An engineering structure full-field deformation detection method based on three-dimensional point cloud comparative analysis
By combining a method based on 3D point cloud comparative analysis with UAV photogrammetry technology and point cloud processing algorithms, the problems of low efficiency and low accuracy in traditional detection methods are solved. This enables rapid and accurate full-field deformation detection of engineering structures, which is applicable to complex field conditions and provides an efficient and low-cost detection solution.
Patent Information
- Application Number
- CN202310367251.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2023-04-07
- Publication Date
- 2025-12-30
- Estimated Expiration
- 2043-04-07
AI Technical Summary
Traditional methods for detecting deformation in engineering structures are inefficient and produce low-quality data under complex field conditions. They cannot acquire deformation information across the entire field, pose safety hazards, and are labor-intensive, failing to meet the demand for rapid, accurate, efficient, and low-cost detection.
By employing a method based on 3D point cloud comparative analysis, combined with UAV close-up photogrammetry technology and 3D point cloud processing algorithms, full-field deformation detection of engineering structures is achieved through point cloud data generation, preprocessing, multi-point cloud data registration, and deformation detection.
It enables rapid and accurate full-field deformation detection of engineering structures, achieving millimeter-level detection accuracy. It is applicable to various field conditions, providing an efficient and low-cost detection method, and ensuring the comprehensiveness and safety of the detection results.
Smart Images
Figure CN116518864B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of building deformation extraction technology, and in particular to a method for detecting full-field deformation of engineering structures based on three-dimensional point cloud comparative analysis. Background Technology
[0002] Under conditions of load, continuous environmental influence, and sudden disasters during construction or service, engineering structures often undergo critical changes such as deformation, displacement, tilting, and torsion, which in turn affect the overall durability and safety of the structure. Traditional methods for quantitatively detecting structural deformation mainly involve deploying sensors and high-precision measuring instruments (such as levels and total stations) to continuously and repeatedly observe key points of the structure, thereby obtaining the critical behavior data.
[0003] Currently used detection methods have the following problems:
[0004] (1) Under complex engineering site conditions, especially in post-disaster emergency assessment, measurement efficiency and data quality cannot be guaranteed. Technical personnel have a large workload and high labor intensity, which is dangerous.
[0005] (2) Engineering structures are often huge in size. Traditional measurement methods that rely on surveying instruments can only observe the key deformation behavior of engineering structures at discrete points, resulting in blind spots and the inability to obtain full-field deformation information of the building. There are certain safety hazards in the detection and monitoring of special key nodes.
[0006] From the perspectives of "the safety of testing personnel, the economy of testing methods, and the comprehensiveness of testing results," the current traditional deformation testing and evaluation technology for single-mode engineering structures can no longer meet the requirements of speed, accuracy, high efficiency, low cost, and adaptability to on-site conditions. Summary of the Invention
[0007] The purpose of this invention is to provide a method for detecting full-field deformation of engineering structures based on three-dimensional point cloud comparative analysis. It integrates close-up photogrammetry technology and three-dimensional point cloud processing algorithms, using close-up photogrammetry technology of UAVs as a non-contact detection medium to achieve full-field detection; and uses point cloud processing algorithms as an intelligent performance perception means, aiming at rapid detection and accurate evaluation of key performance characteristics of engineering structures, with engineering structure deformation as the key performance characteristic, to achieve rapid evaluation and quantification of key performance characteristics of engineering structures.
[0008] The objective of this invention can be achieved through the following technical solutions:
[0009] A method for detecting full-field deformation of engineering structures based on 3D point cloud comparative analysis includes the following steps:
[0010] S1. Point cloud data generation: Select artificial control points and / or natural control points arranged on the surface of the structure as feature points, acquire image sequences of the structure, and complete the reconstruction of the three-dimensional point cloud model based on multi-view optical images to obtain the initial point cloud model.
[0011] S2. Point cloud model preprocessing: Spatial indexing, noise filtering, and downsampling are performed on the initial point cloud model to obtain a point cloud model that is easy to calculate. The point cloud model is then clustered and segmented to obtain a local point cloud model of the structure.
[0012] S3. Multi-point cloud data registration: Based on the RANSAC algorithm, the affine transformation matrix of two sets of local point clouds is roughly calculated to obtain the initial affine transformation matrix. Based on the ICP algorithm, the fine affine transformation matrix is determined using the unsampled point cloud. The point clouds are then finely registered to obtain a local point cloud registration model of two sets of point clouds in the same spatial dimension.
[0013] S4. Point cloud deformation detection: Based on the local point cloud registration model, the deformation values of the structure are determined by using a deformation detection method based on the coordinate system, a deformation detection method based on point cloud registration, or a deformation detection method based on the fusion of coordinate system and point cloud registration for three different scenarios.
[0014] S1 includes the following steps:
[0015] S101, Image Sequence Acquisition:
[0016] Based on the method of close-up photogrammetry, artificial and / or natural image control points are set up, the flight path of the UAV is planned, and high-resolution image sequences of the target are obtained.
[0017] S102, Point Cloud Data Generation:
[0018] Based on the high-resolution image sequence of the target, the motion reconstruction method is used to reconstruct the three-dimensional point cloud and obtain the initial three-dimensional point cloud model.
[0019] S2 includes the following steps:
[0020] S201. Create a spatial index for the initial point cloud model;
[0021] S202, Noise Filtering:
[0022] The acquired structured point cloud model is denoised using a neighborhood algorithm to obtain a denoised point cloud model.
[0023] S203, Point Cloud Downsampling:
[0024] For the denoised point cloud model, point cloud downsampling is performed based on the voxel method, which reduces the number of three-dimensional points while preserving the geometric structural features of the point cloud.
[0025] S204, Point Cloud Clustering and Segmentation:
[0026] S2041. For the downsampled point cloud model, the RANSAC algorithm is used to perform preliminary clustering and segmentation of the point cloud, and the coarsely extracted local analysis point cloud is obtained by fitting the digital model.
[0027] S2042. After obtaining the coarse extraction of local analysis point cloud data, fine extraction of local analysis point cloud is performed, which includes the following steps: determining the height threshold of the structural wall, completing point cloud segmentation based on the density of the point cloud projected along a specific direction, and then removing discrete noise points based on the Kmeans clustering algorithm to complete the fine extraction of the local analysis point cloud of the structure and obtain the local point cloud model of the structure.
[0028] S3 includes the following steps:
[0029] S301. Determine the affine transformation matrix:
[0030] When the relative poses of the point clouds are completely unknown, the initial affine transformation matrices of the two sets of point clouds are roughly calculated by performing a global search matching based on the RANSAC algorithm.
[0031] After obtaining the initial affine transformation matrix, the rigid transformation between two point clouds is estimated by iteratively minimizing the distance difference based on the ICP algorithm. Fine registration is then performed on the unsampled point cloud to obtain the fine affine transformation matrix.
[0032] S302, Point Cloud Data Registration:
[0033] Based on the fine affine transformation matrix, iterative calculations are performed on the point cloud data to obtain fused point cloud data under the same spatial dimension, namely the local point cloud registration model of spatial coordinate system one.
[0034] The RANSAC algorithm in S301 specifically includes the following steps:
[0035] Based on the finely extracted local point cloud, the normal vector of each point in the two sets of point clouds is estimated, and the FPFH feature of each point is calculated based on the normal vector to obtain the geometric feature data of the point cloud.
[0036] Based on the geometric feature data of point clouds, global registration is performed using the RANSAC algorithm. n random points are randomly selected from the source point cloud P, and their corresponding points in the target point cloud Q are detected by querying the nearest neighbors in the 33-dimensional FPFH feature space. The registration point pairs are iterated multiple times, and the optimal result is selected according to the principle of minimizing error. The initial affine transformation matrix is then calculated.
[0037] The initial affine transformation matrix includes a rotation matrix R and a translation matrix t:
[0038] P = {p1, p2, ..., p} n}, Q={q1,q2,…,q n}
[0039] Q = RP + t
[0040] Where P is the source point cloud, Q is the target point cloud, and n is the number of randomly selected points.
[0041] The ICP algorithm in S301 specifically includes the following steps:
[0042] Based on the initial affine transformation matrix, the source point cloud is transformed into the coordinate system of the target point cloud, thus completing the initialization;
[0043] Calculate the difference between the source point cloud and the target point cloud, and use the difference as the evaluation result:
[0044]
[0045] Based on the ICP algorithm, a distance threshold is set. When the distance between corresponding points in two sets of point clouds is less than the threshold, they are considered as corresponding points. Thus, two sets of n new registration points with one-to-one correspondence are obtained in the two sets of point clouds, namely P and Q.
[0046] Update the affine transformation matrix based on the new registration point pair, and repeat the above steps until the evaluation result meets the pre-configured threshold to obtain the fine affine transformation matrix with accurate registration.
[0047] S4 includes the following steps:
[0048] S401. Calculate the overlapping regions of different point clouds:
[0049] Based on the local point cloud registration model, the bounding boxes of different point clouds are calculated, and the intersection of the bounding boxes is used to obtain the overlapping area of the point clouds.
[0050] S402, Point Cloud Deformation Detection:
[0051] After extracting the overlapping regions of the two sets of point clouds, deformation values of the overlapping regions of the structural point clouds are quantified using a coordinate-based deformation detection method, a point cloud registration-based deformation detection method, or a coordinate-and-point-cloud registration fusion deformation detection method for three different scenarios.
[0052] S401 specifically includes the following steps:
[0053] After performing point cloud registration transformation based on the fine affine transformation matrix to achieve coordinate system one for the two sets of point clouds, the maximum and minimum coordinates of the point clouds are calculated respectively to obtain the range of the point clouds and create bounding boxes.
[0054] Based on the bounding boxes of the two sets of point clouds, the union of the bounding boxes is taken in the deformation direction of the point clouds to ensure that the overlapping area of the two sets of point clouds is completely preserved. The intersection of the other two directions is taken to obtain the overlapping area of the point clouds.
[0055] The coordinate system-based deformation detection method is as follows: after verifying the boundary range of the overlapping area of the two sets of point clouds, the point cloud data is gridded by means of the three-dimensional coordinates of the two sets of point clouds, and the structural deformation value is obtained by calculating the one-way C2C distance of the structural wall.
[0056] The deformation detection method based on point cloud registration specifically involves: calculating the coordinate differences between point clouds based on the fine affine transformation matrix to obtain the structural deformation value;
[0057] The deformation detection method based on coordinate system and point cloud registration fusion is as follows: the local unidirectional deformation of the structure is obtained by using the deformation detection method based on the coordinate system, and the rigid body displacement value between the two sets of point clouds is obtained by using the deformation detection method based on point cloud registration; the local unidirectional deformation and the local rigid body displacement value are superimposed to obtain a more accurate structural deformation detection result.
[0058] Compared with the prior art, the present invention has the following beneficial effects:
[0059] (1) High accuracy: The accuracy of the present invention in the deformation extraction of the support wall of the foundation pit project under construction has been verified. The deformation value extracted by the present invention is not much different from the actual value. The comparison and analysis with the reference deformation results of the pre-embedded inclination monitoring points of the project shows that the three proposed deformation detection methods can reach the millimeter level accuracy when optimal, and the local detection accuracy can reach the sub-millimeter level. It realizes the automatic detection of global deformation on the surface of the engineering structure, which meets the needs of engineering practice to a certain extent.
[0060] (2) Wide applicability: The present invention adopts photogrammetry technology, which can provide a low-cost and accurate means for evaluating the key performance of engineering structures. In addition, it uses image control points to control the model coordinate system, which can obtain a high-precision three-dimensional point cloud model even in scenarios with weak satellite positioning signals while ensuring scale accuracy.
[0061] (3) Comprehensive data: This invention realizes the quantitative detection of continuous full-field deformation of engineering structures based on three-dimensional models. Compared with two-dimensional images, the multi-dimensional information of three-dimensional point clouds can more realistically and intuitively display the actual scene, providing a complete data guarantee for the evaluation results of key performance characteristics.
[0062] (4) Fast calculation: Compared with traditional deformation identification methods, this invention is faster and more accurate, and has adaptability to on-site conditions. It can accurately obtain the deformation of the entire structure and realize the rapid evaluation and quantification of the key performance of engineering structures. Attached Figure Description
[0063] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0064] Figure 2 A flowchart illustrating the process of generating point cloud data;
[0065] Figure 3 This is a schematic diagram of the point cloud deformation detection process;
[0066] Figure 4 The above is a surface deformation detection result of the foundation pit wall in one embodiment, wherein (a) is a surface deformation heat map of the foundation pit wall based on the coordinate system, (b) is a surface deformation heat map of the foundation pit wall based on registration, and (c) is a surface deformation heat map of the foundation pit wall based on the fusion of coordinate system and registration.
[0067] Figure 5 This is a comparison of inclinometer data in one embodiment with the deformation amount of various deformation detection methods. Detailed Implementation
[0068] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments. These embodiments are based on the technical solution of the present invention and provide detailed implementation methods and specific operating procedures. However, the scope of protection of the present invention is not limited to the following embodiments.
[0069] This embodiment provides a method for detecting full-field deformation of engineering structures based on three-dimensional point cloud comparative analysis, such as... Figure 1 As shown, it includes the following steps:
[0070] S1, Point Cloud Data Generation
[0071] like Figure 2 As shown, it includes the following steps:
[0072] S101, Image Sequence Acquisition:
[0073] Before collecting image data, a real-world survey of the target is conducted. Points that naturally exist in the scene and are not prone to spatial mapping errors are selected as natural image control points, and target points are artificially set up as artificial image control points. The three-dimensional coordinate information of the image control points is collected.
[0074] Based on close-up photogrammetry, image control points are set up, the flight path of the UAV is planned, and a high-resolution image sequence of the target is acquired.
[0075] S102, Point Cloud Data Generation:
[0076] Based on the high-resolution image sequence of the target, the motion reconstruction method is used to reconstruct the three-dimensional point cloud and obtain the initial three-dimensional point cloud model.
[0077] 3D point cloud data is characterized by disorder, sparsity, and unstructured nature. This means that even though point cloud data contains information such as the three-dimensional spatial attributes of objects, it needs to be processed by a series of algorithms to extract the required information, such as point cloud preprocessing, point cloud segmentation, and point cloud registration.
[0078] S2-S4 are the structural deformation detection steps, and their process is as follows: Figure 3 As shown.
[0079] S2. Point cloud model preprocessing: Spatial indexing, noise filtering, and downsampling are performed on the initial point cloud model to obtain a point cloud model that is easy to calculate. The point cloud model is then clustered and segmented to obtain a local point cloud model of the structure.
[0080] Point cloud models obtained from 3D reconstruction are difficult to directly use for extracting key quantitative indicators of engineering structures. Point cloud preprocessing aims to generate high-quality point clouds for subsequent processing. Point cloud model preprocessing typically includes spatial index creation, noise filtering, downsampling, and clustering segmentation. Specifically, it includes the following steps:
[0081] S201. Create a spatial index for the initial point cloud model.
[0082] Based on the Octree algorithm, large-granular spatial partitioning and index creation are performed. The space enclosed by the bounding box of the point cloud model is divided into eight cubes (eight leaf nodes). Cubes that do not contain any point set are deleted. Then, the cubes containing point set are repeated in the above eight-eight partitioning steps until the edge length of the cube is less than the given leaf node granularity.
[0083] Then, based on KDTree, a subdivision index is performed, the variance of the data in each dimension is calculated, and the dimension with the largest variance is taken as the initial segmentation coordinate axis. Starting from a point in space, the entire space is divided into two by a hyperplane perpendicular to the segmentation coordinate axis. Then, the segmentation steps are repeated in each of the two spaces until all spatial points are processed, and a relatively ordered and structured 3D point cloud is obtained.
[0084] S202, Noise Filtering:
[0085] The obtained structured point cloud model is denoised using a neighborhood algorithm to obtain a denoised point cloud model.
[0086] S203, Point Cloud Downsampling:
[0087] For the denoised point cloud model, point cloud downsampling is performed based on the voxel method, which reduces the number of 3D points while preserving the geometric structure features of the point cloud.
[0088] S204, Point Cloud Clustering and Segmentation:
[0089] S2041. For the downsampled point cloud model, the RANSAC algorithm is used to perform preliminary clustering and segmentation of the point cloud, and the coarsely extracted local analysis point cloud is obtained by fitting the digital model.
[0090] S2042. After obtaining the coarse extraction of local analysis point cloud data, fine extraction of local analysis point cloud is performed, which includes the following steps: determining the height threshold of the structural wall, completing point cloud segmentation based on the density of the point cloud projected along a specific direction, and then removing discrete noise points based on the Kmeans clustering algorithm to complete the fine extraction of the local analysis point cloud of the structure and obtain the local point cloud model of the structure.
[0091] S3, Multi-point Cloud Data Registration
[0092] In computer vision, pattern recognition, and robotics, point cloud registration, also known as scan matching, is the process of finding spatial transformations (e.g., scaling, rotation, and translation) that align two point clouds. The purpose of finding such a transformation involves merging multiple datasets into a globally consistent model (or coordinate system) and mapping new measurements onto known datasets to identify features or estimate their pose.
[0093] Point cloud registration is also an important method for multi-source point cloud data fusion. It can be used to detect small deformations on the surface of the same object and is of great significance in point cloud processing. Point cloud registration algorithms mainly include the RANSAC algorithm and the ICP algorithm.
[0094] Specifically, it includes the following steps:
[0095] S301. Determine the affine transformation matrix:
[0096] S3011. When the relative poses of the point clouds are completely unknown, a global search matching is performed based on the RANSAC algorithm to roughly calculate the initial affine transformation matrices of the two sets of point clouds.
[0097] The RANSAC algorithm specifically includes the following steps:
[0098] Based on the finely extracted local point cloud, the normal vector of each point in the two sets of point clouds is estimated, and the FPFH feature of each point is calculated based on the normal vector to obtain the geometric feature data of the point cloud.
[0099] Based on the geometric feature data of point clouds, global registration is performed using the RANSAC algorithm. n random points are randomly selected from the source point cloud P, and their nearest neighbors are found in the 33-dimensional FPFH feature space to detect their corresponding points in the target point cloud Q. The registration point pairs are iterated multiple times, and the optimal result is selected based on the principle of minimizing error. The initial affine transformation matrix is calculated, including the rotation matrix R and the translation matrix t.
[0100] P = {p1, p2, ..., p}n}, Q={q1,q2,…,q n}
[0101] Q = RP + t
[0102] Where P is the source point cloud, Q is the target point cloud, and n is the number of randomly selected points.
[0103] S3012. After obtaining the initial affine transformation matrix, the rigid transformation between the two point clouds is estimated by minimizing the distance difference through the ICP algorithm in an iterative manner. Fine registration is then performed on the unsampled point cloud to obtain the fine affine transformation matrix.
[0104] The ICP algorithm specifically includes the following steps:
[0105] Based on the initial affine transformation matrix, the source point cloud is transformed into the coordinate system of the target point cloud, thus completing the initialization;
[0106] Calculate the difference between the source point cloud and the target point cloud, and use the difference as the evaluation result:
[0107]
[0108] Based on the ICP algorithm, a distance threshold is set. When the distance between corresponding points in two sets of point clouds is less than the threshold, they are considered as corresponding points. Thus, two sets of n new registration points with one-to-one correspondence are obtained in the two sets of point clouds, namely P and Q.
[0109] Update the affine transformation matrix based on the new registration point pair, and repeat the above steps until the evaluation result meets the pre-configured threshold to obtain the fine affine transformation matrix with accurate registration.
[0110] S302, Point Cloud Data Registration:
[0111] Based on the fine affine transformation matrix, iterative calculations are performed on the point cloud data to obtain fused point cloud data under the same spatial dimension, namely the local point cloud registration model of spatial coordinate system one.
[0112] S4. Point cloud deformation detection: Based on the local point cloud registration model, the deformation values of the structure are determined by using a deformation detection method based on the coordinate system, a deformation detection method based on point cloud registration, or a deformation detection method based on the fusion of coordinate system and point cloud registration for three different scenarios.
[0113] Specifically, it includes the following steps:
[0114] S401. Calculate the overlapping regions of different point clouds:
[0115] After performing point cloud registration transformation based on the fine affine transformation matrix to achieve coordinate system one for the two sets of point clouds, the maximum and minimum coordinates of the point clouds are calculated respectively to obtain the range of the point clouds and create bounding boxes.
[0116] Based on the bounding boxes of the two sets of point clouds, the union of the bounding boxes is taken in the deformation direction of the point clouds to ensure that the overlapping area of the two sets of point clouds is completely preserved. The intersection of the other two directions is taken to obtain the overlapping area of the point clouds.
[0117] S402, Point Cloud Deformation Detection:
[0118] After extracting the overlapping regions of the two sets of point clouds, deformation values of the overlapping regions of the structural point clouds are quantified using a coordinate-based deformation detection method, a point cloud registration-based deformation detection method, or a coordinate-and-point-cloud registration fusion deformation detection method for three different scenarios.
[0119] The deformation detection method based on the coordinate system is as follows: after verifying the boundary range of the overlapping area of two sets of point clouds, the point cloud data is gridded using the mesh slicing method based on the three-dimensional coordinates of the two sets of point clouds, and the structural deformation value is obtained by calculating the one-way C2C distance of the structural wall.
[0120] The deformation detection method based on point cloud registration is as follows: calculate the coordinate differences between point clouds according to the fine affine transformation matrix to obtain the structural deformation value.
[0121] The deformation detection method based on coordinate system and point cloud registration fusion is as follows: the local unidirectional deformation of the structure is obtained by using the coordinate system-based deformation detection method, and the rigid body displacement value between the two sets of point clouds is obtained by using the point cloud registration-based deformation detection method; the local unidirectional deformation and local rigid body displacement value of the structure are superimposed to obtain a more accurate structural deformation detection result.
[0122] To verify the feasibility and accuracy of this invention, a foundation pit support structure under construction was used as the research object. Two image data acquisition experiments were designed and implemented. High-precision point cloud data of the foundation pit wall at different times in the same area were obtained through refined 3D reconstruction. After completing the above series of point cloud processing steps, interference-free foundation pit wall point cloud data was extracted. The overlapping area of the two sets of point clouds was extracted, and the deformation of the overlapping area of the foundation pit wall was quantified using the three deformation detection methods proposed in this invention. The results are as follows. Figure 4 As shown.
[0123] In this project, to monitor the horizontal displacement of the foundation pit soil, the construction team, based on site conditions, followed the basic layout principle of pre-burying one inclinometer tube every 20m along the extension direction of the foundation pit in the retaining wall. Each inclinometer tube had a measuring point every 0.5m downwards along its height, and the horizontal displacement of the soil inside each measuring point was calculated daily using an inclinometer with millimeter-level accuracy. In this experiment, a pre-buried inclinometer point ZQT21 was found in a certain section, which can provide a relatively accurate deformation reference value for verifying the deformation accuracy of this method.
[0124] To verify the deformation detection accuracy of the above methods, the horizontal lateral deformation monitoring results of the pit wall support structure by the inclinometer were used as deformation reference values, and the comparison results of each deformation detection method were obtained as follows: Figure 5 As shown in Table 1, the reference error between the specific detection method of the present invention and the deformation reference value is as follows.
[0125] Table 1. Relative errors between deformation detection results and inclinometer results.
[0126]
[0127]
[0128] Table 1 shows that the results obtained by the registration-based deformation detection method are basically consistent with the inclinometer data; the deformation detection results of each method are similar in shape, but there are different degrees of deviation in the relative deformation amount.
[0129] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A full-field deformation detection method for engineering structures based on comparative analysis of three-dimensional point clouds, characterized in that, The method comprises the following steps: S1, point cloud data generation: selecting artificial and / or natural control points arranged on the structure surface as feature points, collecting image sequences of the structure, and reconstructing a three-dimensional point cloud model based on multi-view optical images to obtain an initial point cloud model; S2, point cloud model preprocessing: creating a spatial index, noise filtering, and downsampling processing on the initial point cloud model to obtain a point cloud model convenient for calculation, and performing cluster segmentation on the point cloud model to obtain a local point cloud model of the structure; S3, multi-point cloud data registration: coarsely calculating an affine transformation matrix of two local point clouds based on a RANSAC algorithm to obtain an initial affine transformation matrix, and determining a fine affine transformation matrix based on an ICP algorithm using the non-downsampled point cloud to finely register the point clouds and obtain a local point cloud registration model of the two point clouds in the same spatial dimension; S4, point cloud deformation detection: based on the local point cloud registration model, three different scene-based deformation detection methods are used to determine the structure deformation value, including a coordinate system-based deformation detection method, a point cloud registration-based deformation detection method, and a coordinate system and point cloud registration fusion-based deformation detection method; The S4 comprises the following steps: S401, calculating the overlapping area of different point clouds: Based on the local point cloud registration model, the boundary boxes of different point clouds are calculated, and the intersection of the boundary boxes is taken to obtain the point cloud overlapping area; The S401 specifically comprises the following steps: After the two point clouds are transformed based on the fine affine transformation matrix to realize the same coordinate system, the maximum and minimum coordinates of the point clouds are calculated to obtain the range of the point clouds, and the boundary boxes are created; Based on the boundary boxes of the two point clouds, the union of the boundary boxes in the deformation direction of the point clouds is taken to ensure that the overlapping area of the two point clouds is completely reserved, and the intersection of the other two directions is taken to obtain the overlapping area of the point clouds; S402, point cloud deformation detection: After the overlapping area of the two point clouds is extracted, three different scene-based deformation detection methods are used to quantify the deformation value of the structure point cloud overlapping area, including a coordinate system-based deformation detection method, a point cloud registration-based deformation detection method, and a coordinate system and point cloud registration fusion-based deformation detection method.
2. The method according to claim 1, wherein, The S1 comprises the following steps: S101, image sequence collection: Based on the close-range photogrammetry method, artificial and / or natural control points are arranged, the flight route of the unmanned aerial vehicle is planned, and the high-resolution image sequence of the target is obtained; S102, point cloud data generation: Based on the high-resolution image sequence of the target, a three-dimensional reconstruction of the point cloud is performed by using the motion recovery structure method to obtain an initial three-dimensional point cloud model.
3. The method according to claim 1, wherein, The S2 comprises the following steps: S201, creating a spatial index for the initial point cloud model; S202, noise filtering: Based on the neighborhood algorithm, the obtained structured point cloud model is denoised to obtain a denoised point cloud model; S203, point cloud downsampling: Based on the voxel method, the denoised point cloud model is downsampled to reduce the number of three-dimensional points while retaining the geometric structure features of the point cloud; S204, point cloud cluster segmentation: S2041, for the down-sampling point cloud model, the RANSAC algorithm is used for preliminary clustering segmentation of the point cloud, and the rough extracted local analysis point cloud is obtained by fitting the digital model; S2042, after obtaining the rough extracted local analysis point cloud data, fine extraction of the local analysis point cloud is carried out, which specifically includes the following steps: determining the height threshold of the structure wall surface, completing point cloud segmentation according to the density of point cloud projection along a specific direction, and removing discrete noise points based on the Kmeans clustering algorithm, to complete fine extraction of the structure local analysis point cloud, and obtain the local point cloud model of the structure.
4. The method according to claim 1, wherein, The S3 includes the following steps: S301, determine the affine transformation matrix: In the case that the relative pose of the point cloud is completely unknown, the RANSAC algorithm is used for global search matching, and the initial affine transformation matrix of the two groups of point clouds is roughly calculated; After obtaining the initial affine transformation matrix, the ICP algorithm is used to estimate the rigid transformation between the two point clouds by iteratively minimizing the distance difference, to finely register the point cloud without down-sampling, and obtain the fine affine transformation matrix; S302, point cloud data registration: Based on the fine affine transformation matrix, the point cloud data is iteratively calculated to obtain the fused point cloud data in the same spatial dimension, i.e. the local point cloud registration model with the same spatial coordinate system.
5. The method according to claim 4, wherein, The RANSAC algorithm in S301 specifically includes the following steps: Based on the fine extracted local point cloud, the normal vector of each point in the two groups of point clouds is estimated, the FPFH feature of each point is calculated based on the normal vector, and the geometric feature data of the point cloud is obtained; Based on the geometric feature data of the point cloud, the RANSAC algorithm is used for global registration, n random points are randomly selected from the source point cloud P, the nearest neighbor in the 33-dimensional FPFH feature space is queried, the corresponding point in the target point cloud Q is detected, the registration point pairs are iterated for multiple times, and the optimal result is selected according to the minimum error principle, and the initial affine transformation matrix is calculated.
6. The method according to claim 5, wherein, The initial affine transformation matrix includes a rotation matrix R and a translation matrix t: P = {p1, p2,..., p n}, Q = {q1, q2,..., q n} Q=RP+t Wherein, P is the source point cloud, Q is the target point cloud, and n is the number of selected random points.
7. The method according to claim 6, wherein, The ICP algorithm in S301 specifically includes the following steps: Based on the initial affine transformation matrix, the source point cloud is transformed into the coordinate system of the target point cloud, and the initialization is completed; The difference between the source point cloud and the target point cloud is calculated, and the difference is taken as the evaluation result: Based on the ICP algorithm, a distance threshold is set, and when the distance between the corresponding points in the two groups of point clouds is less than the threshold, the corresponding points are considered, so that n pairs of new registration points are obtained in the two groups of point clouds P and Q; Based on the new registration point pairs, the affine transformation matrix is updated, and the above steps are repeated until the evaluation result meets the preconfigured threshold, to obtain the fine affine transformation matrix for accurate registration.
8. The method according to claim 1, wherein, The deformation detection method based on the coordinate system specifically includes: after checking the boundary range of the overlapping area of the two groups of point clouds, the point cloud data is gridded by means of the grid slicing method based on the three-dimensional coordinates of the two groups of point clouds, and the one-way C2C distance of the structure wall surface is calculated to obtain the structure deformation value. The deformation detection method based on point cloud registration specifically comprises: calculating the coordinate difference between the point clouds according to the fine affine transformation matrix to obtain a structure deformation value. The deformation detection method based on coordinate system and point cloud registration fusion specifically comprises: obtaining a local one-way deformation of the structure by using the deformation detection method based on the coordinate system, and obtaining a rigid body displacement value between the two groups of point clouds by using the deformation detection method based on the point cloud registration; and superimposing the local one-way deformation of the structure and the local rigid body displacement value to obtain a more accurate structure deformation detection result.
Citation Information
Patent Citations
Tower-type ancient building deformation monitoring and analyzing method and device
CN109916323A