Three-dimensional reconstruction splicing method for large-size component based on structured light measurement

Abstract

A three-dimensional reconstruction splicing method of large-size component based on structured light measurement belongs to the technical field of structured light automatic three-dimensional measurement. The present application aims at the problem that the excessive initial pose deviation of the existing structured light three-dimensional reconstruction algorithm of the component will cause the local optimal solution of the algorithm to converge in advance and reduce the accuracy of point cloud registration when two point cloud blocks are registered. It comprises the following steps: a structured light sensor is used to sequentially scan along the width direction of a large three-dimensional component at a preset scanning width to obtain sequentially marked point clouds; preprocessing is performed to obtain preprocessed point cloud blocks; the adjacent previous point cloud block is sequentially taken as a target block, and the adjacent next point cloud block is taken as a source point cloud block to register the target block, first spatial pose coarse registration is performed, and then the coordinate transformation of the coarse registration is performed on the source point cloud block to register the target block; and then the three-dimensional reconstruction splicing of the large three-dimensional component is realized. The present application is used for the three-dimensional reconstruction of large-size components.

Description

TECHNICAL FIELD

[0001] The present application relates to a large-size component three-dimensional reconstruction splicing method based on structured light measurement, and belongs to the technical field of structured light automatic three-dimensional measurement. BACKGROUND

[0002] Three-dimensional reconstruction based on computer vision has made rapid progress in the past few decades. Three-dimensional reconstruction based on vision is a popular field in the field of computer vision. The concept is to rely on relevant information collection instruments to collect surface information of objects, obtain two-dimensional image information, then analyze and process the information, and finally apply three-dimensional reconstruction related techniques and methods to realize the reconstruction of object surface contour information in the real environment. Three-dimensional reconstruction based on vision is widely used in robot sensing, self-driving cars, medical detection, and industrial automation due to its high efficiency, high precision, simplicity, and real-time performance.

[0003] In industrial production, for example, the skin frame commonly used in wing parts has the characteristics of complex structure, and its three-dimensional contour detection and reconstruction are very important. For the selection of three-dimensional reconstruction methods, the applicability, real-time performance, and three-dimensional reconstruction results of various methods need to be considered. First, the three-dimensional reconstruction of workpieces is generally carried out indoors, which is convenient for building a measurement platform. Second, unlike outdoor scene reconstruction, workpieces require higher reconstruction accuracy and clearer detail performance. Structured light method becomes the first choice for indoor workpiece detection due to its fast reconstruction result rate, high accuracy, low energy consumption, and strong anti-interference. Since the length of a single laser stripe of line structured light is limited, it is difficult to complete the three-dimensional reconstruction of large workpieces such as skin frames in one time. In addition, due to the limitations of point cloud collection equipment in the field of view, the measured object may have occlusions due to its complex shape, causing visual blind spots, so it is difficult to fully describe a complete object through a single scan. Currently, the most common method to solve the problem of incomplete coverage of single camera field of view in the field of three-dimensional vision is to use multiple cameras to extract multi-angle information or use a single camera to scan and then splice the point cloud data in the later stage. The most common solution for machine vision manufacturers is to use a multi-view single-angle line structured light sensor system. This method is expensive, has a large installation volume, requires communication between multiple sensors, has high hardware requirements for the host computer, and reduces the real-time performance of the system.

[0004] When a single structured light sensor is combined with a motion system to realize the traversal scanning of a large and complex workpiece, the difficulty and length of the post-point cloud processing greatly increase, and the real-time performance is difficult to meet, so efficient and high-precision processing of the point cloud becomes a problem to be solved. In order to obtain a complete three-dimensional model, firstly, multiple point cloud data of the object at different angles need to be obtained, and then all the data are correctly aligned by being converted to a common coordinate system through a certain motion relationship, the process of solving the motion relationship is called point cloud splicing, and finally the three-dimensional reconstruction of the overall profile of the large component is performed. This requires multiple multi-angle scanning to obtain point cloud data of each part, and then point cloud registration and splicing operations are performed to realize the detection of the large workpiece.

[0005] Besel first proposed the Iterative Closest Point (ICP) algorithm in the 1990s, as shown in Figure 4 Before implementing the ICP algorithm, a corresponding target function threshold is set according to the desired calculation result, and then the ICP algorithm is applied for iterative calculation. Each iteration optimizes the current calculation result, and continuously approaches the set threshold of the target function. When the iteration reaches a certain time, the value of the target function is less than the set threshold, and the number of iterations is less than the maximum number of iterations, the iteration ends. However, the algorithm has two problems: first, the ICP algorithm has high requirements for the initial pose of the point cloud, and is suitable for the registration of two point clouds with small initial pose difference. If the initial pose difference between the two point clouds is large, the ICP algorithm will greatly increase the number of iterations, greatly prolonging the time required for the entire registration process. In addition, a large initial pose deviation may cause the algorithm to converge prematurely due to a local optimal solution, greatly reducing the accuracy of the point cloud registration process. Second, the classic ICP algorithm is not suitable for the registration of two point clouds with low coincidence degree. The classic ICP algorithm is suitable for the registration of two point clouds with small initial pose difference or most of the same point clouds. For the registration of two point clouds with low coincidence degree, the algorithm will fall into a local optimal solution and cannot continue to register. SUMMARY

[0006] In view of the problem that in the existing structured light three-dimensional reconstruction algorithm for component registration, a large initial pose deviation may cause the algorithm to converge prematurely due to a local optimal solution, reducing the accuracy of the point cloud registration, the present application provides a structured light measurement-based large-size component three-dimensional reconstruction splicing method.

[0007] The structured light measurement-based large-size component three-dimensional reconstruction splicing method provided by the present application comprises,

[0008] Step 1: A structured light sensor is used to sequentially scan along the width direction of a large three-dimensional component at a preset scanning width, and sequentially labeled point clouds are obtained; the adjacent point clouds have a preset width of overlapping area;

[0009] Step two: preprocessing each point cloud data to obtain the preprocessed point cloud block;

[0010] Step three: sequentially registering the adjacent next point cloud block to the target block, including:

[0011] Spatial pose coarse registration: extracting 2-3 groups of corresponding feature points in the target block and the source point cloud block, the feature points including the start point, end point and curvature abnormal point in the scanning direction; calculating the first spatial transformation parameter of the source point cloud block feature point to the target block feature point coordinate system according to the coordinate values of the corresponding feature points, and then transforming the spatial pose of the source point cloud block to the coordinate system of the target block according to the first spatial transformation parameter to obtain the coordinate-transformed source point cloud block coarsely registered to the target block;

[0012] Coordinate-transformed source point cloud block registration with the target block: segmenting the overlapping region on the coordinate-transformed source point cloud block as the source overlapping region, and segmenting the overlapping region on the target block as the target overlapping region;

[0013] Applying ICP algorithm to the source overlapping region and the target overlapping region for registration to obtain the second spatial transformation parameter, and then transforming the spatial pose of the coordinate-transformed source point cloud block to the coordinate system of the target block according to the second spatial transformation parameter to realize the registration of the coordinate-transformed source point cloud block with the target block; and further realizing the three-dimensional reconstruction splicing of the large three-dimensional component.

[0014] According to the three-dimensional reconstruction splicing method of the large-size component based on the structured light measurement, the first spatial transformation parameter in step three includes a translation vector T and a rotation matrix R; the method for obtaining the first spatial transformation parameter includes:

[0015] A seven-parameter coordinate system conversion model is used to calculate the corresponding relationship of the corresponding feature points in the target block and the source point cloud block:

[0016]

[0017] In the formula, X B is the horizontal coordinate of the feature point in the target block, Y B is the vertical coordinate of the feature point in the target block, Z B is the vertical coordinate of the feature point in the target block, X A is the horizontal coordinate of the feature point in the source point cloud block, Y A is the vertical coordinate of the feature point in the source point cloud block, Z A is the vertical coordinate of the feature point in the source point cloud block; T X is the horizontal coordinate translation vector, T Y is the vertical coordinate translation vector, T Z is the vertical coordinate translation vector, and m is the linear length ratio value of the coordinate system conversion, ω Xis an angle of rotation around the horizontal axis when the source point cloud block coordinate system is rotated to be completely coincident with the target block coordinate system, and ω Y is an angle of rotation around the vertical axis when the source point cloud block coordinate system is rotated to be completely coincident with the target block coordinate system, and ω Z is an angle of rotation around the vertical axis when the source point cloud block coordinate system is rotated to be completely coincident with the target block coordinate system, and ω

[0018] A plurality of corresponding feature points are selected to solve a seven-parameter coordinate system conversion model, to obtain a translation vector T (T X , T Y , T Z ) and a rotation matrix R (ω X , ω Y , ω Z ).

[0019] According to the large-size component three-dimensional reconstruction splicing method based on structured light measurement, in step one, the structured light sensor adopts a direct calibration method to calibrate a light plane.

[0020] According to the large-size component three-dimensional reconstruction splicing method based on structured light measurement, in step one, the preset width range of the overlapping area is 10-15 mm.

[0021] According to the large-size component three-dimensional reconstruction splicing method based on structured light measurement, the preprocessing in step two includes:

[0022] The voxel grid is used to perform downsampling processing on each point cloud data, to obtain a gravity point of all point clouds in each voxel grid;

[0023] The gravity point of all voxel grids is further removed by using a radius filtering method to remove noise gravity points, to obtain a point cloud block after preprocessing.

[0024] According to the large-size component three-dimensional reconstruction splicing method based on structured light measurement, the length of the voxel grid is 0.05-0.15 mm.

[0025] According to the large-size component three-dimensional reconstruction splicing method based on structured light measurement, the method for removing the noise gravity points includes:

[0026] A circle with a radius of 0.02-0.1 mm is drawn with each gravity point as a center, if the number of neighboring gravity points in the circle is less than a preset number of neighboring points threshold, the current gravity point is removed, otherwise, the current gravity point is reserved as a main point cloud, and all main point clouds constitute a point cloud block after preprocessing.

[0027] According to the large-size component three-dimensional reconstruction splicing method based on structured light measurement, the number of neighboring points threshold is 2-10.

[0028] The source overlap area and the target overlap area each contain an overlap area with a preset width, and the width of the source overlap area and the target overlap area is 110-140% of the preset width of the overlap area.

[0029] The method of the present application has the following advantages:

[0030] (1) The specific feature points are quickly extracted, the point clouds are roughly registered, the point cloud posture is adjusted, the time consumption in the ICP registration process is reduced, and the iteration of non-convergence or local optimal solution is avoided.

[0031] (2) The point clouds are accurately segmented by presetting the overlap width during scanning and improving the ICP algorithm, which meets the registration requirements of low-coincidence point clouds, reduces the ICP algorithm running amount, realizes the fast splicing of point clouds, and does not need to scan multiple times for comparison and high-repetition scanning of the component, greatly reducing the operation time in the three-dimensional reconstruction process. BRIEF DESCRIPTION OF DRAWINGS

[0032] Figure 1 is a flowchart of the method of three-dimensional reconstruction splicing of large-size components based on structured light measurement according to the present application;

[0033] Figure 2 is a flowchart of spatial pose rough registration;

[0034] Figure 3 is a flowchart of registration of the source point cloud block and the target block after coordinate transformation;

[0035] Figure 4 is a flowchart of the existing ICP algorithm;

[0036] Figure 5 is a schematic diagram of radius filtering;

[0037] Figure 6 is a scanning schematic diagram of the structural member in the specific embodiment. DETAILED DESCRIPTION

[0038] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0039] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0040] The present application will be further described below in conjunction with the drawings and specific embodiments, but not as a limitation of the present application.

[0041] Specific embodiment one, combined Figures 1 to 4 As shown in the drawings, the present application provides a large-size component three-dimensional reconstruction splicing method based on structured light measurement, comprising,

[0042] Step one: using a structured light sensor to sequentially scan along the width direction of a large three-dimensional component according to a preset scanning width, obtaining sequentially marked point clouds; the adjacent point clouds have an overlapping area with a preset width;

[0043] Step two: pre-processing each piece of point cloud data to obtain a pre-processed point cloud block;

[0044] Step three: sequentially taking the adjacent previous point cloud block as a target block, and taking the adjacent next point cloud block as a source point cloud block to register to the target block, comprising:

[0045] Spatial pose coarse registration: extracting 2-3 groups of corresponding feature points in the target block and the source point cloud block, the feature points including the start point, the end point and the curvature abnormal point in the scanning direction; the start point and the end point can present the edge features of the component; calculating the one-time spatial transformation parameters of the source point cloud block feature points to the coordinate system of the target block feature points according to the coordinate values of the corresponding feature points, and then transforming the spatial pose of the source point cloud block to the coordinate system of the target block according to the one-time spatial transformation parameters, to obtain the coordinate-transformed source point cloud block registered to the target block;

[0046] Coordinate-transformed source point cloud block registration with the target block: dividing the overlapping area on the coordinate-transformed source point cloud block as a source overlapping area, and dividing the overlapping area on the target block as a target overlapping area;

[0047] Applying ICP algorithm to the source overlapping area and the target overlapping area for registration to obtain secondary spatial transformation parameters, and then transforming the spatial pose of the coordinate-transformed source point cloud block to the coordinate system of the target block according to the secondary spatial transformation parameters, to realize the registration of the coordinate-transformed source point cloud block with the target block; and further realize the three-dimensional reconstruction splicing of the large three-dimensional component.

[0048] In the process of the structured light sensor scanning, the direction of the distance between the sensor and the large three-dimensional component is taken as the z-axis, the sensor z-axis height is basically kept fixed, and the height direction (longitudinal direction) of the large three-dimensional component is taken as the y-axis direction for scanning; when a piece of point cloud scanning is completed, the sensor is moved along the x-axis (the component width direction) to return to the initial position of the y-axis, and the scanning is repeated to obtain a second piece of point cloud. After the overall scanning of the component is completed, the point cloud 1, the point cloud 2, …, the point cloud n, etc. are spliced and registered according to the scanning sequence.

[0049] Further, in combination with Figure 2 It is shown that the one-time space transformation parameter in step three includes a translation vector T and a rotation matrix R; the method for obtaining the one-time space transformation parameter includes:

[0050] The seven-parameter coordinate system conversion model is adopted to calculate the corresponding relationship of the corresponding feature points in the target block and the source point cloud block:

[0051]

[0052] In the formula, X B is the horizontal coordinate of the feature point in the target block, Y B is the vertical coordinate of the feature point in the target block, Z B is the vertical coordinate of the feature point in the target block, X A is the horizontal coordinate of the feature point in the source point cloud block, Y A is the vertical coordinate of the feature point in the source point cloud block, Z A is the vertical coordinate of the feature point in the source point cloud block; T X is the horizontal coordinate translation vector, T Y is the vertical coordinate translation vector, T z is the vertical coordinate translation vector, m is the linear length ratio value of the coordinate system conversion, ω X is the angle of rotation of the source point cloud block coordinate system around the horizontal axis when the source point cloud block coordinate system is rotated to completely coincide with the target block coordinate system, ω Y is the angle of rotation of the source point cloud block coordinate system around the vertical axis when the source point cloud block coordinate system is rotated to completely coincide with the target block coordinate system, ω Z is the angle of rotation of the source point cloud block coordinate system around the vertical axis when the source point cloud block coordinate system is rotated to completely coincide with the target block coordinate system.

[0053] A plurality of corresponding feature points are selected to solve the seven-parameter coordinate system conversion model, and the translation vector T (T X , T Y , T Z ) and the rotation matrix R (ω X , ω Y , ω Z ) are obtained.

[0054] In the embodiment, as Figure 2As shown, the efficiency and accuracy of registration are improved by extracting feature points for rough registration before ICP registration. The point cloud registration process is essentially to convert the coordinate system of the source point cloud to be consistent with the coordinate system of the target point cloud through coordinate system conversion. After conversion, in the coordinate system of the target point cloud, the coordinates of the points in the overlapping part of the two adjacent point clouds in the target point cloud coordinate system are the same. Therefore, the corresponding feature points of the two point clouds to be registered can be extracted, and after obtaining the coordinates of the corresponding feature points (a, A; b, B; c, C; …), rough registration is performed through coordinate system conversion to adjust the initial pose of the two point clouds, thereby improving the overall registration efficiency.

[0055] In this embodiment, the seven-parameter coordinate system conversion model is used for point cloud rough registration by introducing the non-collinear feature points. The seven-parameter model is a commonly used model for conversion between two different three-dimensional space rectangular coordinate systems, and the solving process is equivalent to solving a system of equations.

[0056] In the seven-parameter coordinate system conversion model, the scale factor m represents the ratio of the length of the same straight line in the two space rectangular coordinate systems, and realizes the proportional conversion of the scale of the two rectangular coordinate systems. In the reconstruction of the same component, m = 1.

[0057] The extracted feature points are solved to obtain the calculation results of the six unknown parameters. The space pose of the source point cloud is transformed according to the rotation matrix and the translation vector, and the rectangular coordinate system in which the source point cloud is located is converted to the rectangular coordinate system in which the target point cloud is located, thereby realizing rough registration based on feature point extraction. After registration, the coordinates of all points in the overlapping area of the source point cloud block and the target point cloud in the target point cloud coordinate system are basically the same.

[0058] In order to solve the adaptability problem of the current ICP algorithm to low repeatability point clouds, the embodiment adopts an automatic point cloud segmentation method for registration. According to the preset overlap area width, the point cloud segmentation method is applied to segment the overlapping area of the two adjacent point clouds, and two sets of point clouds with a coincidence rate of more than 70% are obtained. Then, ICP registration is performed on the overlapping area, and the spatial transformation parameters R and T are output after registration. The space pose of the source point cloud is transformed based on the spatial transformation parameters R and T, and the registration of the two source point clouds is realized. The schematic diagram is shown in Figure 3 .

[0059] The translation vector T and the rotation matrix R in this embodiment are equivalent to the rigid body transformation matrix H in Figure 4 .

[0060] As an example, in step one, the structured light sensor uses a direct calibration method to calibrate the measurement plane.

[0061] In this embodiment, a structured light scanning reconstruction system can be formed by a structured light sensor and an industrial robot, and based on the obtained point cloud, the topography of large components in an industrial site can be reconstructed in three dimensions.

[0062] As an example, in step one, the preset width range of the overlapping region is 10-15mm.

[0063] Further, the preprocessing in step two includes:

[0064] Each piece of point cloud data is down-sampled using a voxel grid to obtain the center of gravity of all point clouds in each voxel grid;

[0065] The center of gravity of all voxel grids is further processed using a radius filtering method to remove noise center of gravity points to obtain the preprocessed point cloud block.

[0066] Using appropriate down-sampling methods and filtering methods can reduce the density of point clouds, improve the computational efficiency of the subsequent reconstruction process, and reduce the time consumption of the algorithm.

[0067] As an example, the edge length of the voxel grid is 0.05-0.15mm.

[0068] Further, in combination with Figure 5 As shown, the method for removing noise center of gravity points includes:

[0069] A circle with a radius of 0.02-0.1mm is drawn with each center of gravity point as the center. If the number of neighboring center of gravity points in the circle is less than the preset number of neighboring points threshold, the current center of gravity point is removed, otherwise the current center of gravity point is retained as the main point cloud, and all main point clouds constitute the preprocessed point cloud block.

[0070] As an example, the number of neighboring points threshold is 2-10.

[0071] Figure 5 In the method, the number of neighboring points threshold is set to 3. After drawing a circle in the main part of the point cloud, the number of neighboring center of gravity points pts=6, which is greater than 3, so the corresponding center of gravity point is retained. After drawing a circle in the outlier part of the point cloud, the number of neighboring center of gravity points pts=2, which is less than 3, so it is filtered out.

[0072] Further, in the embodiment, the source overlapping region and the target overlapping region both contain an overlapping region with a preset width, and the width of the source overlapping region and the target overlapping region is 110-140% of the width of the preset width overlapping region.

[0073] The width of the source overlapping region and the target overlapping region is set to 110-140% of the width of the preset width overlapping region, so that the repetition rate of the overlapping region is controlled at 70-90%, which can minimize the stitching error and meet the fast registration requirements of the ICP algorithm.

[0074] Repetition rate = preset width / segmentation width = 1 / 1.1-1.4 = 0.7-0.9.

[0075] The width of the source overlap region and the target overlap region is selected in relation to the number of feature point groups, if the number of feature point groups is relatively large, the width of the source overlap region and the target overlap region can be relatively small, and the accuracy of the final registration is not affected, if the number of feature point groups is relatively small, the width of the source overlap region and the target overlap region needs to be relatively wide, so as to ensure the accuracy of the final registration.

[0076] The specific implementation process of the method is as follows:

[0077] I: Before measuring the member profile, the structured light sensor is calibrated first to obtain the world coordinate length corresponding to the unit pixel length, including the plane coordinate calibration in the static state and the speed coupling calibration when moving.

[0078] (1) The direct calibration method is used for the measurement light plane calibration of the sensor. A standard block sample with a specific shape is prepared, and the structured light sensor vertically scans the standard block at different distances to obtain a series of data corresponding to the change of the real length represented by the unit pixel and the image plane longitudinal coordinate. Regression analysis is performed on the calibration results to directly establish the relationship between the real length represented by the unit pixel and the image plane longitudinal coordinate on the structured light scanning light plane.

[0079] (2) The motion measurement system is calibrated. The standard block is fixed, the robot is moved, the sensor scans the forward direction, that is, the y-axis direction of the coordinate system, the moving speed of the robot, that is, the relative scanning speed of the sensor is set as v, after the numerical control machine tool starts to move, the scanning time is t, and the initial position of the sensor along the y-axis is y0 at the initial moment.

[0080] y = vt + y0;

[0081] Wherein the initial position y0, the end point y coordinate and the time t are known, and the image scanning speed corresponding to the robot motion speed can be obtained. Thus, the calibration of the entire measurement system is completed. The three-dimensional measurement system after calibration can output the three-dimensional world coordinates (x, y, z) of any point on the surface of the measured workpiece at any time t.

[0082] II: The robot moves to measure the measured member, and the structured light sensor moves at a uniform speed relative to the measured member. The moving speed is the same as the moving speed used for calibration. Different positions of the large three-dimensional member are scanned, and the overlapping width of different regions is 10-15mm;

[0083] III: The voxel grid algorithm is used for down-sampling processing of the point cloud to reduce the number of point clouds and improve the calculation efficiency. The voxel grid edge length is set to 0.05-0.15mm, the center of gravity of all points in each voxel grid is calculated, which represents all points in the grid, so that all points in each voxel grid are simplified to the center of gravity point;

[0084] 4. Radius filtering is used to filter the centroids of the downsampled point cloud to remove noise points. The algorithm traverses all centroids of the point cloud, draws a circle with each centroid as the center, and sets the radius of the circle, i.e. the neighborhood radius, to be 0.02 to 0.1 mm. A threshold for the number of nearest neighbors (determined according to the characteristics of the point cloud, 2 to 10) is set within this circle. If the number of nearest neighbors does not reach the threshold, it will be regarded as an outlier and removed from the point set. If it reaches the threshold, it will be regarded as the main part of the point cloud and retained.

[0085] 5. The filtered point cloud blocks are registered and stitched together. First, the point cloud features are extracted and coarse registration is performed by spatial coordinate transformation. Then, the improved ICP algorithm is used to re-register the low overlap point cloud after coarse registration. The overlapping area point cloud is segmented, ICP registration is performed, and the main body of the point cloud is registered after obtaining the transformed coordinates.

[0086] Six: Register and stitch all point cloud blocks to output the registered complete 3D model. Specific implementation examples:

[0088] The following specific examples demonstrate the effectiveness of this invention:

[0089] This embodiment uses a structured light sensor to measure a large mesh structure as an example. Figure 6 As shown, the scanning plane size of the structural component is approximately 430*1500mm, and the total number of scans is 1. The equipment used is a Keyence LJ-8200 structured light probe and a KUKAR16 robot, with a reference distance of 270mm, a scanning width of 80mm, and a sampling frequency of 40 frames / second.

[0090] Three-dimensional reconstruction was performed using both the conventional ICP registration method and the method of the present invention. The specific method of the present invention is as follows:

[0091] Step 1: Before measuring the surface profile, calibrate the structured light sensor.

[0092] (1) The sensor's measurement light plane is calibrated using a direct calibration method. A standard block sample of a specific shape is prepared and scanned perpendicular to the structured light sensor at a distance of 250-290 mm. A series of data on the change of the true width represented by a unit pixel and the vertical coordinate of the image plane are obtained. Regression analysis is performed on the calibration results to directly establish the relationship between the true width represented by a unit pixel and the vertical coordinate of the image plane on the structured light scanning light plane.

[0093] (2) Calibrate the motion measurement system. Fix the standard block and move the robot at a speed of 10 mm / s to obtain the system coordinate correspondence at this moving speed.

[0094] Step two: the robot drives the structured light sensor to move at a uniform speed relative to the component to be measured, and the moving speed is 10 mm / s. Different positions of the three-dimensional component are scanned respectively, the overlapping width of different areas is 10 mm, the lateral scanning is from top to bottom, and the total point cloud block number is 6, and the point number in each point cloud is about 63000-72000;

[0095] Step three: the point cloud is down-sampled by using the voxel grid algorithm. The length of the voxel grid is set to 0.1 mm, the gravity center of all points in each voxel grid is calculated, which represents all points in the grid, so that all points in each voxel grid are simplified to the gravity center point, and the point number in the point cloud is reduced by 73.3%;

[0096] Step four: the down-sampled point cloud is filtered by using the radius filtering, the filtering radius is 0.05 mm, and the adjacent point threshold is 5.

[0097] Step five: the filtered point cloud block is registered and spliced. First, the point cloud is feature extracted, and the spatial coordinate transformation is performed for coarse registration, then the improved ICP algorithm is used to register the obtained low-overlapping point cloud, the overlapping area point cloud is segmented, the width is 12 mm, the ICP registration is performed, and the main point cloud is registered after the conversion coordinates are obtained.

[0098] Step six: all point cloud blocks are registered and spliced, and the complete three-dimensional model after registration is output.

[0099] Table 1 is a comparison of registration accuracy and time consumption under different methods. The ICP algorithm is directly used to register the component, the point cloud overlapping degree is low, which does not meet the algorithm requirement, cannot converge, and the registration fails; after the point cloud is directly segmented, the ICP algorithm can be used for registration, but due to the large initial attitude difference of the point cloud, the registration time is long; after the method of the application is used, the registration time is reduced by about 35%, and the accuracy is improved to a certain extent.

[0100] Table 1 comparison of registration results of different methods

[0101]

[0102] It is obvious to those skilled in the art that the present application is not limited to the details of the above exemplary embodiments, and the present application can be implemented in other specific forms without departing from the spirit or essential characteristics of the present application. Therefore, the embodiments should be regarded as exemplary and non-limiting, the scope of the present application is defined by the appended claims rather than the above description, and therefore all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0103] Furthermore, it should be understood that although the specification is described in terms of embodiments, not every embodiment includes every feature or implementation described herein. The specification can include implicit combinations of explicitly mentioned features and / or implicit combinations of implicitly mentioned features. Such combinations are also expressly included within the scope of the specification and an embodiment.

Claims

1. A large-size component three-dimensional reconstruction splicing method based on structured light measurement, characterized by The application relates to a large three-dimensional component three-dimensional reconstruction splicing method based on structured light measurement. Step one: sequentially scanning along the width direction of a large three-dimensional component by using a structured light sensor with a preset scanning width to obtain sequentially marked point clouds; the adjacent point clouds have a preset width of overlapping areas; Step two: preprocessing each piece of point cloud data to obtain a preprocessed point cloud block; Step three: sequentially taking an adjacent previous point cloud block as a target block and taking an adjacent next point cloud block as a source point cloud block to register the target block, which comprises the following steps: Spatial pose coarse registration: extracting 2-3 groups of corresponding feature points in the target block and the source point cloud block, wherein the feature points comprise a starting point, an ending point and a curvature abnormal point in the scanning direction; calculating the first spatial transformation parameter of the source point cloud block feature points to the target block feature point coordinate system according to the coordinate values of the corresponding feature points, and then transforming the spatial pose of the source point cloud block to the coordinate system of the target block according to the first spatial transformation parameter to obtain the coordinate-transformed source point cloud block coarsely registered to the target block; Coordinate-transformed source point cloud block and target block registration: dividing the overlapping area in the coordinate-transformed source point cloud block as a source overlapping area and dividing the overlapping area in the target block as a target overlapping area; Applying the ICP algorithm to the source overlapping area and the target overlapping area to register the two, obtaining the second spatial transformation parameter, and then transforming the spatial pose of the coordinate-transformed source point cloud block to the coordinate system of the target block according to the second spatial transformation parameter to realize the registration of the coordinate-transformed source point cloud block and the target block; and further realizing the three-dimensional reconstruction splicing of the large three-dimensional component.

2. The method according to claim 1, wherein, The first spatial transformation parameter in step three comprises a translation vector T and a rotation matrix R; the method for obtaining the first spatial transformation parameter comprises the following steps: Using a seven-parameter coordinate system conversion model to calculate the corresponding relationship of the corresponding feature points in the target block and the source point cloud block: wherein X B is the horizontal coordinate of the feature point in the target block, Y B is the vertical coordinate of the feature point in the target block, Z B is the vertical coordinate of the feature point in the target block, X A is the horizontal coordinate of the feature point in the source point cloud block, Y A is the vertical coordinate of the feature point in the source point cloud block, Z A is the vertical coordinate of the feature point in the source point cloud block; T X is the horizontal coordinate translation vector, T Y is the vertical coordinate translation vector, T Z is the vertical coordinate translation vector, m is the linear length ratio of coordinate system conversion, ω X is the angle of rotation of the source point cloud block coordinate system around the horizontal axis when rotated to completely coincide with the target block coordinate system, ω Y is the angle of rotation of the source point cloud block coordinate system around the vertical axis when rotated to completely coincide with the target block coordinate system, ω Z is the angle of rotation of the source point cloud block coordinate system around the vertical axis when rotated to completely coincide with the target block coordinate system. A plurality of corresponding feature points are selected to solve a seven-parameter coordinate system conversion model to obtain a translation vector T (T X , T Y , T Z ) and a rotation matrix R (ω X , ω Y , ω Z ).

3. The large-size component three-dimensional reconstruction splicing method based on structured light measurement according to claim 1 or 2, characterized in that, In step one, the structured light sensor is measured by using a direct calibration method to calibrate the light plane.

4. The large-size component three-dimensional reconstruction splicing method based on structured light measurement according to claim 1 or 2, characterized in that, In step one, the preset width of the overlapping area is 10-15 mm.

5. The large-size component three-dimensional reconstruction splicing method based on structured light measurement according to claim 1 or 2, characterized in that, The preprocessing in step two comprises the following steps: Using a voxel grid to downsample each piece of point cloud data to obtain the gravity center point of all the point clouds in each voxel grid; Then, using a radius filtering method to remove the noise gravity center points of all the voxel grids to obtain the preprocessed point cloud block.

6. The method according to claim 5, wherein, The length of the voxel grid is 0.05-0.15 mm.

7. The method according to claim 6, wherein, The method for removing the noise gravity center points comprises the following steps: Respectively taking each gravity center point as the center to draw a circle with a radius of 0.02-0.1 mm, if the number of the adjacent gravity center points in the circle is less than a preset adjacent point number threshold, the current gravity center point is removed, otherwise, the current gravity center point is reserved as a main point cloud, and all the main point clouds constitute the preprocessed point cloud block.

8. The method according to claim 7, wherein, The adjacent point number threshold is 2-10.

9. The large-size component three-dimensional reconstruction splicing method based on structured light measurement according to claim 1, wherein The source overlapping area and the target overlapping area both comprise a preset width of overlapping area, and the width of the source overlapping area and the target overlapping area is 110-140% of the preset width of the overlapping area.

Images