A deformation detection method directly based on TLS high-density point cloud measured values

Abstract

This invention relates to the field of surveying and mapping science and technology, and discloses a deformation detection method directly based on measured values ​​of high-density point clouds (TLS). The method includes: 1) point cloud voxel segmentation based on an octree structure; 2) registration of point cloud data from two consecutive periods; and 3) deformation comparison between the voxels of the later point cloud and the earlier point cloud. Each point in the later point cloud is located in the space of the earlier point cloud voxels according to the octree index, containing the voxel or the voxel closest to that point. The distance from each point to its corresponding voxel's local differential plane is calculated, thus obtaining the deformation at the corresponding point. This invention can comprehensively locate deformed areas and statistically analyze the deformation. It overcomes the problem that the applicability of existing methods is limited by the distribution of target points or feature points, and can effectively extract image feature points in smooth areas where texture changes are not significant.

Description

Technical Field

[0001] This invention relates to the field of surveying and mapping science and technology, and in particular to a deformation detection method based directly on measured values ​​of TLS high-density point clouds. Background Technology

[0002] Currently, large military and civilian equipment / facilities such as antennas, oil tanks, bridges, and tunnels gradually undergo deformation over long periods of use, typically reaching the millimeter level or higher. The magnitude of this deformation determines whether the equipment meets safety standards, making research on deformation detection methods for such large equipment of significant importance. Deformation detection generally occurs in localized areas of the tested equipment, requiring periodic deformation localization and measurement. This invention addresses the problem of non-contact deformation detection at the millimeter level or higher for such large equipment / facilities.

[0003] The core of deformation detection is the comparison problem, that is, comparing the changes between the measured values ​​and the design values, or comparing the changes between the measured values ​​of the previous period and the measured values ​​of the next period.

[0004] References [Chen Yangbo, Yi Guodong, Zhang Shuyou. A method for detecting warping deformation of curved surfaces based on point cloud feature comparison [J]. Journal of Zhejiang University (Engineering Science), 2021, 55(01):81-88.] propose a deformation detection method based on point cloud feature comparison for the problem of warping deformation of curved surfaces. This method uses the comparison of the feature point positions of the measured point cloud and the template point cloud to locate the deformed parts. The template point cloud needs to be generated based on the three-dimensional model of the device under test. References [Yan Tianhao. Deformation detection and three-dimensional modeling of vertical rescue well wall based on machine vision [D]. Chang'an University, 2020.] use image feature point extraction and matching to reconstruct the three-dimensional region of the target area, construct the cross section of the target area, and then detect the deformed parts based on the difference between the measured value and the design value. References [Li Guinan. Research on stereo vision deformation detection of asphalt pavement using mobile cloud service [D]. China University of Mining and Technology, 2019.] generate a dense three-dimensional point cloud of the target under test based on stereo vision imaging method, and perform deformation detection of the target area based on the local flatness of the point cloud.

[0005] References [Yang Qiao, Jin Banghu. Intelligent Detection System for Deformation of Large Building Structures Based on Laser Sensors [J]. Laser Journal, 2020, 41(08):224-229.] address the deformation problem of large building structures by deploying cooperative optical fibers at the deformation sites and acquiring the displacement of each point along the optical fibers through multiple laser sensors to detect the deformation amount; Reference [Guan Yuanyuan. Deformation Detection of Large Objects Based on Binocular Stereo Vision [D]. Xi'an University of Technology, 2018.] uses a cooperative target method to measure the positional changes of the target site and monitor the deformation of the target site; Reference [Zhao Jiaxing. Research on Image-Based Bridge Deformation Detection Method [D]. Chang'an University, 2020.] obtains the three-dimensional point cloud of the target area based on image matching theory, and then detects the deformation amount by matching the feature points of two phases of data.

[0006] Existing methods can be summarized as follows: 1) Detection based on measured values ​​of the target area and known models (or local geometric regularities, such as planes or circles). This type of method requires prior knowledge of the three-dimensional model or geometric characteristics of the target area and can only detect the relative deformation of the target area, thus having a limited scope of application; 2) Detection of target area point movement based on cooperative target points or extracted image feature points. This type of method specifically measures the deformation of surface points in a portion of the target area, is easily limited by the distribution of target points or feature points, and is difficult to effectively extract image feature points in smooth areas where texture changes are not significant.

[0007] The Terrestrial Laser Scanning System (TLS), also known as 3D laser imaging radar, Light Detect and Ranging (LiDAR), employs the principle of active polar coordinate laser imaging. It senses the distance from the target to the laser emission point based on the time difference or phase difference between laser emission and reception. It measures the 3D coordinates of the target point by measuring laser range and angle using a angular scale. It can acquire high-density 3D point coordinates on the surface of objects in the measured scene with high measurement accuracy. A typical TLS can achieve a point measurement accuracy of ±2mm within a range of 100 meters. Summary of the Invention

[0008] To address the limitations of existing deformation detection methods, such as narrow applicability or limited effectiveness, this invention proposes a deformation detection method directly based on measured values ​​of TLS high-density point clouds.

[0009] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0010] A deformation detection method based directly on measured values ​​of high-density point clouds in TLS, comprising the following steps:

[0011] 1) Point cloud voxel segmentation based on octree structure: First, the previous point cloud is segmented into voxel sets. In the image, pixels are cubes, and one pixel corresponds to a region in the measured space. A voxel is a cube in three-dimensional space. A voxel space contains several three-dimensional coordinate points. The point cloud uses three-dimensional points to represent three-dimensional objects, while voxels are represented by cubes. During the voxel generation process, information statistics are performed based on the point set within the voxel to give the voxel more dimensional information: including normal, curvature, principal components, dimensional feature descriptors, and point density. After the voxel structure is generated, the distribution of points within each voxel has a consistent coplanarity, which is called a supervoxel.

[0012] 2) Registration of point cloud data from two different periods: Point cloud comparison requires that the two sets of point clouds (data from two different periods) be registered to a unified coordinate system, i.e., "aligned" the two sets of point clouds. To achieve this, the point cloud data P = {p1, p2, ..., p...} n} and Q = {q1,q2,...,q n The unification of the coordinate system requires the following calculation formula:

[0013]

[0014] The rotation matrix R in PQ Translation vector T PQ Among them, point p(x) p ,y p ,z p ) and q(x q ,y q ,z q For a pair of points with the same name: solve the above system of equations using 4 or more pairs of common points, where R PQ Represented using Rodrigue matrix and solved using singular value decomposition;

[0015] Point cloud registration involves a coarse-to-fine process, i.e., coarse registration → fine registration. In the coarse registration stage, a simpler method is used: manually selecting at least four pairs of common points and calculating the coordinate system transformation parameters between the two point cloud sets based on these pairs. PQ and T PQ ;

[0016] 3) Comparison of voxels between later point cloud and earlier point cloud. The premise of realizing the comparison of deformation between the two data is to determine the correspondence between data points in the two data.

[0017] By constructing voxels from the previous point cloud using an octree structure, the correspondence between points in the later point cloud and voxels from the previous point cloud is directly found based on the octree index: for a point P = (x, y, z) in the later point cloud, the distribution space contains voxels LN of that point. PThe position of the leaf node is determined based on its point coordinates and the octree encoding method. Let the deepest layer number of the octree subdivision be n, and calculate the code code of the leaf node at layer n where point P is located. P , and then perform layer-by-layer addressing of LN P according to this code; if LN P is at the mth (m < n) layer, then the node obtained when addressing to the mth layer is the voxel LN P ;

[0018] Let LN P be at the sibling nodes. The same upper-level node divides into 8 sub-nodes, and the three-axis direction numbers in the sibling nodes are respectively (i nx , i ny , i nz ), i n ∈ [1, 2], then the total number in the three-axis direction of LN P is:

[0019]

[0020] And

[0021]

[0022] Let the side length of the voxel at layer n be w o , then the relationship between the total number in the three-axis direction and point P is LN P

[0023]

[0024] where floor() is the rounding function;

[0025] The total number in the three axes of the neighboring voxels of LN P is

[0026]

[0027] According to the above formula, the indexing of neighboring voxels can be achieved;

[0028] In the later stage, each point in the point cloud is located in the space of the previous point cloud voxel according to the octree index, including the voxel containing the point or the voxel closest to the point; calculate the distance from each point to the local differential plane patch of each corresponding voxel, and the deformation amount at the corresponding point position can be obtained.

[0029] A deformation detection method based directly on measured values ​​of high-density point clouds using TLS (Tree of Truth) is proposed. The voxel segmentation of the point cloud employs point cloud plane extraction using octree voxel growth. Based on the point cloud density and thickness distribution, the approximate plane of the local voxel flatness is used as the segmentation termination index. The point cloud is initially segmented into a series of uniform octree voxels using a fixed voxel size. The number of initial segmentation layers is adjusted according to the average voxel point density to adapt the octree structure to the point cloud density distribution. Then, through calculation and statistics of each voxel information, the voxel set with the most approximate planar distribution is obtained, and the segmentation termination condition of the octree structure is statistically determined. Finally, through recursive segmentation, an octree voxel set adaptive to the point cloud density and thickness distribution is obtained.

[0030] Due to the adoption of the technical solution described above, the present invention has the following advantages:

[0031] This invention employs TLS to acquire high-density 3D point cloud data of the surface of the target device under test. Then, by directly comparing the point cloud data from two different periods, it comprehensively locates deformed areas and calculates the deformation amount. Point cloud data can be described as scattered 3D coordinate data points on the visible surface of the tested scene. Compared to an image, a point cloud can be simply understood as pixel coordinates containing depth (distance) information.

[0032] The deformation detection method proposed in this invention is not limited by the shape and texture of the target object. It can use two-phase point cloud data of TLS non-contact measurement to directly and completely obtain deformation detection information such as the deformation area and deformation amount of the entire surface of the target. Attached Figure Description

[0033] Figure 1 Point cloud map of large oil tank equipment;

[0034] Figure 2 Voxel sets are distinguished by color;

[0035] Figure 3 A schematic diagram of coarse registration of point clouds in two phases based on four sets of common points;

[0036] Figure 4 Two phases of point cloud data were generated, with noise artificially added to the later phase of the point cloud data.

[0037] Figure 5 Image after registration of point clouds in two phases;

[0038] Figure 6 Deformation detection results. Detailed Implementation

[0039] like Figure 1 , 2 As shown in Figures 3, 4, 5, and 6, a deformation detection method directly based on measured values ​​of high-density point clouds in TLS comprises the following steps:

[0040] 1) Point cloud voxel segmentation based on an octree structure: In the image, pixels are represented as cubes, and each pixel corresponds to a region in the measured space. A voxel is a cube in three-dimensional space. A voxel contains several three-dimensional coordinate points. Point clouds use three-dimensional points to represent three-dimensional objects, while voxels are represented by cubes. During voxel generation, information is statistically analyzed based on the point set within the voxel, which can endow the voxel with more dimensional information (including normal, curvature, principal components, dimensional feature descriptors, point density, etc.). After the voxel structure is generated, the distribution of points within each voxel is consistent (e.g., coplanar), which is called a supervoxel (still simply referred to as a voxel).

[0041] This invention first segments the previous point cloud into a voxel set. The voxel segmentation of the point cloud is referenced in the literature [Point Cloud Plane Extraction Using Octree Voxel Growth. Optics and Precision Engineering. 2018, 26(1):172-183.]. An octet tree structure is used, and based on the point cloud density and thickness distribution, the local (within-voxel) flatness (planar approximation) of the voxel segmentation is used as the segmentation termination index. The point cloud is initially segmented into a series of uniform octet tree voxels using a fixed voxel size. The number of initial segmentation layers is adjusted according to the average voxel point density to adapt the octet tree structure to the point cloud density distribution. Then, through calculation and statistics of each voxel information, the voxel set with the closest planar distribution is obtained, and the segmentation termination condition of the octet tree structure is statistically obtained. Finally, through recursive segmentation, an octet tree voxel set adaptive to the point cloud density and thickness distribution is obtained. Figure 1 and Figure 2 As shown.

[0042] 2) Registration of point cloud data from two different periods: Point cloud comparison requires that the two sets of point clouds (data from two different periods) be registered to a unified coordinate system, i.e., "aligned" the two sets of point clouds. To achieve this, the point cloud data P = {p1, p2, ..., p...} n} and Q = {q1,q2,...,q n The unification of the coordinate system requires the calculation formula.

[0043]

[0044] The rotation matrix R in PQ Translation vector T PQ Among them, point p(x) p ,y p ,z p ) and q(x q ,y q ,z q Let R be a pair of points with the same name. Solve the above system of equations using four or more pairs of common points, where R... PQRepresented by the Rodrigues matrix and solved by the singular value decomposition method [Yang Fan, Li Guangyun, Wang Li. Research on three-dimensional coordinate transformation method [J]. Bulletin of Surveying and Mapping, 2010, 0(6): 5-7.].

[0045] The registration of point clouds generally needs to go through a process from coarse to fine, that is, coarse registration → fine registration. In the coarse registration stage, a relatively simple method is adopted, that is, manually select more than 4 groups of common point pairs, and calculate the coordinate transformation parameters (R PQ and T PQ ) between the two groups of point clouds based on the common point pairs. As Figure 3 shown.

[0046] The fine registration is implemented by using the Iterative Closest Point (ICP) related algorithm proposed in the literature [Besl, P.J. and McKay, N.D. A Method for Registration of 3-D Shapes [J]. Transactions on Pattern Analysis and Machine Intelligence, 1992, 14(2): 239-256.]. The effects before and after point cloud registration are as Figures 4-5 shown.

[0047] 3) The deformation comparison between the later point cloud and the voxels of the previous point cloud. The premise for realizing the deformation comparison of the two-phase data is to determine the corresponding relationship between the data points in the two-phase data. Since the octree structure is used to establish voxels for the previous point cloud, directly find the corresponding relationship between each point in the later point cloud and the previous voxels based on the octree index: For a point P=(x, y, z) in the later point cloud, the position of the voxel (LN P , leaf node) that contains this point in the distribution space is determined according to its point coordinates and the octree coding method. Let the deepest layer number of the octree subdivision be n, calculate the code P of the leaf node at the nth layer where the P point is located, and then perform layer-by-layer addressing of LN P according to this code. If LN P is located at the mth (m < n) layer, the node obtained when addressing to the mth layer is the voxel LN P . <00...​​​​​​​​​​​​​

[0049]

[0050] and

[0051]

[0052] Let the side length of the n-layer voxel be w. o Then the relationship between the total number in the three axes and point P is LN P

[0053]

[0054] floor() is the floor function. LN P The triaxial total number of the neighboring voxels is

[0055]

[0056] The above formula can be used to index neighboring voxels.

[0057] In the later point cloud, each point can be located in the earlier point cloud voxels using an octree index, either as the voxel containing that point or the closest voxel. Calculating the distance from each point to the local differential plane patch of its corresponding voxel yields the deformation at that point.

Claims

1. A deformation detection method directly based on measured values ​​of high-density point clouds in TLS, characterized by: Its steps as follows: 1) Point cloud voxel segmentation based on octree structure: First, the previous point cloud is segmented into voxel sets. Pixels in the image are individual cubes, and one pixel corresponds to a region in the measured space. A voxel is a cube in three-dimensional space. A voxel space contains several three-dimensional coordinate points. The point cloud uses three-dimensional points to represent three-dimensional objects, while voxels are represented by cubes. During the voxel generation process, information statistics are performed based on the point set within the voxel to give the voxel more dimensional information: including normal, curvature, principal components, dimensional feature descriptors, and point density. After the voxel structure is generated, the distribution of points within each voxel has a consistent coplanarity, which is called a supervoxel. 2) Registration of point cloud data between two periods: The prerequisite for point cloud comparison is that the two sets of point clouds are registered to a unified coordinate system, that is, the two sets of point clouds are "aligned". To achieve the registration of point cloud data between the two periods, P = {p1, p2, ..., p...} n } and Q = {q1,q2,...,q n The unification of the coordinate system requires the following calculation formula: The rotation matrix R in PQ Translation vector T PQ Among them, point p(x) p ,y p ,z p ) and q(x q ,y q ,z q For a pair of points with the same name: solve the above system of equations using 4 or more pairs of common points, where R PQ Represented using Rodrigue matrix and solved using singular value decomposition; Point cloud registration involves a coarse-to-fine process, i.e., coarse registration → fine registration. In the coarse registration stage, a simpler method is used: manually selecting at least four pairs of common points and calculating the coordinate system transformation parameters between the two point cloud sets based on these pairs. PQ and T PQ ; 3) Comparison of voxels between later point cloud and earlier point cloud. The premise of realizing the comparison of deformation between the two data is to determine the correspondence between data points in the two data. By constructing voxels from the previous point cloud using an octree structure, the correspondence between points in the later point cloud and voxels from the previous point cloud is directly found based on the octree index: for a point P = (x, y, z) in the later point cloud, the distribution space contains voxels LN of that point. P The position of the leaf node is determined based on its coordinates and the octree encoding method. Let the deepest level of the octree be n, then calculate the encoding code of the leaf node at level n where point P is located. P Then perform LN according to this encoding. P Layer-by-layer addressing; if the LNP is located at the m-th layer, where m is less than n, then the node obtained when addressing to the m-th layer is the voxel LN. P ; Let LN P In sibling nodes, the same parent node splits into 8 child nodes, which are sibling nodes with their three-axis direction numbers as follows (i nx i ny i nz ),i n If ∈[1,2], then LN P The overall numbering in the three-axis direction is: and Let the side length of the n-layer voxel be w. o Then the relationship between the total number in the three axes and point P is LN P floor() is the floor function; LN P The triaxial total number of the neighboring voxels is Based on the above formula, indexing of neighboring voxels can be achieved; In the later point cloud, each point is located in the space of the voxels of the earlier point cloud according to the octree index, which is the voxel containing the point or the voxel closest to the point; the distance of each point to the local differential plane patch of each corresponding voxel is calculated, thus obtaining the deformation at the corresponding point.

2. The deformation detection method based directly on measured values ​​of TLS high-density point clouds according to claim 1, characterized in that: The voxel segmentation of the point cloud is performed using point cloud plane extraction from octree voxel growth. Based on the point cloud density and thickness distribution, the flatness of the local voxel plane is used as the segmentation termination index. The point cloud is initially segmented into a series of uniform octree voxels with a fixed voxel size. The number of initial segmentation layers is adjusted according to the average voxel point density to adapt the octree structure to the point cloud density distribution. Then, through calculation and statistics of each voxel information, a set of voxels with a planar distribution is obtained, and then the segmentation termination condition of the octree structure is statistically obtained. Finally, through recursive segmentation, an octree voxel set adaptive to the point cloud density and thickness distribution is obtained.

Images