a continuous and smooth interpolation result.

7. The method for automatic detection and quantitative analysis of bridge defects based on point cloud data processing according to claim 1 is characterized by: The step six specifically includes: Step 6.1: After the point cloud defect reconstruction is completed, the original defective part and the reconstructed part form a closed hole, and the surface area and volume of the closed hole are calculated using Delaunay triangulation and Delaunay tetrahedron partitioning algorithms respectively.

Images