Three-dimensional point cloud preprocessing method, system and medium for bridge emergency detection and evaluation

Abstract

The application relates to the field of three-dimensional laser scanning data processing technology and discloses a three-dimensional point cloud preprocessing method and system for bridge emergency detection and evaluation and a medium, the method comprising the following steps: constructing a noise distribution model corresponding to original point cloud data; calculating adaptive filtering parameters according to the noise distribution model, performing initial filtering on the original point cloud data according to the adaptive filtering parameters, and obtaining final denoising point cloud; generating a point cloud distribution heat map of the final denoising point cloud; performing pseudo-disturbance voxel down-sampling based on density feedback according to the point cloud distribution heat map, obtaining down-sampling point cloud data, and calculating a point cloud quality evaluation index of the down-sampling point cloud data; and performing feedback optimization according to the point cloud quality evaluation index, and obtaining target point cloud data. The application can improve the efficiency and fidelity of bridge three-dimensional point cloud data preprocessing, and realizes fast and accurate original point cloud data preprocessing.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional laser scanning data processing technology, specifically a three-dimensional point cloud preprocessing method, system, and medium for emergency detection and assessment of bridges after disasters. Background Technology

[0002] The scale of highway bridge infrastructure continues to expand, with many bridges located in high-intensity earthquake zones, flood-prone areas, or busy traffic sections. After a disaster, bridge structures may experience partial collapse, bearing displacement, beam cracking, or overall deformation. A rapid assessment of their safety status within the "golden 72 hours" is urgently needed to guide traffic control, evacuation, and priority of repairs. However, a long-standing conflict exists between "speed" and "accuracy" in engineering practice: pursuing processing speed often sacrifices geometric precision, while high-precision processing struggles to meet emergency response timelines. While 3D laser scanning technology can rapidly acquire 3D point cloud data of bridges non-contactly and comprehensively, providing a new tool for post-disaster assessment, the original point clouds generally suffer from severe noise interference, high data redundancy, and the easy loss of key damage features.

[0003] Current mainstream point cloud preprocessing methods struggle to balance efficiency and fidelity. For example, existing research (202211357364.6) discloses a multi-scale filtering method for structural point cloud data that considers the dynamic influence of the environment. This method mainly uses the "K" nearest neighbor method to dynamically filter the original point cloud data. Then, based on the improved principal component analysis algorithm (LMSRPCA), the point cloud data is divided into flat regions and abrupt change regions. Statistical filtering based on local surface fitting is used for flat regions, while spatial adaptive bilateral filtering is used for abrupt change regions. However, this method lacks the ability to identify minute but fatal geometric anomalies such as crack edges, support misalignment, and concrete spalling, often misjudging them as noise and removing them. Paper 202510406588.9 discloses an efficient 3D reconstruction method for component surfaces based on adaptive octree sampling. It adaptively divides octree nodes based on local point cloud density and curvature features, and then performs voxel downsampling or spatial adaptive bilateral filtering on different regions according to the curvature of each node. While fine filtering based on curvature or normal vectors can preserve details, it is computationally complex and time-consuming, failing to meet the needs of rapid post-disaster response. Conventional downsampling methods often pursue uniform point cloud distribution, neglecting structural semantics, and further weakening the geometric representation of key regions while compressing data.

[0004] Point cloud preprocessing, as a crucial link between on-site scanning and intelligent assessment, directly determines the reliability of subsequent damage identification and the scientific nature of emergency decision-making. Improper preprocessing can delay repair efforts or even lead to structural misjudgments and secondary disasters. Therefore, there is an urgent need for a new point cloud preprocessing method that is both fast and high-fidelity, designed for post-disaster emergency scenarios, to truly achieve the data preprocessing goal of being both fast and accurate, simple and precise. Summary of the Invention

[0005] The technical problem to be solved by this invention is how to improve the efficiency and fidelity of 3D point cloud data preprocessing in scenarios such as emergency detection and assessment of bridges after disasters, so as to achieve the goal of fast and accurate data preprocessing.

[0006] The present invention solves the above-mentioned technical problems through the following technical means:

[0007] Acquire raw point cloud data, construct the minimum bounding box corresponding to the raw point cloud data, and construct a noise distribution model of the raw point cloud data based on the minimum bounding box. Based on the noise distribution model, adaptive filtering parameters are calculated, and distance filtering, statistical filtering, and multi-feature fusion iterative filtering are performed on the original point cloud data according to the adaptive filtering parameters to obtain the final denoised point cloud. Calculate the spatial density of the final denoised point cloud, and generate a point cloud distribution heatmap based on the spatial density; Based on the point cloud distribution heatmap, the final denoised point cloud is downsampled using pseudo-perturbation voxels based on density feedback to obtain downsampled point cloud data. Calculate the point cloud quality evaluation index of the downsampled point cloud data, and perform feedback optimization on the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data; The target point cloud data is converted into a standard format file, and a structured report containing processing parameters, operation logs, and quality assessment charts is generated simultaneously.

[0008] Optionally, constructing the minimum bounding box corresponding to the original point cloud data includes: Calculate the data centroid of the original point cloud data, and construct a decentralized coordinate matrix based on the data centroid; Calculate the covariance matrix of the original point cloud data based on the decentralized coordinate matrix; The covariance matrix is ​​decomposed into eigenvectors to obtain eigenvectors, and a rotation matrix is ​​constructed based on the eigenvectors. Calculate the minimum and maximum coordinates of the original point cloud data on the coordinate axes based on the rotation matrix and the data centroid; The minimum bounding box corresponding to the original point cloud data is determined based on the minimum and maximum coordinates.

[0009] Optionally, constructing a noise distribution model of the original point cloud data based on the minimum bounding box includes: The minimum bounding box is divided into multiple voxel intervals along the coordinate axis, and the density histogram of each voxel interval is calculated. Calculate the density histogram using the following formula:

[0010] in, Indicates the first Density histogram of individual element intervals, Indicates the first The total number of original point cloud data in the individual pixel interval. Indicates the first Individual element interval along the first The interval length of each coordinate axis Indicates the first The original point cloud data in the individual pixel interval is perpendicular to The average cross-sectional area of ​​the shaft. These represent the principal axis directions of the horizontal axis, vertical axis, and depth axis, respectively. The density gradient is calculated based on the density histogram, and the noise distribution model is determined based on the density gradient.

[0011] Optionally, the step of performing distance filtering, statistical filtering, and multi-feature fusion iterative filtering on the original point cloud data according to the adaptive filtering parameters to obtain the final denoised point cloud includes: Calculate the Euclidean distance from each of the original point cloud data to the nearest scan origin, and remove the original point cloud data whose Euclidean distance is greater than a preset distance threshold to obtain distance-denoised point cloud; Calculate the average Euclidean distance and standard deviation of each distance-denoised point cloud in its corresponding neighborhood, and calculate the global mean and global standard deviation of the distance-denoised point cloud based on the average Euclidean distance and the standard deviation. If the average Euclidean distance is greater than the sum of the global mean and the global standard deviation which is a multiple of a set threshold, then the distance-denoised point cloud is determined to be an outlier and removed to obtain a statistically filtered point cloud. Based on a preset initial neighborhood size, the local density, local curvature, and average angle of the normal vectors of each statistically filtered point cloud are calculated to construct a weighted composite anomaly score. Points with weighted composite anomaly scores higher than the preset current dynamic threshold are removed to form the current round of filtering results. The initial neighborhood size and the dynamic threshold are then adaptively updated based on the local density of the point cloud in the current round of filtering results. The current filtering result is iteratively denoised based on the updated initial neighborhood size and dynamic threshold to obtain the final denoised point cloud.

[0012] Optionally, the step of calculating the local density, local curvature, and average angle of the normal vectors of each statistically filtered point cloud based on a preset initial neighborhood size, and constructing a weighted composite anomaly score, includes: The local density is obtained by calculating the ratio of the number of neighborhood points to the average distance within the initial neighborhood size of the statistically filtered point cloud. Construct a decentralized coordinate matrix of the neighborhood points and perform eigenvalue decomposition to obtain eigenvalues. Calculate the ratio of the smallest eigenvalue to the sum of all eigenvalues ​​to obtain the local curvature. The unit normal vector is determined based on the eigenvector corresponding to the largest eigenvalue among the eigenvalues, and the average angle between the normal vectors of the statistically filtered point cloud and the neighboring points is calculated based on the unit normal vector. The average angle between the normal vectors of the statistically filtered point cloud and the neighboring point cloud is calculated using the following formula:

[0013] in, Indicates the first Statistical Filtered Point Cloud The average angle between the normal vectors of the points and the neighboring points. Indicates the neighborhood size as The set of time-neighboring points, Indicates the first Neighboring points, Indicates the first The unit normal vector of a statistically filtered point cloud. Indicates the first The unit normal vector of each neighboring point; The local density, local curvature, and average angle of the normal vector are normalized and then summed according to preset weights to obtain a weighted composite anomaly score.

[0014] Optionally, the step of performing density feedback-based pseudo-perturbation voxel downsampling on the final denoised point cloud according to the point cloud distribution heatmap to obtain downsampled point cloud data includes: Construct the variable perturbation domain corresponding to each regular voxel grid in the point cloud distribution heatmap; Calculate the spatial local density of each final denoised point cloud within the variable perturbation domain; Calculate density weights based on the spatial local density of the final denoised point cloud. Representative point cloud data are selected from the variable perturbation domain according to the density weights; Representative point cloud data is selected using the following formula:

[0015] in, This indicates that the data represents point cloud data. Indicates the first The final denoised cloud. Indicates the first The density weights corresponding to the final denoised point cloud. Represents a regular voxel grid. Represents the variable perturbation domain; Based on the representative point cloud data and the point cloud distribution heatmap, the filtered voxel interval is subjected to density-based adaptive downsampling to obtain downsampled point cloud data.

[0016] Optionally, the calculation of the point cloud quality evaluation index of the downsampled point cloud data includes: The downsampled point cloud data is divided into a set of voxels. The normalized standard deviation of the downsampled point cloud data in each voxel within the set of voxels is calculated. The normalized standard deviation is used as a point cloud uniformity index. The point cloud uniformity index is calculated using the following formula:

[0017] in, This represents the point cloud uniformity index. Indicates the first The standard deviation of the number of downsampled point cloud data in individual elements. No. The mean number of downsampled point cloud data in individual pixels. This represents a preset constant; Calculate the relative change in local density of each downsampled point cloud data to obtain a density coefficient change index; The edge point set of the original point cloud data is detected by vector mutation, the nearest neighbor point is found from the downsampled point cloud data, and the structural edge retention rate index is calculated based on the edge point set and the nearest neighbor point. The edge retention rate metric is calculated using the following formula:

[0018] in, This represents the percentage of structural edges retained. This represents a set of downsampled point cloud data. Indicates the first One set of raw point cloud data, Indicates the first The nearest neighbor of each original point cloud data point in the downsampled point cloud data point. Represents the set of edge points; The proportion of noise points in the downsampled point cloud data is calculated using the statistical filtering residual, and the proportion of noise points is used as the noise residual rate index. Calculate the normal vector distribution change index based on the mean and standard deviation of the angle between the normal vectors of the original point cloud data and the downsampled point cloud data; The standard deviation of the point spacing is calculated based on the average distance standard deviation between each of the original point cloud data and each of the downsampled point cloud data and the nearest point cloud. By combining the point cloud uniformity index, the density coefficient change index, the structure edge retention rate index, the noise residual rate index, the normal vector distribution change index, and the point spacing standard deviation change index, the point cloud quality evaluation index of the downsampled point cloud data is obtained.

[0019] Optionally, the step of optimizing the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data includes: A weighted comprehensive score is calculated based on the point cloud uniformity index, density coefficient change index, structure edge retention rate index, and noise residue rate index in the point cloud quality evaluation index. The weighted composite score is calculated using the following formula:

[0020] in, This represents the weighted overall score. , indicating the preset indicator weights, This represents the point cloud uniformity index. Indicators representing changes in density coefficient This represents the percentage of structural edges retained. Indicator representing noise residual rate; A feedback method is generated when the weighted comprehensive score is greater than a preset score threshold or when any of the point cloud quality assessment indicators does not meet the preset indicator range. The adaptive filtering parameters are optimized according to the feedback method to obtain optimized filtering parameters; The downsampled point cloud data is optimized based on the optimized filtering parameters to obtain the target point cloud data.

[0021] To address the aforementioned problems, this invention also proposes a three-dimensional point cloud preprocessing system for bridge emergency detection and assessment, the system comprising: The noise distribution model construction module is used to acquire raw point cloud data, construct the minimum bounding box corresponding to the raw point cloud data, and construct the noise distribution model of the raw point cloud data based on the minimum bounding box. The filtering module is used to calculate adaptive filtering parameters based on the noise distribution model, and perform distance filtering, statistical filtering, and multi-feature fusion iterative filtering on the original point cloud data based on the adaptive filtering parameters to obtain the final denoised point cloud. The point cloud distribution heatmap generation module is used to calculate the spatial density of the final denoised point cloud and generate a point cloud distribution heatmap based on the spatial density. The voxel downsampling module is used to perform pseudo-perturbation voxel downsampling based on density feedback on the final denoised point cloud according to the point cloud distribution heatmap, so as to obtain downsampled point cloud data. The feedback adjustment module is used to calculate the point cloud quality evaluation index of the downsampled point cloud data, and to perform feedback optimization on the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data. The data encapsulation and report generation module is used to convert the target point cloud data into a standard format file and simultaneously generate a structured report containing processing parameters, operation logs, and quality assessment charts.

[0022] The present invention also provides a computer-readable storage medium, characterized in that the computer-readable storage medium stores computer instructions, the computer instructions causing the computer to perform the above-described method for three-dimensional point cloud preprocessing for bridge emergency detection and assessment.

[0023] The advantages of this invention are: This invention constructs a noise distribution model using the minimum bounding box corresponding to the original point cloud data, accurately acquiring the principal scale, orientation, and spatial attitude of the structure, providing guidance for setting adaptive filtering parameters. Based on the adaptive filtering parameters, distance filtering, statistical filtering, and multi-feature fusion iterative filtering are applied to the original point cloud data, enabling the processing of millions of point clouds within 5 minutes with a noise residual rate of less than 2% and an edge preservation rate improved by more than 25%, truly achieving post-disaster "fast yet efficient, simple yet comprehensive" results. The final denoised point cloud undergoes voxel downsampling in a variable perturbation domain, compressing the data volume by 21% while effectively preserving supports, cracks, beam ends, etc. The geometric integrity of high-risk areas is maintained to achieve sparsity of point clouds while preserving the overall geometric structure and features of the point clouds. By constructing a multi-dimensional quality assessment system for downsampled point cloud data, including point cloud uniformity, density variation coefficient, edge retention rate, noise residue rate, normal vector distribution stability, and standard deviation variation of point spacing, the preprocessing effect of the original point cloud data can be objectively quantified, avoiding the risk of misjudgment due to improper preprocessing. In addition, combined with error feedback, the system automatically adjusts the filtering radius, sampling ratio, and density compensation coefficient to improve the efficiency and fidelity of 3D point cloud data preprocessing, achieving fast and accurate preprocessing of original point cloud data. Attached Figure Description

[0024] Figure 1This is a flowchart illustrating a three-dimensional point cloud preprocessing method for bridge emergency detection and assessment in one embodiment of the present invention; Figure 2 This is a schematic diagram of the overall bridge point cloud data obtained by a 3D laser scanner in one embodiment of the present invention; Figure 3 This is based on one embodiment of the present invention. Figure 2 A schematic diagram of the original point cloud data corresponding to the bridge components obtained from the complete bridge shown. Figure 4 This is a schematic diagram of the minimum bounding box of the original point cloud data of a bridge component in one embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the effect of filtering the original point cloud data of bridge components in one embodiment of the present invention. Figure 6 This is a schematic diagram illustrating the effect of downsampling the original point cloud data of a bridge component in one embodiment of the present invention; Figure 7 This is a functional module diagram of a three-dimensional point cloud preprocessing system for bridge emergency detection and assessment provided in one embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Reference Figure 1 The diagram shown is a flowchart illustrating a three-dimensional point cloud preprocessing method for bridge emergency detection and assessment provided in an embodiment of the present invention. In this embodiment, the three-dimensional point cloud preprocessing method for bridge emergency detection and assessment includes: S1. Obtain the original point cloud data, construct the minimum bounding box corresponding to the original point cloud data, and construct the noise distribution model of the original point cloud data based on the minimum bounding box.

[0027] In this embodiment of the invention, the original point cloud data is bridge point cloud data obtained by a 3D laser scanner, used to test the effect of the 3D point cloud data preprocessing of the present invention. The source of the original point cloud data can be, for example,... Figure 2 The overall bridge point cloud shown is as follows. Figure 3 Based on Figure 2 The diagram shows the original point cloud data of the bridge components obtained from the overall bridge.

[0028] In one optional embodiment of the present invention, some basic point cloud information of the original point cloud data can be shown in Table 1.

[0029] Table 1

[0030] The minimum bounding box is the smallest volume geometry that contains all the points of the original point cloud data, used to compactly describe the spatial extent of the point cloud.

[0031] Specifically, constructing the minimum bounding box corresponding to the original point cloud data includes: Calculate the data centroid of the original point cloud data, and construct a decentralized coordinate matrix based on the data centroid; Calculate the covariance matrix of the original point cloud data based on the decentralized coordinate matrix; The covariance matrix is ​​decomposed into eigenvectors to obtain eigenvectors, and a rotation matrix is ​​constructed based on the eigenvectors. Calculate the minimum and maximum coordinates of the original point cloud data on the coordinate axes based on the rotation matrix and the data centroid; The minimum bounding box corresponding to the original point cloud data is determined based on the minimum and maximum coordinates.

[0032] In this embodiment of the invention, a set of original point cloud data is defined. ,in, This represents the total number of original point cloud data. This indicates that each original point cloud data point is represented by three-dimensional spatial coordinates.

[0033] The centroid of the original point cloud data is calculated using the following formula:

[0034] in, Indicates the centroid of the data. This represents the total number of original point cloud data. Indicates the first One set of raw point cloud data.

[0035] Furthermore, with Construct a decentralized coordinate matrix for row data Then, the covariance matrix of the original point cloud data is calculated using the following formula:

[0036] in, Represents the covariance matrix. This represents the total number of original point cloud data. Represents a decentralized coordinate matrix. This indicates transpose.

[0037] For covariance matrix Eigenvalue decomposition can be represented as:

[0038] in, Represents the eigenvalues ​​of the covariance matrix. This represents the eigenvector.

[0039] Based on the feature vector Construct the rotation matrix:

[0040] in, Represents the rotation matrix. Represents the eigenvector. This indicates transpose.

[0041] Furthermore, the original point cloud data is projected onto the principal axis coordinate system of the rotation space based on the rotation matrix. The coordinate range of the minimum bounding box is obtained by finding the minimum and maximum coordinates of the point cloud along each axis (x-axis, y-axis, z-axis) in the principal axis coordinate system, thus yielding the minimum bounding box corresponding to the original point cloud data. Specifically, a schematic diagram of the minimum bounding box is shown below. Figure 4 As shown, the original point cloud of the bridge components is 7,262,047, and the minimum bounding box size is: X: 78.9390 (-63.7397: 15.1993), Y: 111.1480 (-54.9849: 56.1630), Z: 76.8296 (0.0830: 76.9126).

[0042] Specifically, the minimum and maximum coordinates of the original point cloud data on the coordinate axes are calculated using the following formula:

[0043]

[0044]

[0045] in, Indicates in Minimum coordinates on the coordinate axes Represents the rotation matrix. Indicates the centroid of the data. Indicates the first One set of raw point cloud data, Represents a collection of raw point cloud data. Indicates in The maximum coordinate on the coordinate axis Indicates in Minimum coordinates on the coordinate axes Indicates in The maximum coordinate on the coordinate axis Indicates in Minimum coordinates on the coordinate axes Indicates in The maximum coordinate on the coordinate axis.

[0046] In detail, determine the extent and volume of the minimum bounding box in the principal axis space. Represented as:

[0047] In this embodiment of the invention, by aligning the minimum bounding box along the main direction, the spatial fit of the bridge structure is significantly improved, and the main dimensions, orientation, and spatial attitude of the structure are accurately obtained.

[0048] In this embodiment of the invention, the noise distribution model is used to determine the distribution pattern of outliers in the original point cloud data in three-dimensional space.

[0049] Specifically, constructing the noise distribution model of the original point cloud data based on the minimum bounding box includes: The minimum bounding box is divided into multiple voxel intervals along the coordinate axis, and the density histogram of each voxel interval is calculated. The density gradient is calculated based on the density histogram, and the noise distribution model is determined based on the density gradient.

[0050] In this embodiment of the invention, along That is, the three principal axes: the horizontal axis, the vertical axis, and the depth axis. The minimum bounding box is divided into voxel intervals, and the number of points in each interval is counted to obtain a density histogram, as shown in the following formula:

[0051] in, Indicates the first Density histogram of individual element intervals, Indicates the first The total number of original point cloud data in the individual pixel interval. Indicates the first Individual element interval along the first The interval length of each coordinate axis Indicates the first The original point cloud data in the individual pixel interval is perpendicular to The average cross-sectional area of ​​the axis is calculated from the projected area of ​​the original point cloud data.

[0052] Further calculation of the density gradient:

[0053] in, Indicates the first Density gradient in the voxel interval, Indicates the first A preliminary histogram of individual element intervals. Indicates the first Density histogram of individual element intervals.

[0054] Specifically, if a sudden drop in density occurs in the boundary region of the smallest bounding box, i.e. If the gradient threshold is set, it is determined that there is an outlier cluster in the voxel interval, including flying points caused by strong reflection, edge mismatch caused by scanning occlusion, and false acquisition caused by external interference, which provides a basis for subsequent spatial partitioning filtering.

[0055] Furthermore, define the boundary extension region. To extend the safety margin outward from the minimum bounding box is In the spatial domain, the boundary extension region is obtained, and the density of the point cloud within this boundary extension region is calculated. Compared with the density of the main structure (the density of the original point cloud data in the actual occupied area). The ratio:

[0056] like If so, then significant noise is determined to exist. This is a preset empirical threshold, representing the maximum acceptable ratio of the density of the outer region to the density of the main region. It is used to determine whether the point cloud density in the boundary extension region is abnormally high, and is typically set to... If the value exceeds the empirical threshold range, it indicates that the boundary extension region is the second noise region.

[0057] Further classification by spatial direction, Axial direction abnormality: in Low-density clusters of dots appearing outside the boundaries at both ends of the axis, with their elevations differing from the main structure by a greater than a preset threshold, can be identified as elevation error noise, mostly caused by ground undulations, scanner attitude drift, and multiple reflections; horizontal plane Edge coefficient point clusters: distributed outside the horizontal boundary of the smallest bounding box, with low density and spatial dispersion, mostly due to reflection from obstructions and flying points; internal low-density voids: appearing in the internal region of the structure. The missing range may be due to occlusion and needs to be filled in during subsequent modeling.

[0058] In this embodiment of the invention, the distribution of noise in the original point cloud data is determined based on the density gradient, thereby generating a noise distribution model representing the distribution pattern of outliers in the original point cloud data in three-dimensional space. This noise distribution model provides guidance for setting adaptive filtering parameters, increasing the filtering radius or distance threshold in high-noise regions; maintaining a conservative filtering intensity in the main structure region to prevent over-denoising; and for... An elevation constraint filter is applied to outlier points, and a dynamic elevation tolerance is set.

[0059] S2. Calculate the adaptive filtering parameters according to the noise distribution model, and perform distance filtering, statistical filtering and multi-feature fusion iterative filtering on the original point cloud data according to the adaptive filtering parameters to obtain the final denoised point cloud.

[0060] In this embodiment of the invention, the adaptive filtering parameters include parameters for distance filtering, statistical filtering, and multi-feature fusion iterative filtering, including distance threshold for distance filtering, proportional threshold for statistical filtering, elimination threshold for geometric feature fusion filtering, and sampling threshold for subsequent voxel downsampling, etc., which can be set according to the noise distribution model using preset rules or expert experience.

[0061] In detail, the step of performing distance filtering, statistical filtering, and multi-feature fusion iterative filtering on the original point cloud data according to the adaptive filtering parameters to obtain the final denoised point cloud includes: Calculate the Euclidean distance from each of the original point cloud data to the nearest scan origin, and remove the original point cloud data whose Euclidean distance is greater than a preset distance threshold to obtain distance-denoised point cloud; Calculate the average Euclidean distance and standard deviation of each distance-denoised point cloud in its corresponding neighborhood, and calculate the global mean and global standard deviation of the distance-denoised point cloud based on the average Euclidean distance and the standard deviation. If the average Euclidean distance is greater than the sum of the global mean and the global standard deviation which is a multiple of a set threshold, then the distance-denoised point cloud is determined to be an outlier and removed to obtain a statistically filtered point cloud. Based on a preset initial neighborhood size, the local density, local curvature, and average angle of the normal vectors of each statistically filtered point cloud are calculated to construct a weighted composite anomaly score. Points with weighted composite anomaly scores higher than the preset current dynamic threshold are removed to form the current round of filtering results. The initial neighborhood size and the dynamic threshold are then adaptively updated based on the local density of the point cloud in the current round of filtering results. The current filtering result is iteratively denoised based on the updated initial neighborhood size and dynamic threshold to obtain the final denoised point cloud.

[0062] In detail, in order to remove far-field spurious points caused by long-distance measurement errors, multiple reflections and environmental interference, this invention takes into account that bridge point clouds are usually stitched together by multi-station scanning. If a single origin is used as a reference, it may lead to the accidental deletion of real structural points. Therefore, distance filtering based on the scanning origin is used to perform distance filtering on the original point cloud data.

[0063] Wherein, let the set of origin points corresponding to each scanning station when collecting raw point cloud data be . For each raw point cloud data Calculate its Euclidean distance to the nearest scan origin:

[0064] in, Indicates the first The Euclidean distance from the original point cloud data to the nearest scan origin. Indicates the first One scan origin, This indicates the total number of scan origin points. Represents the norm.

[0065] Furthermore, combined with the maximum effective range of the laser scanning equipment The minimum bounding box space range is constructed, and a dynamic distance threshold is set, which can be expressed as:

[0066] in, Indicates the distance threshold. This represents the distance from the center of the smallest bounding box to the farthest corner point. This indicates the preset safety margin.

[0067] Eliminate all conditions that meet the distance threshold. The original point cloud data is used to obtain distance-denoised point clouds. Distance denoising can quickly remove isolated pseudo-points far from the main structure in the preprocessing stage, significantly improving the accuracy of subsequent neighborhood analysis and providing a reliable data foundation for statistical filtering.

[0068] In this embodiment of the invention, after obtaining the distance-denoised point cloud, a statistical outlier removal method is used to further identify and remove residual local outlier noise. At this point, far-field pseudopoints have been eliminated, the local point distribution tends to be more reasonable, and the subsequent neighborhood size calculation is more accurate.

[0069] In detail, for each distance-denoised point cloud, a k-nearest neighbor algorithm is constructed based on a preset neighborhood range to obtain a neighborhood point cloud. First, the average Euclidean distance between the distance-denoised point cloud and the neighborhood point cloud is calculated, and the standard deviation of the Euclidean distance is calculated based on the average Euclidean distance to obtain the distance standard deviation.

[0070] Specifically, the standard deviation of the distance is calculated using the following formula:

[0071]

[0072] in, Indicates the first Noise removal cloud at a distance The average Euclidean distance to neighboring point clouds. This represents the number of neighboring point clouds. Indicates the first A set of neighborhood point clouds for distance-denoised point clouds. Indicates the first A neighborhood point cloud, This represents the distance from the standard deviation.

[0073] Calculate the average Euclidean distance and standard deviation of all distance-denoised point clouds using the above formula, and let... and For the mean and standard deviation of all distance-denoised point clouds, respectively, set a proportional threshold for the standard deviation. If a denoised cloud at a certain distance satisfies:

[0074] These are then identified as outliers and removed, resulting in a statistically filtered point cloud.

[0075] In this embodiment of the invention, in order to improve the denoising accuracy, the invention adopts a multi-round iterative filtering algorithm, which integrates geometric features such as local density, curvature, and normal vector consistency to construct a comprehensive anomaly criterion, so as to realize the set feature fusion filtering, and dynamically adjusts the neighborhood parameters and the elimination threshold in each iteration.

[0076] In detail, during each iteration, the neighborhood region of each statistical filter point is cloudified according to the preset neighborhood size, and multidimensional local geometric features are extracted based on the neighborhood region. The multidimensional local geometric features include local curvature, the average angle with the normal vector of the neighboring points, and the average Euclidean distance between the neighboring point cloud and the neighborhood region.

[0077] In each iteration, a dynamically changing rejection threshold is set. This threshold starts from the initial global statistical value and gradually tightens with each iteration, achieving progressive noise reduction from coarse to fine. At the same time, the neighborhood size for the next round is adaptively updated based on the local density of the remaining point cloud. In high-density areas, the neighborhood is appropriately reduced to improve detail preservation, while in low-density or edge areas, the minimum neighborhood range is limited to prevent holes, thus balancing the filtering effect with the integrity of the geometric structure.

[0078] Specifically, after each round of statistical filtering point cloud removal where the weighted composite anomaly score exceeds a preset removal threshold, the first filtered point cloud data is obtained. The neighborhood size and removal threshold for the next round are then adaptively updated based on the local density of the remaining points. The neighborhood size is updated according to the local density of the point cloud in the neighborhood region. The first filtered point cloud data is then iteratively denoised based on the updated neighborhood size until the maximum number of iterations is reached or the point cloud stabilizes, resulting in the final denoised point cloud. This iterative process continues until the preset maximum number of iterations is reached, or the removal rate is lower than a set threshold for two consecutive rounds. Finally, the filtered point cloud data optimized through multiple rounds is output.

[0079] Specifically, the step of calculating the composite anomaly score of the statistically filtered point cloud based on the neighborhood region includes: The process involves calculating the local density, local curvature, and average angle of the normal vectors of each statistically filtered point cloud based on a preset initial neighborhood size, and constructing a weighted composite anomaly score, including: The local density is obtained by calculating the ratio of the number of neighborhood points to the average distance within the initial neighborhood size of the statistically filtered point cloud. Construct a decentralized coordinate matrix of the neighborhood points and perform eigenvalue decomposition to obtain eigenvalues. Calculate the ratio of the smallest eigenvalue to the sum of all eigenvalues ​​to obtain the local curvature. The unit normal vector is determined based on the eigenvector corresponding to the largest eigenvalue among the eigenvalues, and the average angle between the normal vectors of the statistically filtered point cloud and the neighboring points is calculated based on the unit normal vector. The local density, local curvature, and average angle of the normal vector are normalized and then summed according to preset weights to obtain a weighted composite anomaly score.

[0080] In this embodiment of the invention, by calculating the nearest neighborhood of each point and extracting local geometric features, the local density is defined as the ratio of the number of points in the neighborhood to the average distance, which can characterize the sparsity of the spatial distribution.

[0081] Specifically, following the steps described above for constructing a decentralized coordinate matrix, a decentralized coordinate matrix for the point cloud data within the neighborhood region is constructed, denoted as [missing information]. Calculate the covariance matrix Then, perform eigenvalue decomposition on it to obtain eigenvalues. Local curvature is defined as the ratio of the minimum eigenvalue to the sum of all eigenvalues. This index can effectively distinguish between planar regions (where local curvature tends to zero) and edge or corner regions (where local curvature tends to one).

[0082] Furthermore, the unit normal vector The eigenvector corresponding to the largest eigenvalue is used to determine the average angle between the eigenvector and the normal vector of the neighboring point cloud data. To measure the consistency of local surface orientation, the average angle between the normal vectors of the statistically filtered point cloud and the neighboring point cloud data is calculated using the following formula:

[0083] in, Indicates the first Statistical Filtered Point Cloud The average angle between the normal vectors of the data and the neighboring point cloud data. Indicates the neighborhood size as A collection of temporal neighborhood point cloud data. Indicates the first Neighborhood point cloud data, Indicates the first The unit normal vector of a statistically filtered point cloud. Indicates the first The unit normal vector of the neighborhood point cloud data.

[0084] Preferably, when the gradient of the local curvature or normal vector exceeds a threshold, the elimination decision for that point is skipped to prevent the loss of key structural features.

[0085] Furthermore, the weighted composite anomaly score of the statistically filtered point cloud is calculated using the following formula:

[0086] in, Indicates a composite abnormality score. , , These represent the preset weights. Indicates the first Statistical Filtered Point Cloud The average Euclidean distance between the data and neighboring point cloud data. This represents the preset average distance reference value. Indicates local curvature. Indicates the first Statistical Filtered Point Cloud The average angle between the normal vectors of the data and the neighboring point cloud data.

[0087] In this embodiment of the invention, The average distance reference value for the neighboring region is taken from the neighboring region. Mean, weighting coefficient satisfy It can be adaptively configured based on the structural region type to enhance the ability to identify different types of noise. Then, a dynamic rejection threshold is set. initial value Determined based on global statistics, and with each iteration round Gradually decrease:

[0088] To achieve progressive noise reduction from coarse to fine. All satisfy... Points that are considered outliers are removed, resulting in a first-filtered point cloud. To adapt to the uniformity of the point cloud's spatial distribution, the neighborhood size is adjusted. Dynamically update based on local density after each iteration:

[0089] in, This is the proportionality coefficient. and As a preset boundary, To optimize the local density of the point cloud in the neighborhood, the filtering stability is enhanced in high-density areas, while avoiding empty neighborhood spaces in low-density or edge areas. The iterative process continues until the relative change in noise removal rate between two consecutive iterations is lower than a preset threshold. or reaching the maximum number of iterations. The final denoised point cloud can effectively remove noise while preserving the key geometric features of the bridge structure to the greatest extent.

[0090] Among them, you can refer to Figure 5 The image shows a schematic diagram of the effect of filtering the original point cloud data. The number of points in the filtered point cloud is 6,796,206, and the minimum bounding box size is: X: 8.5925 (-8.2385: 0.3540), Y: 18.2936 (-9.6109: 8.6827), Z: 6.0621 (0.2736: 6.3357).

[0091] S3. Calculate the spatial density of the final denoised point cloud, and generate a point cloud distribution heatmap based on the spatial density.

[0092] In this embodiment of the invention, spatial density refers to the number of points per unit volume (or area) of the final denoised point cloud, thereby reflecting the spatial density of the initial filtered point cloud data.

[0093] In this embodiment of the invention, for example, the voxel size is set. The final denoised point cloud 3D space is divided into several non-overlapping cubic voxel units to form a regular voxel mesh:

[0094] Among them, voxels Corresponding spatial region:

[0095] The final denoised cloud , Based on their spatial coordinates, they are mapped to the corresponding voxels to form a set of points within the voxels:

[0096] The final denoised point cloud for each regular voxel grid Set the search radius around the center. Construct a spherical neighborhood:

[0097] in, express The neighborhood point set, This is the preset neighborhood radius.

[0098] Count the number of points in this neighborhood. :

[0099] Calculate the volume of the neighborhood space (Volume of a sphere):

[0100] Estimated points Spatial local density (spatial density) :

[0101] Iterate through all point clouds in each regular voxel grid and calculate their local spatial density. Construct a global density field The local spatial density values ​​are mapped to color gradients, with blue to red representing low to high density, thus generating a point cloud density distribution heatmap for visualizing the distribution of structural features.

[0102] In this embodiment of the invention, by generating a point cloud distribution heatmap, key areas can be identified during subsequent downsampling, and adaptive adjustment can be performed based on the density field.

[0103] S4. Based on the point cloud distribution heatmap, perform pseudo-perturbation voxel downsampling on the final denoised point cloud to obtain downsampled point cloud data.

[0104] In this embodiment of the invention, voxel downsampling involves dividing the space into uniform cubes (voxels), with each voxel retaining only one representative point. Existing voxel downsampling methods perform voxel downsampling on each non-empty voxel. The set centroid of the internal point set is calculated as the initial representative point of the voxel. :

[0105] The set of representative points of all voxels constitutes the downsampled point cloud. This method effectively reduces the number of points through spatial clustering, significantly improving computational efficiency while maintaining the overall geometric structure of the point cloud.

[0106] While traditional voxel mesh downsampling is efficient, the fixed meshing method can easily lead to regular artifacts in the downsampling results, such as mesh distribution and step effect. This is especially true at structural edges or in areas with drastic curvature changes, which can cause local geometric information loss or distortion, affecting the accuracy of subsequent feature extraction and defect detection.

[0107] To overcome the aforementioned shortcomings, this invention proposes a density-based pseudo-perturbation voxel downsampling method. Based on standard voxel downsampling, a local nondeterministic perturbation method is introduced to construct a variable perturbation domain, breaking the spatial periodicity of voxel partitioning and making the selection of representative points more spatially applicable. For example... Figure 6 As shown, Figure 6 After mid-sampling, the number of point clouds is 5,741,065, and the minimum bounding box dimensions are X: 8.4898 (-8.1537: 0.3361), Y: 16.528 (-8.2431: 8.2849), Z: 5.3912 (0.2775: 5.6687). Figure 6 The diagram illustrates the effect of downsampling on the original point cloud data, visually demonstrating the advantages of this invention in significantly reducing point cloud density and improving subsequent computational efficiency while preserving structural details.

[0108] Specifically, the step of performing density-feedback-based pseudo-perturbation voxel downsampling on the final denoised point cloud based on the point cloud distribution heatmap to obtain downsampled point cloud data includes: Construct the variable perturbation domain corresponding to each regular voxel grid in the point cloud distribution heatmap; Calculate the spatial local density of each final denoised point cloud within the variable perturbation domain; Calculate density weights based on the spatial local density of the final denoised point cloud. Representative point cloud data are selected from the variable perturbation domain according to the density weights; Based on the representative point cloud data and the point cloud distribution heatmap, the regular voxel grid is subjected to density-based adaptive downsampling to obtain downsampled point cloud data.

[0109] In this embodiment of the invention, for each regular voxel grid corresponding to the point cloud distribution heatmap At its geometric center Based on this, a variable perturbation domain is constructed. Defined as Centered on, with a preset disturbance radius of spherical region:

[0110] Among them, the disturbance radius Non-fixed values ​​are determined by the local density of the point cloud within a regular voxel mesh. Dynamic adjustment.

[0111] Specifically, in the subsequent downsampling process, adaptive adjustment can be made according to the global density field: more structural details are retained in high-density areas and the downsampling intensity is reduced; sampling compression is appropriately enhanced in low-density areas to improve the overall processing efficiency, thereby achieving efficient and intelligent point cloud simplification while ensuring geometric fidelity.

[0112] In this embodiment of the invention, the spatial local density of each final denoised point cloud within the variable perturbation domain can be calculated using the above-described method for calculating spatial local density, and then the density weight can be calculated using the following formula:

[0113] in, Indicates the first The spatial local density of the final denoised cloud The corresponding density weights This indicates the preset parameters.

[0114] Representative point cloud data is selected using the following formula:

[0115] in, This indicates that the data represents point cloud data. Indicates the first The final denoised cloud. Indicates the first The density weights corresponding to the final denoised point cloud. Represents a regular voxel grid. This represents the variable perturbation domain.

[0116] Preferably, representative point cloud data is used as downsampled point cloud data. To further improve the adaptability of 3D point cloud data preprocessing to multi-scale geometric features, the present invention achieves adaptive control of disturbance intensity through local density feedback.

[0117] Disturbance radius Based on the average spatial local density within the rule-based voxel grid Dynamically adjust using the following formula:

[0118] in, This represents the average local density within the voxel; The Sigmoid normalization function maps the density to... interval; , These are the minimum and maximum values ​​for the preset disturbance radius.

[0119] For example, in areas with a high density of bridges, Larger areas increase the disturbance radius, expand the sampling candidate range, reduce local point density, and suppress data redundancy; in low-density areas or edge areas of the bridge, Smaller size reduces the disturbance radius, shrinks the disturbance range, weakens the disturbance intensity, and preserves key structural features.

[0120] S5. Calculate the point cloud quality evaluation index of the downsampled point cloud data, and perform feedback optimization on the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data.

[0121] In this embodiment of the invention, to further ensure the geometric fidelity and engineering usability of the preprocessing results, the point cloud uniformity index, density coefficient change index, structural edge retention rate index, noise residual rate index, normal vector distribution change index, and point spacing standard deviation change index of the downsampled point cloud data are calculated.

[0122] Specifically, the calculation of the point cloud quality evaluation index for the downsampled point cloud data includes: The downsampled point cloud data is divided into a set of voxels. The normalized standard deviation of the downsampled point cloud data in each voxel within the set of voxels is calculated. The normalized standard deviation is used as a point cloud uniformity index. Calculate the relative change in local density of each downsampled point cloud data to obtain a density coefficient change index; The edge point set of the original point cloud data is detected by vector mutation, the nearest neighbor point is found from the downsampled point cloud data, and the structural edge retention rate index is calculated based on the edge point set and the nearest neighbor point. The proportion of noise points in the downsampled point cloud data is calculated using the statistical filtering residual, and the proportion of noise points is used as the noise residual rate index. Calculate the normal vector distribution change index based on the mean and standard deviation of the angle between the normal vectors of the original point cloud data and the downsampled point cloud data; The standard deviation of the point spacing is calculated based on the average distance standard deviation between each of the original point cloud data and each of the downsampled point cloud data and the nearest point cloud. By combining the point cloud uniformity index, the density coefficient change index, the structure edge retention rate index, the noise residual rate index, the normal vector distribution change index, and the point spacing standard deviation change index, the point cloud quality evaluation index of the downsampled point cloud data is obtained.

[0123] In this embodiment of the invention, the point cloud uniformity index The closer the value is to 1, the more uniform the distribution. It is used to measure the spatial distribution uniformity of the point cloud after filtering and downsampling. It is defined as the normalized standard deviation of the number of points within all non-empty voxels. The point cloud uniformity index is calculated using the following formula:

[0124] in, This represents the point cloud uniformity index. Indicates the first The standard deviation of the number of downsampled point cloud data in individual elements. No. The mean number of downsampled point cloud data in individual pixels. This represents a preset constant.

[0125] The density coefficient change index is used to quantify the relative change in local density before and after preprocessing. For each downsampled point cloud data, before performing distance filtering, statistical filtering, multi-feature fusion iterative filtering, and voxel downsampling, the local density in the original point cloud data is... The local density after treatment is The density coefficient change index Represented as:

[0126] in, The smaller the value, the more stable the density distribution of the downsampled point cloud data, and the more complete the information of key areas is preserved.

[0127] Furthermore, the structure edge retention rate metric is used to evaluate the ability of filtering and voxel downsampling to preserve structure edge features. Edge point sets in the original point cloud data are detected through normal vector mutation. Then, find its nearest neighbor in the downsampled point cloud data. If the Euclidean distance is less than a preset threshold If the margin is not specified, it is considered retained, and the marginal retention rate can be calculated using the following formula:

[0128] in, This represents the percentage of structural edges retained. This represents a set of downsampled point cloud data. Indicates the first One set of raw point cloud data, Indicates the first The nearest neighbor of each original point cloud data point in the downsampled point cloud data point. This represents the set of edge points.

[0129] A larger value indicates that the edge structure is better preserved.

[0130] The noise residual rate is used to evaluate the denoising effect. The proportion of noisy points in the processed point cloud is calculated by statistically analyzing the filter residual.

[0131] in, This represents the set of raw point cloud data that is determined to be noise. This represents a set of downsampled point cloud data.

[0132] The above indicators are scored using a weighted composite method. Quantifying overall quality: The normal vector distribution change index calculates the mean and standard deviation of the angle between the normal vectors of the original point cloud data before processing, as well as the mean and standard deviation of the angle between the normal vectors of the downsampled point cloud data after processing.

[0133] in, Indicator of change in normal vector distribution Indicates the first The normal vector of the original point cloud data. Indicates the first The normal vector of the downsampled point cloud data corresponding to the original point cloud data. This represents the total number of downsampled point cloud data. This indicates the calculation of the angle between vectors.

[0134] in, The smaller the value, the smaller the surface geometric deformation.

[0135] The standard deviation of the inter-point spacing index is used to calculate the average standard deviation of the distance between each downsampled point cloud data and its k nearest neighbors. Compare the changes before and after the treatment:

[0136] in, This represents the index indicating the change in the standard deviation of the point spacing. This represents the standard deviation of the average distance between downsampled point cloud data and its nearest neighbor point clouds. This represents the standard deviation of the average distance between the original point cloud data and its nearest neighbor point clouds. The smaller the value, the less likely the filtering and downsampling have damaged the local density structure.

[0137] Specifically, the step of optimizing the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data includes: A weighted comprehensive score is calculated based on the point cloud uniformity index, density coefficient change index, structure edge retention rate index, and noise residue rate index in the point cloud quality evaluation index. A feedback method is generated when the weighted comprehensive score is greater than a preset score threshold or when any of the point cloud quality assessment indicators does not meet the preset indicator range. The adaptive filtering parameters are optimized according to the feedback method to obtain optimized filtering parameters; The downsampled point cloud data is optimized based on the optimized filtering parameters to obtain the target point cloud data.

[0138] The weighted composite score is calculated using the following formula:

[0139] in, This represents the weighted overall score. , indicating the preset indicator weights, This represents the point cloud uniformity index. Indicators representing changes in density coefficient This represents the percentage of structural edges retained. This indicates the noise residual rate.

[0140] In this embodiment of the invention, the comprehensive score If the value is less than the preset threshold or any indicator exceeds the standard, the system triggers a feedback method to automatically adjust the adaptive filtering parameters. The adjusted parameters are used to optimize the downsampled point cloud data until the point cloud quality assessment parameters meet the requirements and the target point cloud data is obtained.

[0141] For example, adjusting the filter radius: If If the threshold is too high, it indicates insufficient noise reduction. Increase the statistical filter's proportional threshold or distance threshold using the gradient ascent method. .

[0142] Sampling ratio adjustment: If or The density field calculated using the above steps does not meet the requirements. For low-density areas ( Reduce the downsampling intensity to decrease the disturbance radius in this area. .

[0143] Density compensation coefficient adjustment: Introduce preset weights at the edges or coefficient regions. And compensation introduction Dynamically adjust the generation of representative points:

[0144] in, Follow By reducing and increasing, low-density areas rely more on weighted centers to preserve the structure.

[0145] This feedback method achieves closed-loop control of "evaluation-diagnosis-optimization", which significantly improves the robustness of point cloud data preprocessing.

[0146] S6. Convert the target point cloud data into a standard format file and simultaneously generate a structured report containing processing parameters, operation logs, and quality assessment charts.

[0147] Furthermore, after preprocessing and optimization, the raw point cloud data yields the target point cloud data in .pcd or .las format, ensuring compatibility with mainstream point cloud processing platforms. Simultaneously, a structured processing report is generated, including: parameter settings (initial values ​​and adjustment records for parameters such as filtering, downsampling, perturbation, and compensation); processing logs (input and output point counts, time consumption, and key operations at each stage); and quality assessment results (four indicator values ​​and a comprehensive score). Error analysis charts.

[0148] In one optional embodiment of the present invention, a user interface module can be constructed to receive user input parameters, monitor processing progress, and visualize the point cloud processing process and results. The system is integrated into on-site computing devices as lightweight software or deployed on vehicle / mobile terminals, supporting rapid startup and processing within 72 hours after a disaster.

[0149] In this embodiment of the invention, a multi-dimensional quality assessment system is constructed, covering point cloud uniformity, density variation coefficient, edge retention rate, noise residual rate, normal vector distribution stability, and standard deviation variation of point spacing. This system can objectively quantify the preprocessing effect of the original point cloud data. At the same time, combined with error feedback, the system automatically adjusts the filtering radius, sampling ratio, and density compensation coefficient to complete the closed-loop optimization and standardized output of high-quality point clouds, ensuring that the data can be directly used for post-disaster damage identification models.

[0150] In one embodiment, the method of the present invention preprocesses the original point cloud data, which can complete the processing of millions of point cloud data within 5 minutes with a noise residual rate of less than 2% and an edge retention rate of more than 25%, truly achieving "fast but not rough, simple but not missing" post-disaster treatment; breaking the regularity constraints of traditional voxel meshes, while compressing the data volume by 21%, it effectively preserves the geometric integrity of high-risk areas such as supports, cracks, and beam ends. Furthermore, by constructing quantifiable multi-dimensional point cloud quality assessment indicators, it supports closed-loop parameter optimization, ensuring that the output point cloud can be directly input into the intelligent damage identification model, avoiding the risk of misjudgment due to improper preprocessing; and it is specifically designed for emergency assessment of bridges after disasters such as earthquakes, floods, and impacts. Its efficiency and robustness have been verified in test scenarios, and it can serve as a key technical support for the rapid determination of bridge safety and traffic capacity within the "golden 72 hours".

[0151] like Figure 7 The diagram shown is a functional block diagram of a three-dimensional point cloud preprocessing system for bridge emergency detection and assessment provided in an embodiment of the present invention.

[0152] The 3D point cloud preprocessing system 100 for bridge emergency detection and assessment described in this invention can be installed in a processing device. Depending on the functions implemented, the 3D point cloud preprocessing system 100 may include a noise distribution model construction module 101, a filtering module 102, a variable disturbance domain construction module 103, a point cloud quality assessment index calculation module 104, an iterative filtering module 105, and a data encapsulation and report generation module 106. The module described in this invention can also be called a unit, which refers to a series of computer program segments that can be executed by an electronic device processor and perform a fixed function, stored in the memory of the electronic device.

[0153] The specific execution methods for the steps in each of the above modules are the same as the corresponding execution steps in the three-dimensional point cloud preprocessing method for bridge emergency detection and assessment described above.

[0154] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example.

[0155] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A three-dimensional point cloud preprocessing method for emergency bridge inspection and assessment, characterized in that, include: Acquire raw point cloud data, construct the minimum bounding box corresponding to the raw point cloud data, and construct a noise distribution model of the raw point cloud data based on the minimum bounding box. Based on the noise distribution model, adaptive filtering parameters are calculated, and distance filtering, statistical filtering, and multi-feature fusion iterative filtering are performed on the original point cloud data according to the adaptive filtering parameters to obtain the final denoised point cloud. Calculate the spatial density of the final denoised point cloud, and generate a point cloud distribution heatmap based on the spatial density; Based on the point cloud distribution heatmap, the final denoised point cloud is downsampled using pseudo-perturbation voxels based on density feedback to obtain downsampled point cloud data. Calculate the point cloud quality evaluation index of the downsampled point cloud data, and perform feedback optimization on the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data; The target point cloud data is converted into a standard format file, and a structured report containing processing parameters, operation logs, and quality assessment charts is generated simultaneously.

2. The three-dimensional point cloud preprocessing method for bridge emergency detection and assessment as described in claim 1, characterized in that, The construction of the minimum bounding box corresponding to the original point cloud data includes: Calculate the data centroid of the original point cloud data, and construct a decentralized coordinate matrix based on the data centroid; Calculate the covariance matrix of the original point cloud data based on the decentralized coordinate matrix; The covariance matrix is ​​decomposed into eigenvectors to obtain eigenvectors, and a rotation matrix is ​​constructed based on the eigenvectors. Calculate the minimum and maximum coordinates of the original point cloud data on the coordinate axes based on the rotation matrix and the data centroid; The minimum bounding box corresponding to the original point cloud data is determined based on the minimum and maximum coordinates.

3. The three-dimensional point cloud preprocessing method for bridge emergency detection and assessment as described in claim 1, characterized in that, The step of constructing a noise distribution model for the original point cloud data based on the minimum bounding box includes: The minimum bounding box is divided into multiple voxel intervals along the coordinate axis, and the density histogram of each voxel interval is calculated. Calculate the density histogram using the following formula: in, Indicates the first Density histogram of individual element intervals, Indicates the first The total number of original point cloud data in the individual pixel interval. Indicates the first Individual element interval along the first The interval length of each coordinate axis Indicates the first The original point cloud data in the individual pixel interval is perpendicular to The average cross-sectional area of ​​the shaft. These represent the principal axis directions of the horizontal axis, vertical axis, and depth axis, respectively. The density gradient is calculated based on the density histogram, and the noise distribution model is determined based on the density gradient.

4. The three-dimensional point cloud preprocessing method for bridge emergency detection and assessment as described in claim 1, characterized in that, The step of performing distance filtering, statistical filtering, and multi-feature fusion iterative filtering on the original point cloud data according to the adaptive filtering parameters to obtain the final denoised point cloud includes: Calculate the Euclidean distance from each of the original point cloud data to the nearest scan origin, and remove the original point cloud data whose Euclidean distance is greater than a preset distance threshold to obtain distance-denoised point cloud; Calculate the average Euclidean distance and standard deviation of each distance-denoised point cloud in its corresponding neighborhood, and calculate the global mean and global standard deviation of the distance-denoised point cloud based on the average Euclidean distance and the standard deviation. If the average Euclidean distance is greater than the sum of the global mean and the global standard deviation which is a multiple of a set threshold, then the distance-denoised point cloud is determined to be an outlier and removed to obtain a statistically filtered point cloud. Based on a preset initial neighborhood size, the local density, local curvature, and average angle of the normal vectors of each statistically filtered point cloud are calculated to construct a weighted composite anomaly score. Points with weighted composite anomaly scores higher than the preset current dynamic threshold are removed to form the current round of filtering results. The initial neighborhood size and the dynamic threshold are then adaptively updated based on the local density of the point cloud in the current round of filtering results. The current filtering result is iteratively denoised based on the updated initial neighborhood size and dynamic threshold to obtain the final denoised point cloud.

5. The three-dimensional point cloud preprocessing method for bridge emergency detection and assessment as described in claim 4, characterized in that, The process involves calculating the local density, local curvature, and average angle of the normal vectors of each statistically filtered point cloud based on a preset initial neighborhood size, and constructing a weighted composite anomaly score, including: The local density is obtained by calculating the ratio of the number of neighborhood points to the average distance within the initial neighborhood size of the statistically filtered point cloud. Construct a decentralized coordinate matrix of the neighborhood points and perform eigenvalue decomposition to obtain eigenvalues. Calculate the ratio of the smallest eigenvalue to the sum of all eigenvalues ​​to obtain the local curvature. The unit normal vector is determined based on the eigenvector corresponding to the largest eigenvalue among the eigenvalues, and the average angle between the normal vectors of the statistically filtered point cloud and the neighboring points is calculated based on the unit normal vector. The average angle between the normal vectors of the statistically filtered point cloud and the neighboring point cloud is calculated using the following formula: in, Indicates the first Statistical Filtered Point Cloud The average angle between the normal vectors of the points and the neighboring points. Indicates the neighborhood size as The set of time-neighboring points, Indicates the first Neighboring points, Indicates the first The unit normal vector of a statistically filtered point cloud. Indicates the first The unit normal vector of each neighboring point; The local density, local curvature, and average angle of the normal vector are normalized and then summed according to preset weights to obtain a weighted composite anomaly score.

6. The three-dimensional point cloud preprocessing method for bridge emergency detection and assessment as described in claim 1, characterized in that, The step of performing density feedback-based pseudo-perturbation voxel downsampling on the final denoised point cloud according to the point cloud distribution heatmap to obtain downsampled point cloud data includes: Construct the variable perturbation domain corresponding to each regular voxel grid in the point cloud distribution heatmap; Calculate the spatial local density of each final denoised point cloud within the variable perturbation domain; Calculate density weights based on the spatial local density of the final denoised point cloud. Representative point cloud data are selected from the variable perturbation domain according to the density weights; Representative point cloud data can be selected using the following formula: in, This indicates that the data represents point cloud data. Indicates the first The final denoised cloud. Indicates the first The density weights corresponding to the final denoised point cloud. Represents a regular voxel grid. Represents the variable perturbation domain; Based on the representative point cloud data and the point cloud distribution heatmap, the filtered voxel interval is subjected to density-based adaptive downsampling to obtain downsampled point cloud data.

7. The three-dimensional point cloud preprocessing method for bridge emergency detection and assessment as described in claim 1, characterized in that, The calculation of the point cloud quality evaluation index for the downsampled point cloud data includes: The downsampled point cloud data is divided into a set of voxels. The normalized standard deviation of the downsampled point cloud data in each voxel within the set of voxels is calculated. The normalized standard deviation is used as a point cloud uniformity index. The point cloud uniformity index is calculated using the following formula: in, This represents the point cloud uniformity index. Indicates the first The standard deviation of the number of downsampled point cloud data in individual elements. No. The mean number of downsampled point cloud data in individual pixels. This represents a preset constant; Calculate the relative change in local density of each downsampled point cloud data to obtain a density coefficient change index; The edge point set of the original point cloud data is detected by vector mutation, the nearest neighbor point is found from the downsampled point cloud data, and the structural edge retention rate index is calculated based on the edge point set and the nearest neighbor point. The edge retention rate metric is calculated using the following formula: in, This represents the percentage of structural edges retained. This represents a set of downsampled point cloud data. Indicates the first One set of raw point cloud data, Indicates the first The nearest neighbor of each original point cloud data point in the downsampled point cloud data point. Represents the set of edge points; The proportion of noise points in the downsampled point cloud data is calculated using the statistical filtering residual, and the proportion of noise points is used as the noise residual rate index. Calculate the normal vector distribution change index based on the mean and standard deviation of the angle between the normal vectors of the original point cloud data and the downsampled point cloud data; The standard deviation of the point spacing is calculated based on the average distance standard deviation between each of the original point cloud data and each of the downsampled point cloud data and the nearest point cloud. By combining the point cloud uniformity index, the density coefficient change index, the structure edge retention rate index, the noise residual rate index, the normal vector distribution change index, and the point spacing standard deviation change index, the point cloud quality evaluation index of the downsampled point cloud data is obtained.

8. The three-dimensional point cloud preprocessing method for bridge emergency detection and assessment as described in claim 1, characterized in that, The step of optimizing the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data includes: A weighted comprehensive score is calculated based on the point cloud uniformity index, density coefficient change index, structure edge retention rate index, and noise residue rate index in the point cloud quality evaluation index. The weighted composite score is calculated using the following formula: in, This indicates the weighted overall score. , indicating the preset indicator weights, This represents the point cloud uniformity index. Indicators representing changes in density coefficient This represents the percentage of structural edges retained. Indicator representing noise residual rate; A feedback method is generated when the weighted comprehensive score is greater than a preset score threshold or when any of the point cloud quality assessment indicators does not meet the preset indicator range. The adaptive filtering parameters are optimized according to the feedback method to obtain optimized filtering parameters; The downsampled point cloud data is optimized based on the optimized filtering parameters to obtain the target point cloud data.

9. A three-dimensional point cloud preprocessing system for bridge emergency detection and assessment, characterized in that, include: The noise distribution model construction module is used to acquire raw point cloud data, construct the minimum bounding box corresponding to the raw point cloud data, and construct the noise distribution model of the raw point cloud data based on the minimum bounding box. The filtering module is used to calculate adaptive filtering parameters based on the noise distribution model, and perform distance filtering, statistical filtering, and multi-feature fusion iterative filtering on the original point cloud data based on the adaptive filtering parameters to obtain the final denoised point cloud. The point cloud distribution heatmap generation module is used to calculate the spatial density of the final denoised point cloud and generate a point cloud distribution heatmap based on the spatial density. The voxel downsampling module is used to perform pseudo-perturbation voxel downsampling based on density feedback on the final denoised point cloud according to the point cloud distribution heatmap, so as to obtain downsampled point cloud data. The feedback adjustment module is used to calculate the point cloud quality evaluation index of the downsampled point cloud data, and to perform feedback optimization on the downsampled point cloud data based on the point cloud quality evaluation index to obtain the target point cloud data. The data encapsulation and report generation module is used to convert the target point cloud data into a standard format file and simultaneously generate a structured report containing processing parameters, operation logs, and quality assessment charts.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1-7.

Images