A weak feature point cloud registration method and system based on geometric distance constraints

Abstract

This invention belongs to the field of robot measurement technology and specifically discloses a weak feature point cloud registration method and system based on geometric distance constraints. The method includes: acquiring vehicle body shape point cloud data; converting the vehicle body shape point cloud data to a robot base coordinate system for coarse registration; constructing a robust weak feature point cloud registration algorithm framework; substituting the coarsely registered point cloud data into the robust weak feature point cloud registration algorithm framework for fine registration to output locally complete 3D point cloud data of the vehicle body; substituting the finely registered 3D point cloud data of the vehicle body and vehicle body design model data into the robust weak feature point cloud registration algorithm framework to accurately calculate the vehicle body pose matrix and complete the vehicle body localization. This invention features high point cloud registration accuracy, fast convergence speed, and good robustness.

Description

Technical Field

[0001] This invention belongs to the field of robot measurement technology, and more specifically, relates to a weak feature point cloud registration method and system based on geometric distance constraints. Background Technology

[0002] In robotic machining scenarios, accurate workpiece pose identification (workpiece positioning) can reduce machining errors and improve manufacturing precision. Point cloud registration technology, a commonly used method for workpiece positioning, works by registering the measured point cloud of the workpiece with the 3D CAD design model to the same coordinate system. However, for components with weak geometric features, such as high-speed rail car bodies, turbine blades, and engine rotors, whose geometric features include normal vectors and curvature, the geometric features of the point cloud do not change significantly at different locations. Point cloud registration technology suffers from the following problems: weak feature points lead to unstable solutions for rigid transformation matrices; traditional evaluation metrics such as RMSE and α-recall may register two point clouds with similar features, which does not reflect reality. Therefore, more reasonable evaluation methods and more robust algorithms are needed to process and calculate point clouds of workpieces with weak features.

[0003] The Intermediate Point Collation (ICP) algorithm, a commonly used fine point cloud registration algorithm, traditionally defines distance residuals in three ways: point-to-point, point-to-area, and squared distance. In practice, all three methods suffer from slow convergence, sensitivity to initial values, and susceptibility to noise. To achieve better fine registration, symmetric distance residuals can achieve faster convergence and greater robustness to noise, outliers, and overlapping regions. However, in the registration of weak feature point clouds where most normal vectors are coplanar, the aforementioned residual definitions, due to their small error fields, lead to slow convergence and a tendency to get trapped in local optima. Therefore, weak feature point cloud registration algorithms need to define a distance residual metric with a larger error field.

[0004] To address the aforementioned issues, Chinese patent CN115908517A discloses a low-overlap point cloud registration method. This method, targeting mobile robot environmental perception, combines neural networks to extract feature vectors from the robot's source and target point clouds. These vectors are then input into a feature matching module to obtain a matching matrix for corresponding points. Finally, the matching matrix is ​​input into an attention point fusion module to complete the registration of the robot's large-scale sparse environmental point clouds. Additionally, Chinese patent CN116245921A discloses a weak-feature 3D point cloud registration method incorporating planar constraints. This algorithm employs a coarse-fine registration strategy, finding specific planar structures between two point clouds and combining planar constraints with the ICP algorithm to achieve effective fine registration of the two point clouds. The methods proposed in Chinese patents CN115908517A and CN116245921A are only applicable to their respective working conditions. The method proposed in Chinese patent CN115908517A is only applicable to sparse point clouds and 3D point clouds with obvious features; while the method proposed in Chinese patent CN116245921A has significant limitations, requiring the extraction of planar region features from the source and target point clouds. This method is no longer suitable for registration of cylindrical ruled surface structures such as engine rotors. Therefore, there is currently no suitable solution to the above problems. Summary of the Invention

[0005] To address the aforementioned deficiencies or improvement needs of existing technologies, this invention provides a weak feature point cloud registration method and system based on geometric distance constraints. Combining the characteristics of robot measurement itself and the features of weak feature point cloud registration, a corresponding weak feature point cloud registration method based on geometric distance constraints is designed. By constructing a robust weak feature point cloud registration algorithm framework, the coarsely registered point cloud data is substituted into the framework for fine registration, outputting locally complete 3D point cloud data of the vehicle body. Then, the finely registered 3D point cloud data of the vehicle body and the vehicle body design model data are substituted into the robust weak feature point cloud registration algorithm framework to accurately calculate the vehicle body pose matrix and complete the vehicle body localization. Accordingly, by weighting and fusing the point-to-point distance residuals and point-to-area distance residuals, the algorithm convergence speed is accelerated and slippage is prevented. Simultaneously, by introducing a robust function and adjusting the value of parameter μ to dynamically adjust the number of point pairs used for calculation, the point cloud registration accuracy is improved. Therefore, this invention possesses the characteristics of high point cloud registration accuracy, fast convergence speed, and good robustness.

[0006] To achieve the above objectives, according to one aspect of the present invention, a weak feature point cloud registration method based on geometric distance constraints is proposed, comprising the following steps:

[0007] Step 1: Collect point cloud data of vehicle body shape;

[0008] Step 2: Convert the vehicle body shape point cloud data to the robot base coordinate system for coarse registration of the point cloud data;

[0009] Step 3: Construct a robust registration algorithm framework for weak feature point clouds, and substitute the coarsely registered point cloud data into the robust registration algorithm framework for fine registration to output three-dimensional point cloud data of the vehicle body with local complete shape.

[0010] Step four: Substitute the finely registered 3D point cloud data of the vehicle body and the vehicle body design model data into the weak feature point cloud robust registration algorithm framework to accurately calculate the vehicle body pose matrix and complete the vehicle body localization work.

[0011] As a further preferred option, step one includes the following steps:

[0012] (11) Construct a robot measurement system, which includes a robot and a monocular structured light scanner located at the end of the robot;

[0013] (12) Use a robot measurement system to collect point cloud data of the high-speed train body from multiple poses;

[0014] (13) Denoise the point cloud data of the high-speed rail body to obtain the point cloud data of the body shape.

[0015] As a further preferred option, step two includes the following steps:

[0016] (21) Perform hand-eye calibration on the robot to obtain the robot hand-eye matrix;

[0017] (22) Calculate the homogeneous transformation matrix from the monocular structured light scanner to the base coordinate system by using the robot's hand-eye matrix and kinematic matrix;

[0018] (23) Convert the point cloud data of the vehicle body shape measured multiple times to the robot base coordinate system to complete the coarse registration of the point cloud data.

[0019] As a further preferred embodiment, step three, which involves constructing a robust registration algorithm framework for weak feature point clouds, includes:

[0020] (31) Use weighting coefficients to fuse the distance residuals between point-to-point and point-to-area pairs to construct a weighted distance function;

[0021] (32) Based on the geometric distance constraint of the baseball bat, the weighted distance function is used as the variable to be optimized, and the distance between the source point cloud and the target point cloud and the standard value of the distance between the center of the baseball bat in the two point clouds are used as constraints to construct the optimization function;

[0022] (33) The solution of the optimization function is transformed into a quadratic constrained quadratic programming problem by using the linear perturbation of the homogeneous transformation matrix T, and then the interior point method is used to solve it.

[0023] (34) In the solution process of step (33), a dynamic relaxation constraint algorithm is adopted. By judging the initial distance between the source point cloud and the target point cloud and the standard value of the distance between the two point clouds, the upper and lower bounds of the optimization function constraint are adjusted, so as to ensure the strength of the ball-and-stick geometric distance constraint in real time and ensure that the algorithm can converge correctly, thereby solving the homogeneous transformation matrix between the two point clouds.

[0024] More specifically, in step three, the calculation process for constructing the robust registration algorithm framework for weak feature point clouds is as follows:

[0025] First, input the manually labeled source point cloud data O1 and target point cloud data O2, as well as the standard distance d between the two point cloud data, and set the number of iterations K and k.

[0026] Then, using the homogeneous transformation matrix T obtained in step (22), the nearest point is found iteratively.

[0027] The weights w1 and w2 for point-to-point and point-to-multipoint distance metrics are calculated using a weighted distance function, and an intermediate variable T is defined. temp =I.

[0028] The robust loss function is constructed using the Welsch function, and the parameters A and b are calculated as a quadratic constrained quadratic programming (QCQP) problem.

[0029] Calculate the lower bound of the constraints in the quadratic constrained quadratic programming (QCQP) problem, and solve the QCQP problem to obtain the homogeneous transformation matrix T between the two point clouds. k .

[0030] If the homogeneous transformation matrix T obtained at the Kth order is... k The homogeneous transformation matrix T obtained in the (K+1)th iteration k+1 If the tolerance does not meet the set threshold ε, then return to readjust the parameters of the loss function. If the requirement is met, then let T... temp =T k+1 T temp Determine whether ||T|| is satisfied. temp -I|| F <εor T K+1 =T temp T K If yes, the operation ends; otherwise, the next iteration begins, and the value of parameter μ is adjusted. In one embodiment of the invention, parameter μ is adjusted to μ / 2.

[0031] As a further preferred embodiment, in step (31), the weighted distance function includes:

[0032]

[0033] In the formula, e i Let R be the distance residual, R be the rotation matrix between the two point clouds, and p be the distance residual. i For the source point cloud measurement point, q i Let be the target point cloud measurement point, and t be the translation vector between the two point clouds. This is the normal vector of the corresponding measurement point in the target point cloud;

[0034] Preferably, in step (32), the optimization function includes:

[0035]

[0036]

[0037] In the formula, For p in the k-th iteration i The nearest point; O1 and O2 are artificially marked points in the source point cloud and target point cloud, respectively; δ is the tolerance of the distance constraint, M is the number of corresponding points in the point cloud, and d is the standard value of the distance between the center of the ball and the center of the baseball.

[0038] The specific formula for ρ(r) is:

[0039]

[0040] In the formula: r is the residual, and μ is the parameter of the function.

[0041] As a further preferred embodiment, in step (31), the weight calculation formula is:

[0042]

[0043] In the formula: K is the number of iterations, I max This is the set value.

[0044] As a further preferred option, step (33) specifically includes:

[0045] The optimization function is solved by using a linear perturbation of the homogeneous transformation matrix T:

[0046]

[0047]

[0048] Through mathematical transformation, the optimization function is converted into a quadratic constrained quadratic programming (QCQP) problem, which is then solved using the interior-point method. The specific formula is as follows:

[0049]

[0050]

[0051] In the formula: A0 is a small variable; A1 The correlation coefficient matrix; d lb d ub These are the upper and lower bounds of the transformed constraints.

[0052] As a further preferred embodiment, in step (34), the upper and lower bounds of the adjustment optimization function constraint include:

[0053] Assume ||O1O2||2≤d,

[0054]

[0055]

[0056] In the formula: k is the number of iterations; As a dynamic lower bound, to satisfy small perturbations and geometric distance constraints, the dynamic lower bound should satisfy:

[0057]

[0058] definition

[0059]

[0060] In the formula, K is the number of iterations, τ is a certain integer, and τ is less than the maximum number of iterations.

[0061] According to another aspect of the present invention, a weak feature point cloud registration system based on geometric distance constraints is also provided, comprising:

[0062] The first main control module is used to collect point cloud data of vehicle body shape;

[0063] The second main control module is used to convert the vehicle body topography point cloud data to the robot base coordinate system for coarse registration of the point cloud data.

[0064] The third main control module is used to construct a robust registration algorithm framework for weak feature point clouds. It substitutes the coarsely registered point cloud data into the robust registration algorithm framework for fine registration, so as to output three-dimensional point cloud data of the vehicle body with local complete shape.

[0065] The fourth main control module is used to input the finely registered 3D point cloud data of the vehicle body and the vehicle body design model data into the weak feature point cloud robust registration algorithm framework to accurately calculate the vehicle body pose matrix and complete the vehicle body positioning work.

[0066] As a further preferred embodiment, the first main control module includes a robot measurement system, which includes a robot and a monocular structured light scanner located at the end of the robot.

[0067] In summary, compared with the prior art, the above-described technical solutions conceived by this invention mainly possess the following technical advantages:

[0068] 1. This invention constructs a robust registration algorithm framework for weak feature point clouds. The coarsely registered point cloud data is substituted into this framework for fine registration, outputting locally complete 3D point cloud data of the vehicle body. Then, the finely registered 3D point cloud data and the vehicle body design model data are substituted into the robust registration algorithm framework to accurately calculate the vehicle body pose matrix, completing the vehicle body localization. Correspondingly, by weighting point-to-point and point-to-area pairs, the algorithm convergence speed is accelerated and slippage is prevented. Simultaneously, by introducing a robust function and adjusting the value of parameter μ to dynamically adjust the number of point pairs used in the calculation, the point cloud registration accuracy is improved. Therefore, this invention possesses characteristics such as high point cloud registration accuracy, fast convergence speed, and good robustness.

[0069] 2. This invention combines the advantages of the ICP algorithm and the TDM algorithm by collecting weighted distance metrics, thereby accelerating the convergence speed of the algorithm and preventing slippage during the registration of weak feature point clouds.

[0070] 3. This invention employs the Welsch robust function, which dynamically adjusts the number of point pairs used for calculation by adjusting the value of parameter μ, thereby improving the robustness of point cloud registration.

[0071] 4. By introducing geometric distance constraints between the ball and the stick, this invention can reduce the effective solution set region and improve the convergence speed of the algorithm.

[0072] 5. This invention employs a dynamic relaxation constraint algorithm to adjust the upper and lower bounds of the ball-and-stick constraint in real time to control the strength of the constraint and prevent the algorithm from converging incorrectly. Attached Figure Description

[0073] Figure 1 This is a flowchart of a weak feature point cloud registration method based on geometric distance constraints according to an embodiment of the present invention;

[0074] Figure 2 This is a schematic diagram of the robot measurement system structure according to an embodiment of the present invention;

[0075] Figure 3 This is a flowchart illustrating the positioning process of the high-speed rail body-in-white according to an embodiment of the present invention;

[0076] Figure 4 This is a flowchart illustrating the coarse registration process described in an embodiment of the present invention.

[0077] Figure 5 This is a flowchart of a robust registration algorithm for weak feature point clouds based on geometric distance constraints, as described in an embodiment of the present invention.

[0078] In all the accompanying drawings, the same reference numerals denote the same technical features, specifically: 1-robot, 2-surface structured light scanner. Detailed Implementation

[0079] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0080] like Figure 1 and Figure 3 As shown in the figure, this invention provides a weak feature point cloud registration method based on geometric distance constraints, and the robot measurement system is as follows. Figure 2 As shown, the robot measurement system consists of an AUBO-i10 series collaborative robot and a self-developed monocular structured light scanner. The self-developed monocular structured light scanner includes a DLPLightCrafter 4500 projector with a resolution of 1920×1080 and a Basler-ACA2440-20gm industrial camera. The scanner is rigidly connected to the robot's end effector flange. By adjusting the robot's end effector, the scanner's pose can be controlled, enabling multi-pose measurement.

[0081] like Figure 4 As shown, the vehicle positioning process includes:

[0082] A practical robot equipped with a surface structured light scanner forms a robot measurement system, which collects point cloud data of the high-speed rail body shape multiple times.

[0083] The vehicle body shape point cloud data is transformed into the robot coordinate system through the robot hand-eye matrix and kinematic matrix for coarse registration.

[0084] The weak feature point cloud robust registration algorithm framework proposed in this invention is used for iterative calculation to accurately identify the vehicle body pose.

[0085] More specifically, because the high-speed rail car body is a large-sized, low-feature component, the area that a scanner can measure in a single operation is limited. Therefore, it is necessary to control a robot to collect point cloud data of the high-speed rail car body from multiple poses. After collection, the acquired point cloud needs to be coarsely stitched together. Due to the limitations of the area structured light scanner's ability to collect all point clouds of the object's surface within the scanning area and the precision of the camera equipment, noise will be introduced into the collected point cloud data. Therefore, it is necessary to filter the collected point cloud and remove unwanted topographic point cloud data.

[0086] This paper employs a coarse stitching method based on the robot's base coordinate system, transferring the point clouds from two measurements to this system using the robot's hand-eye matrix and kinematics matrix. Therefore, hand-eye calibration of the robot is necessary. The hand-eye calibration model used in this paper is the "eye on hand" model. Images of the calibration board are captured by the robot in multiple poses, the extrinsic parameters of the scanner camera are solved, and the robot's kinematics matrix corresponding to the camera extrinsic parameters is recorded. A hand-eye calibration equation is then constructed and solved to obtain the hand-eye calibration matrix. The specific steps of the coarse registration are as follows:

[0087] Preliminary work such as filtering and removing noise points is performed on the acquired topographic point cloud data of the high-speed rail body;

[0088] Perform hand-eye calibration on the robot to obtain the robot's hand-eye matrix;

[0089] The homogeneous transformation matrix from the scanner to the base coordinate system is calculated using the robot's hand-eye matrix and kinematic matrix.

[0090] The point cloud data of the vehicle body shape measured multiple times were uniformly converted to the robot's base coordinate system to complete the coarse registration.

[0091] In this embodiment, the coarsely stitched point cloud data is substituted into the robust registration algorithm framework for weak feature point clouds of the present invention to accurately calculate the homogeneous transformation matrix T between different vehicle body point clouds, thereby completing the fine registration of the vehicle body point clouds and obtaining 3D point cloud data of the vehicle body with a locally complete topography. That is, a robust registration algorithm framework for weak feature point clouds is constructed, and the coarsely registered point cloud data is substituted into this framework for fine registration to output 3D point cloud data of the vehicle body with a locally complete topography.

[0092] like Figure 5 As shown, the above steps, including constructing a robust registration algorithm framework for weak feature point clouds, include:

[0093] (31) Use weighting coefficients to fuse the distance residuals between point-to-point and point-to-area pairs to construct a weighted distance function;

[0094] (32) Based on the geometric distance constraint of the ball and stick, the weighted distance function is used as the variable to be optimized, and the distance constraint tolerance between the source point cloud and the target point cloud and the standard value of the distance between the two point clouds is used as the constraint to construct the optimization function;

[0095] (33) The solution of the optimization function is transformed into a quadratic constrained quadratic programming problem by using the linear perturbation of the homogeneous transformation matrix T, and then the interior point method is used to solve it.

[0096] (34) In the solution process of step (33), a dynamic relaxation constraint algorithm is adopted. By judging the initial distance between the source point cloud and the target point cloud and the standard value of the distance between the two point clouds, the upper and lower bounds of the optimization function constraint are adjusted, so as to ensure the strength of the ball-and-stick geometric distance constraint in real time and ensure that the algorithm can converge correctly, thereby solving the homogeneous transformation matrix between the two point clouds.

[0097] Furthermore, since the robust registration algorithm for weak feature point clouds is an iterative operation, after solving the homogeneous transformation matrix between two point clouds, if the difference between the current homogeneous transformation matrix and the previous homogeneous transformation matrix satisfies a set threshold, i.e., ||T k+1 -T k If ||<ε, then let T temP =T k+1 T temp Otherwise, return to step (33). Let T temp =T k+1 T temp Then, determine T. temp -I|| F <εor T K+1 =T temp T K If the condition is met, the homogeneous transformation matrix is ​​output and the operation ends; otherwise, the parameter μ is adjusted and the nearest point is recalculated until the output homogeneous transformation matrix meets the threshold requirement.

[0098] The precisely registered 3D point cloud data of the vehicle body and the vehicle body design model data are substituted into the robust registration algorithm framework for weak feature point clouds to accurately calculate the vehicle body pose matrix and complete the vehicle body localization. More specifically, the measured 3D point cloud of the vehicle body is registered with the design CAD model. Similarly, it is substituted into the robust registration algorithm framework for weak feature point clouds of this invention to accurately calculate the vehicle body pose matrix T and complete the vehicle body localization.

[0099] In the above steps, since both the fine registration and the registration of the vehicle body 3D point cloud with the design CAD model use the weak feature point cloud robust registration algorithm, the specific calculation process of this algorithm will be introduced in the following text. The algorithm calculation process is as follows: Figure 4 As shown.

[0100] (1) To better introduce the subsequent process of the algorithm, this paper first defines the relevant mathematical symbols of the algorithm:

[0101] Representation of special orthogonal group SO(3) and special Euclidean group SE(3):

[0102] Lie groups and Lie algebras are tools for representing rigid body transformations, and (·)^ represents the mapping process from Lie algebras to Lie groups.

[0103]

[0104]

[0105] Record Point cloud 1 is P = {p1, ..., p} M Point cloud 2 is Q = {q1, ..., q} N The traditional ICP algorithm defines the distance residual in three ways:

[0106] 1) Point-to-point:

[0107]

[0108] 2) Point at the opposite side:

[0109]

[0110] 3) Symmetrical distance:

[0111]

[0112] Among the three distance residual definition methods mentioned above, point-to-point and point-to-area definitions have drawbacks such as low convergence speed, sensitivity to initial values, and susceptibility to noise. The symmetric distance definition method, compared to the other two, has faster convergence speed and is more robust to noise, outliers, and partial overlap. However, in the registration of weak feature point clouds with mostly coplanar normal vectors, the residual definition methods related to the normal vectors all have small error fields, and registration is prone to slippage, resulting in a low convergence rate and getting trapped in local optima.

[0113] To address the aforementioned issues, the algorithm proposed in this paper defines the distance residual as a weighted distance function:

[0114]

[0115] The optimization objective function is as follows:

[0116]

[0117] In the formula: r k Let θ be the distance residual at the k-th iteration, and Θ be the variable to be optimized.

[0118] (2) In order to better register the two point clouds, the geometric distance constraint of the ball and stick is introduced, and the optimization problem is constructed by combining equation (6) and equation (7).

[0119]

[0120]

[0121] In the formula: For p in the k-th iterationi The closest point; O1 and O2 are artificially marked points of the source point cloud and the target point cloud, respectively, which are the spherical point clouds at both ends of the bat AB; δ is the tolerance of the distance constraint.

[0122] The specific formula for ρ(r) is:

[0123]

[0124] In the formula: r is the residual, and μ is the parameter of the function.

[0125] To facilitate the solution of the problem, a linear perturbation of the homogeneous transformation matrix T is used to solve equation (8).

[0126]

[0127]

[0128] Through mathematical transformation, equation (9) can be converted into a quadratic constrained quadratic programming problem (QCQP), which is then solved using the interior point method. The specific formula is as follows:

[0129]

[0130] In the formula: A0 is a small variable; A1 The correlation coefficient matrix; d lb d ub These are the upper and lower bounds of the transformed constraints.

[0131] (3) In the process of solving equation (10), the added linear perturbation can be regarded as a first-order Taylor approximation. In order to satisfy the effectiveness of the first-order Taylor approximation, it is necessary to ensure that the perturbation ξ is small. However, when the initial value ||O1-O2||2 deviates from the standard value d, the perturbation ξ obtained by solving the QCQP problem may be large and thus fail to meet the geometric distance constraint of the soccer ball, resulting in the algorithm failing to converge correctly.

[0132] Therefore, the algorithm of this invention employs a dynamic constraint relaxation algorithm during the solution process, thereby ensuring the strength of the geometric distance constraint between the ball and the bat in real time, and guaranteeing that the algorithm can converge correctly. For example... Figure 2 As shown, the dynamic relaxation constraint algorithm can effectively guarantee the registration accuracy of the algorithm. The dynamic relaxation constraint algorithm adjusts the upper and lower bounds of the constraint in equation (10) by judging the initial ||O1-O2||2 and d. The specific process is as follows:

[0133] This paper assumes that ||O1O2||2≤d

[0134]

[0135] In the formula: k is the number of iterations; This serves as a dynamic lower bound. To satisfy small perturbations and geometric distance constraints, the dynamic lower bound should meet the following requirements:

[0136]

[0137] definition

[0138]

[0139] (4) Definition of parameter μ and weights w1 and w2

[0140] The parameter μ affects the robustness of the objective function in equation (7). Experiments have shown that a smaller μ assigns lower weights to noise and outliers, thus reducing their impact on the optimization results. Therefore, in the early stages of algorithm iteration, using a larger μ can utilize more nearest point pairs, preventing the algorithm from getting trapped in local optima. As iteration progresses, the value of μ needs to be gradually reduced to weaken the influence of noise and outliers. Weight calculation formula:

[0141]

[0142] In the formula: K is the number of iterations, I max This is the set value.

[0143] Parameters w1 and w2 determine the contribution of point-to-point and point-to-area distance residual measurement methods to the residuals. In the early stages of iteration, a smaller w1 value is assigned to adopt a point-to-area distance residual measurement method to accelerate the iteration speed of the algorithm. As the iteration progresses, w1 is gradually increased to adopt a point-to-point distance residual measurement method to prevent the algorithm from getting stuck in a low convergence rate and from slippage after registration.

[0144] Furthermore, according to another aspect of the present invention, a weak feature point cloud registration system based on geometric distance constraints is also provided, comprising:

[0145] The first main control module is used to collect point cloud data of vehicle body shape; wherein, the first main control module includes a robot measurement system, which includes a robot and a monocular structured light scanner located at the end of the robot.

[0146] The second main control module is used to convert the vehicle body topography point cloud data to the robot base coordinate system for coarse registration of the point cloud data.

[0147] The third main control module is used to construct a robust registration algorithm framework for weak feature point clouds. It substitutes the coarsely registered point cloud data into the robust registration algorithm framework for fine registration, so as to output three-dimensional point cloud data of the vehicle body with local complete shape.

[0148] The fourth main control module is used to input the finely registered 3D point cloud data of the vehicle body and the vehicle body design model data into the weak feature point cloud robust registration algorithm framework to accurately calculate the vehicle body pose matrix and complete the vehicle body positioning work.

[0149] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A weak feature point cloud registration method based on geometric distance constraints, characterized in that, Includes the following steps: Step 1: Collect point cloud data of vehicle body shape; Step 2: Convert the vehicle body shape point cloud data to the robot base coordinate system for coarse registration of the point cloud data; Step 3: Construct a robust registration algorithm framework for weak feature point clouds, and substitute the coarsely registered point cloud data into the robust registration algorithm framework for fine registration to output three-dimensional point cloud data of the vehicle body with local complete shape. The framework for constructing a robust registration algorithm for weak feature point clouds includes: (31) Use weighting coefficients to fuse the distance residuals between point-to-point and point-to-area pairs to construct a weighted distance function; (32) Based on the geometric distance constraint of the ball and bat, the weighted distance function is used as the variable to be optimized, and the distance between the source point cloud and the target point cloud and the standard value of the distance between the ball and bat point cloud in the two point clouds are used as constraints to construct the optimization function; In step (32), the optimization function includes: ， In the formula, During the k-th iteration The closest point; Artificially marked points for the source and target point clouds; The tolerance for distance constraints is M, where M is the number of points in the point cloud. d This refers to the standard distance between the center of the baseball and the center of the ball. For source point cloud measurement points, For the target point cloud measurement points, Let be the normal vector of the corresponding measurement point in the target point cloud, and T be the homogeneous transformation matrix. and These are the weighting coefficients; in, The specific formula is: ， In the formula: r For residuals, μ for The parameters of the function; (33) The solution of the optimization function is transformed into a quadratic constrained quadratic programming problem by using the linear perturbation of the homogeneous transformation matrix T, and then the interior point method is used to solve it; The optimization function is solved by using a linear perturbation of the homogeneous transformation matrix T: ， Through mathematical transformation, The problem is transformed into a quadratic constraint quadratic programming problem, which is then solved using the interior point method. The specific formula is as follows: ， In the formula: For small variables; This is the correlation coefficient matrix; These are the upper and lower bounds of the transformed constraints; (34) In the solution process of step (33), a dynamic relaxation constraint algorithm is adopted. By judging the initial distance between the source point cloud and the target point cloud and the standard value of the distance between the two point clouds, the upper and lower bounds of the optimization function constraint are dynamically adjusted, thereby solving the homogeneous transformation matrix between the two point clouds. Step four: Substitute the finely registered 3D point cloud data of the vehicle body and the vehicle body design model data into the weak feature point cloud robust registration algorithm framework to accurately calculate the vehicle body pose matrix and complete the vehicle body localization work.

2. The weak feature point cloud registration method based on geometric distance constraints according to claim 1, characterized in that, Step one includes the following steps: (11) Construct a robot measurement system, which includes a robot and a monocular structured light scanner located at the end of the robot; (12) Use a robot measurement system to collect point cloud data of the high-speed rail body from multiple poses; (13) Denoise the point cloud data of the high-speed rail body to obtain the point cloud data of the body shape.

3. The weak feature point cloud registration method based on geometric distance constraints according to claim 2, characterized in that, Step two includes the following steps: (21) Perform hand-eye calibration on the robot to obtain the robot's hand-eye matrix; (22) Calculate the homogeneous transformation matrix from the monocular structured light scanner to the base coordinate system by using the robot's hand-eye matrix and kinematic matrix; (23) Convert the point cloud data of the vehicle body shape measured multiple times to the robot base coordinate system to complete the coarse registration of the point cloud data.

4. The weak feature point cloud registration method based on geometric distance constraints according to claim 3, characterized in that, In step (31), the weighted distance function includes: ， In the formula, Let R be the distance residual, and R be the rotation matrix between the two point clouds. For source point cloud measurement points, For the target point cloud measurement points, Let be the translation vectors of the two point clouds. Let be the normal vector of the corresponding measurement point in the target point cloud. and These are the weighting coefficients.

5. The weak feature point cloud registration method based on geometric distance constraints according to claim 4, characterized in that, In step (31), the weight calculation formula includes: ， In the formula: K For the number of iterations, I max This is the set value.

6. The weak feature point cloud registration method based on geometric distance constraints according to claim 5, characterized in that, In step (34), the upper and lower bounds of the adjustment optimization function constraints include: Assumption , ， In the formula: k is the number of iterations; As a dynamic lower bound, to satisfy small perturbations and geometric distance constraints, the dynamic lower bound should satisfy: ， definition : ， In the formula, K is the number of iterations. To determine the integer, and Less than the maximum number of iterations.

7. A weak feature point cloud registration system based on geometric distance constraints, characterized in that, include: The first main control module is used to collect point cloud data of vehicle body shape; The second main control module is used to convert the vehicle body topography point cloud data to the robot base coordinate system for coarse registration of the point cloud data. The third main control module is used to construct a robust registration algorithm framework for weak feature point clouds. It substitutes the coarsely registered point cloud data into the robust registration algorithm framework for fine registration, so as to output three-dimensional point cloud data of the vehicle body with local complete shape. The framework for constructing a robust registration algorithm for weak feature point clouds includes: ， In the formula, During the k-th iteration The closest point; Artificially marked points for the source and target point clouds; The tolerance for distance constraints is M, where M is the number of points in the point cloud. d This refers to the standard distance between the center of the baseball and the center of the ball. For source point cloud measurement points, For the target point cloud measurement points, Let be the normal vector of the corresponding measurement point in the target point cloud, and T be the homogeneous transformation matrix. and These are the weighting coefficients; in, The specific formula is: ， In the formula: r For residuals, μ for The parameters of the function; (33) The solution of the optimization function is transformed into a quadratic constrained quadratic programming problem by using the linear perturbation of the homogeneous transformation matrix T, and then the interior point method is used to solve it; The optimization function is solved by using a linear perturbation of the homogeneous transformation matrix T: ， Through mathematical transformation, The problem is transformed into a quadratic constraint quadratic programming problem, which is then solved using the interior point method. The specific formula is as follows: ， In the formula: For small variables; This is the correlation coefficient matrix; These are the upper and lower bounds of the transformed constraints; (34) In the solution process of step (33), a dynamic relaxation constraint algorithm is adopted. By judging the initial distance between the source point cloud and the target point cloud and the standard value of the distance between the two point clouds, the upper and lower bounds of the optimization function constraint are dynamically adjusted, thereby solving the homogeneous transformation matrix between the two point clouds. The fourth main control module is used to input the finely registered 3D point cloud data of the vehicle body and the vehicle body design model data into the weak feature point cloud robust registration algorithm framework to accurately calculate the vehicle body pose matrix and complete the vehicle body positioning work.

8. The weak feature point cloud registration system based on geometric distance constraints according to claim 7, characterized in that, The first main control module includes a robot measurement system, which includes a robot and a monocular structured light scanner located at the end of the robot.

Images