A tool point cloud precise registration method and system based on topographic closed constraint

Abstract

A tool point cloud fine registration method and system based on topographic closed constraint, relate to the field of point cloud registration. The existing three-dimensional point cloud model of the industrial tool is only obtained through the rotation relationship, and the quality of the tool coarse registration three-dimensional point cloud is still poor, and further fine registration is needed, and other problems are solved, the method is characterized in that the nearest point distance between the registration point clouds is calculated first, and the coincidence area is screened in combination with the differentiated threshold. Subsequently, the segmented farthest point sampling strategy is adopted to perform equalization processing on the source point cloud, so as to highlight the effective point cloud features. On this basis, the global topographic closed constraint is introduced into the registration optimization framework, and a parameter step-by-step optimization strategy is designed, which effectively alleviates the error accumulation problem of the algorithm. The application is also suitable for application fields such as industrial detection, reverse engineering and equipment state monitoring.

Description

Technical Field

[0001] This invention relates to the field of point cloud registration technology, specifically to a method and system for fine registration of tool point clouds based on topographic closure constraints. Background Technology

[0002] With the rapid development of intelligent manufacturing and digital twin technologies, high-precision 3D reconstruction has become a core technology in industrial inspection, reverse engineering, and equipment condition monitoring. Industrial cutting tools, as key components in cutting processes, directly determine machining quality and lifespan assessment results based on their geometric accuracy. Therefore, acquiring high-fidelity 3D point cloud models of cutting tools has significant engineering value for achieving automated dimensional inspection, wear identification, and lifespan prediction.

[0003] In current industrial applications, 3D scanners are commonly used in conjunction with rotating fixtures to acquire 3D point clouds of cutting tools from multiple perspectives. A complete 3D model of the cutting tool is then obtained through point cloud registration. Coarse registration of the tool point cloud can be achieved using the rotational relationships provided by the rotating fixture. However, due to interference factors such as fitting errors of the rotation axis, positioning errors of the rotating fixture, and system noise, layering still exists between the tool point clouds after coarse registration, making it difficult to guarantee the accuracy of subsequent tool parameter measurements. Therefore, after coarse registration, the tool point cloud still requires finer registration to obtain a perfect 3D point cloud model.

[0004] Common industrial cutting tools consist of a cutting section and a tool holder. The cutting section has a complex structure but exhibits helical repetition, while the tool holder has a simple structure but occupies a large area. These characteristics present new challenges to commonly used point cloud registration schemes. Due to the limited overlap between adjacent tool point clouds, the classic iterative nearest-point algorithm is prone to getting stuck in local convergence. Introducing feature descriptors such as FPFH and SHOT, or deep learning methods such as PointNetLK and DCP, results in helical misalignment of adjacent tool point clouds during registration.

[0005] To address the aforementioned issues, existing technologies have disclosed methods that utilize topological or geometric prior information to assist registration. For thin-walled parts, existing technologies propose a registration method with multiple positioning and assembly constraints, setting registration weights according to the assembly requirements of different regions, significantly improving the measurement accuracy of key assembly areas. Existing technologies introduce periodic symmetry constraints in gear point cloud reconstruction, improving splicing accuracy by optimizing tooth profile consistency. Existing technologies embed tool axisymmetry as a soft constraint into the ICP objective function, achieving good results in milling cutter reconstruction. However, these methods largely rely on strict symmetry, periodicity, and other specific geometric assumptions, making them difficult to generalize to irregular tools such as irregularly shaped turning tools and drills.

[0006] Furthermore, the aforementioned methods typically decompose multi-view registration into independent pairwise registration optimization problems, neglecting the inherent physical constraint of global closure. This leads to the continuous propagation of registration errors, resulting in significant layering between the first and last point clouds. Considering the special constraints brought about by tool morphology closure, existing technologies propose a multi-view registration framework based on loop closure detection. This framework applies a global consistency constraint when a closed loop is formed during scanning, effectively suppressing error accumulation. However, the registration process is still prone to getting stuck in local convergence. Therefore, the 3D point cloud models of existing industrial tools are usually obtained by multi-view scanning and registration algorithms using a 3D scanner and a rotating fixture. The quality of the coarsely registered 3D point cloud of the tool obtained solely through rotational relationships is still poor, requiring further fine registration. Summary of the Invention

[0007] This invention addresses the problem that existing technologies for obtaining 3D point cloud models of industrial cutting tools typically rely on multi-view scanning and registration algorithms using a 3D scanner and rotating fixture. The resulting coarsely registered 3D point cloud model obtained solely through rotational relationships still suffers from poor quality, necessitating further fine registration. To resolve this, this invention proposes a method and system for fine registration of cutting tool point clouds based on topographic closure constraints. The invention achieves these technical problems through the following technical solutions: Option 1: This invention proposes a method for precise registration of tool point clouds based on topographic closure constraints. The method includes the following steps: Step 1: Statistical filtering is used to remove outliers from the original point cloud, and a double threshold strategy is used to filter the overlapping areas of point clouds from adjacent viewpoints. Segmented sampling of the farthest point is performed along the tool axis to obtain a balanced set of feature points. Step 2: Based on the balanced feature point set obtained in Step 1, an alternating iterative strategy of matching point pair search and transformation parameter solution is adopted to complete the pairwise registration of point clouds from adjacent viewpoints and obtain the initial transformation matrix set. Step 3: Based on the initial transformation matrix set obtained in Step 2, construct a multi-view global topography closure constraint, and solve the global optimal transformation matrix by optimizing the rotation parameters individually, optimizing the translation parameters individually, optimizing the overall parameters jointly, and using the LM algorithm.

[0008] Furthermore, a preferred embodiment is provided, wherein the method for filtering overlapping regions of point clouds from adjacent viewpoints using a dual-threshold strategy in step 1 is as follows: Based on the coarse registration results provided by the rotating fixture, the nearest point distance between the source point cloud and the target point cloud is calculated, and two nearest point distance thresholds are set. and , for 6-10 times.

[0009] Furthermore, a preferred embodiment is provided, wherein the method for performing segmented sampling of the farthest point along the tool axis in step 1 to obtain a uniformly distributed set of feature points is as follows: The source point cloud is evenly divided into 10 continuous segments along the tool axis; the farthest point sampling is performed on each segment, and 1000 feature points are collected for each segment. If the number of point clouds in a segment is less than 1000, all points are retained; all segment sampling points are merged to obtain the equalized source point cloud feature set.

[0010] Furthermore, a preferred embodiment is provided, wherein step 2 employs an alternating iterative strategy of matching point pair search and transformation parameter solution to complete pairwise registration of adjacent viewpoint point clouds and obtain the initial transformation matrix set as follows: Step 2.1: First, take the nearest point in the target point cloud after spatial transformation of the source point cloud as the corresponding matching point; The source point cloud is transformed to the target point cloud coordinate system using the current transformation matrix. KD-Tree is then used to quickly search for the nearest point of each source point, forming a set of matching point pairs. Step 2.2: Based on the matching points determined in Step 2.1, the optimal transformation matrix is ​​solved using the least squares iterative nearest point algorithm. R , t and residual function L ; Step 2.3: Repeat steps 2.1 to 2.2 until the residual change is reached. Output single-view registration transformation matrix T ij , where i and j are the adjacent viewpoint numbers.

[0011] Furthermore, a preferred embodiment is provided, wherein the least squares iterative nearest point algorithm described in step 2.2 is as follows: Step 2.2.1: Determine the optimal rotation matrix. , It is used to remove the centroid from point clouds; Step 2.2.2: Compute the covariance matrix H of the centroid-removed points. ; Step 2.2.3: Perform singular value decomposition on the covariance matrix H. Then the optimal transformation for: , .

[0012] Furthermore, a preferred embodiment is provided, wherein the method for constructing multi-view global topography closure constraints in step 3 is as follows: Suppose that a total of n point clouds are collected from different viewpoints, and the transformation matrix between adjacent viewpoints is T. 12 T 23 ... T (n -1)n Then the global closure constraint is ,in This is the transformation matrix from the first viewpoint to the nth viewpoint.

[0013] Furthermore, a preferred embodiment is provided, wherein the method in step 3 for solving the globally optimal transformation matrix through individual optimization of rotation parameters, individual optimization of translation parameters, joint optimization of overall parameters, and the LM algorithm is as follows: Rotation parameters optimized separately: all translation vectors are fixed, and only the rotation matrices R1, R2, ..., R are optimized. (n-1) Construct the rotational residual function L R The LM algorithm is used to solve for the optimal set of rotation matrices; Translation parameters are optimized separately: the optimal rotation matrix is ​​fixed, and only the translation vectors t1, t2, ..., t are optimized. (n-1) Construct the translation residual function L G The LM algorithm is used to solve for the optimal set of translation vectors; Overall parameter joint optimization: Using the above results as initial values, all rotation matrices and translation vectors are optimized simultaneously.

[0014] Option 2: A tool point cloud fine registration system based on topographic closure constraints, the system comprising: The balanced feature point set acquisition module is used to remove outliers from the original point cloud using statistical filtering, and to filter overlapping areas of point clouds from adjacent viewpoints using a dual threshold strategy. It also performs segmented sampling of the farthest point along the tool axis to obtain a balanced feature point set. The initial transformation matrix module is used to complete the pairwise registration of point clouds from adjacent viewpoints based on the distributed balanced feature point set obtained by the distributed balanced feature point set acquisition module, using an alternating iterative strategy of matching point pair search and transformation parameter solution, to obtain the initial transformation matrix set. The optimization module is used to construct multi-view global topography closure constraints based on the initial transformation matrix set obtained from the initial transformation matrix module. It solves the global optimal transformation matrix through individual optimization of rotation parameters, individual optimization of translation parameters, joint optimization of global parameters, and the LM algorithm.

[0015] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0016] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0017] The advantages of this invention are: This invention discloses a tool point cloud fine registration method and system based on topographic closure constraints. It calculates the nearest point distance between registered point clouds and uses a differential threshold to filter overlapping regions. Furthermore, it employs a segmented farthest point sampling strategy to equalize the source point cloud, highlighting effective point cloud features. Building upon this, global topographic closure constraints are introduced into the registration optimization framework, and a step-by-step parameter optimization strategy is designed to effectively alleviate the error accumulation problem in the algorithm.

[0018] This invention is also applicable to industrial testing, reverse engineering, and equipment condition monitoring. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the tool point cloud fine registration method based on topographic closure constraints described in Implementation Method 1.

[0020] Figure 2 This is a schematic diagram of the SVD-ICP algorithm described in Implementation Method 1.

[0021] Figure 3 This is a schematic diagram of the parameter step-by-step optimization algorithm described in Implementation Method 1. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0023] Implementation Method 1, see [link] Figures 1 to 3 This embodiment proposes a method and system for fine registration of tool point clouds based on topographic closure constraints. It includes three core steps: tool point cloud preprocessing, iterative search of matching point pairs and single-view geometric iterative optimization, and step-by-step parameter optimization based on global topographic closure constraints. The details are as follows: 1. Tool point cloud preprocessing, which mainly includes two steps: registration point cloud screening and segmented farthest point sampling, aims to ensure both the accuracy and computational efficiency of subsequent registration. First, statistical filtering is used to remove outliers from the original point cloud. Then, a dual threshold strategy is used to screen overlapping regions of point clouds from adjacent viewpoints. Finally, segmented farthest point sampling is performed along the tool axis to obtain a balanced set of feature points.

[0024] Dual-threshold overlapping region filtering: Based on the coarse registration results provided by the rotating fixture, the nearest point distance between the source point cloud and the target point cloud is calculated. This ensures that the target point cloud retains dense points to guarantee the reliability of the matching benchmark, while the source point cloud only retains its nearest neighbors to reduce computational complexity. Therefore, two nearest point distance thresholds are set. and , Should be 6-10 times.

[0025] Axial segmentation equalization downsampling: The source point cloud is evenly divided into 10 continuous segments along the tool axis; farthest point sampling (FPS) is performed on each segment, and 1000 feature points are collected for each segment. If the number of point clouds in a segment is less than 1000, all points are retained; all segment sampling points are merged to obtain the equalized source point cloud feature set.

[0026] 2. Iterative search for matching point pairs and iterative optimization of single-view geometry An alternating iterative strategy of "matching point pair search - transformation parameter solution" is adopted to complete the pairwise registration of point clouds from adjacent viewpoints and obtain the initial transformation matrix set.

[0027] Initialization: Use the coarse registration transformation matrix provided by the rotating fixture as the initial pose, and set the convergence threshold. .

[0028] Matching point pair search: Transform the source point cloud to the target point cloud coordinate system using the current transformation matrix, and use KD-Tree to quickly search for the nearest point of each source point to form a set of matching point pairs.

[0029] SVD-ICP Transform Solution: Remove the centroids from the matching point pairs and calculate the covariance matrix. H And perform singular value decomposition. H=U ΣV Solving for the optimal rigid body transformation ( R, t and residual function L .

[0030] Iterative convergence: Repeat the above process until the change in residual is less than Output single-view registration transformation matrix T ij (i and j are the adjacent viewpoint numbers).

[0031] 3. Stepwise parameter optimization based on global topography closure constraints A multi-view global topography closure constraint is constructed. The global optimal transformation matrix is ​​solved by a three-step strategy of optimizing rotation parameters individually, optimizing translation parameters individually, and jointly optimizing global parameters. The LM (Levenberg-Marquardt) algorithm is used.

[0032] Global closure constraint construction: Assume common acquisition n Point clouds from various viewpoints, with adjacent viewpoint transformation matrices as follows: T 12 , T 23 … T(n-1)n Then the global closure constraint is ,in From the first perspective to the second n Transformation matrix for each viewpoint.

[0033] Rotation parameters optimized separately: all translation vectors are fixed, and only the rotation matrix is ​​optimized. R 1. R 2、…、 R (n-1) Constructing the rotational residual function L R The LM algorithm is used to solve for the optimal set of rotation matrices.

[0034] Translation parameters are optimized separately: the optimal rotation matrix is ​​fixed, and only the translation vector is optimized. t 1. t 2、…、 t (n-1) Construct the translation residual function L G The LM algorithm is used to solve for the optimal set of translation vectors.

[0035] Overall parameter joint optimization: Using the above results as initial values, all rotation matrices and translation vectors are optimized simultaneously.

[0036] The SVD-ICP algorithm is as follows.

[0037] Assuming there are n groups of points with the same name, the source point cloud and the target point cloud are represented as follows: , Then the residual function L as follows,

[0038] The goal of SVD-ICP is to find rigid body transformations ( R , t This minimizes the residual function L of the current matched point pair. First, the optimal rotation matrix is ​​determined by removing the centroid from the point cloud to eliminate the influence of translation parameters. Specifically, as follows: , Then, the covariance matrix H of the points after centroid removal is calculated using the following formula:

[0039] covariance matrix H It depicts some clouds. What linear transformation can be used to map to a point cloud? And rotation matrix R This is precisely the optimal part of the linear transformation that changes only the direction and not the scale. For the covariance matrix... H Perform singular value decomposition Then the optimal transformation As shown in the following formula.

[0040]

[0041]

[0042] This algorithm provides the optimal spatial transformation relationship for single point cloud registration.

[0043] Clearly, this algorithm cannot be used for the point clouds at both ends due to geometric closed-loop constraints. Furthermore, error accumulation will occur between the beginning and end points. Therefore, the registration process still needs to introduce global closed-loop constraints for unified optimization. The global closed-loop constraint is as follows.

[0044]

[0045] Step 3 also includes a pixel-by-pixel calibration process, the specific method of which is as follows: The selected matching point pairs are decentralized to eliminate the influence of translation parameters. The optimal rotation matrix is ​​calculated using the covariance matrix of the decentralized point cloud. Finally, the optimal rotation matrix satisfying the geometric closed-loop constraint is obtained using the LM optimization algorithm. At this point, the residual function... L R as follows.

[0046]

[0047] Calculate the translation vectors corresponding to the globally optimal rotation matrix. Use the LM optimization algorithm to find the optimal translation vector that satisfies the geometric closed-loop constraint. At this point, the residual function remains... L G However, the optimization parameters only include translation parameters.

[0048]

[0049] Using the above calculation results as initial values, the globally optimal rotation matrix and translation vector satisfying the geometric closed-loop constraint are obtained through the LM optimization algorithm. At this point, the residual function remains... L G .

[0050] In summary, the method described in this embodiment first calculates the nearest point distance between registered point clouds and then filters overlapping regions using a differential threshold. Subsequently, a segmented farthest point sampling strategy is employed to equalize the source point cloud, highlighting effective point cloud features. Based on this, global topographic closure constraints are introduced into the registration optimization framework, and a step-by-step parameter optimization strategy is designed to effectively alleviate the error accumulation problem of the algorithm.

[0051] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or claims of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0052] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention. Clearly, those skilled in the art can make various alterations and modifications to the invention without departing from its spirit and scope. Thus, if these modifications and modifications of the invention fall within the scope of the claims and their equivalents, the invention is also intended to include these modifications and modifications.

Claims

1. A method for precise registration of tool point clouds based on topographic closure constraints, characterized in that, The method includes the following steps: Step 1: Statistical filtering is used to remove outliers from the original point cloud, and a double threshold strategy is used to filter the overlapping areas of point clouds from adjacent viewpoints. Segmented sampling of the farthest point is performed along the tool axis to obtain a balanced set of feature points. Step 2: Based on the balanced feature point set obtained in Step 1, an alternating iterative strategy of matching point pair search and transformation parameter solution is adopted to complete the pairwise registration of point clouds from adjacent viewpoints and obtain the initial transformation matrix set. Step 3: Based on the initial transformation matrix set obtained in Step 2, construct a multi-view global topography closure constraint, and solve the global optimal transformation matrix by optimizing the rotation parameters individually, optimizing the translation parameters individually, optimizing the overall parameters jointly, and using the LM algorithm.

2. The tool point cloud fine registration method based on topographic closure constraints according to claim 1, characterized in that, The method for filtering overlapping regions of point clouds from adjacent viewpoints using a dual-threshold strategy in step 1 is as follows: Based on the coarse registration results provided by the rotating fixture, the nearest point distance between the source point cloud and the target point cloud is calculated, and two nearest point distance thresholds are set. and , for 6-10 times.

3. The tool point cloud fine registration method based on topographic closure constraints according to claim 1, characterized in that, The method for obtaining a uniformly distributed set of feature points by performing segmented sampling along the tool axis in step 1 is as follows: The source point cloud is evenly divided into 10 continuous segments along the tool axis; the farthest point sampling is performed on each segment, and 1000 feature points are collected for each segment. If the number of point clouds in a segment is less than 1000, all points are retained; all segment sampling points are merged to obtain the equalized source point cloud feature set.

4. The tool point cloud fine registration method based on topographic closure constraints according to claim 1, characterized in that, Step 2 employs an alternating iterative strategy of matching point pairs and solving for transformation parameters to complete pairwise registration of point clouds from adjacent viewpoints, obtaining the initial set of transformation matrices as follows: Step 2.1: First, take the nearest point in the target point cloud after spatial transformation of the source point cloud as the corresponding matching point; The source point cloud is transformed to the target point cloud coordinate system using the current transformation matrix. KD-Tree is then used to quickly search for the nearest point of each source point, forming a set of matching point pairs. Step 2.2: Based on the matching points determined in Step 2.1, the optimal transformation matrix is ​​solved using the least squares iterative nearest point algorithm. R , t and residual function L ; Step 2.3: Repeat steps 2.1 to 2.2 until the residual change is reached. Output single-view registration transformation matrix T ij , where i and j are the adjacent viewpoint numbers.

5. The tool point cloud fine registration method based on topographic closure constraints according to claim 4, characterized in that, The least squares iterative nearest point algorithm described in step 2.2 is as follows: Step 2.2.1: Determine the optimal rotation matrix. , It is used to remove the centroid from point clouds; Step 2.2.2: Compute the covariance matrix H of the centroid-removed points. ; Step 2.2.3: Perform singular value decomposition on the covariance matrix H. Then the optimal transformation for: , .

6. The tool point cloud fine registration method based on topographic closure constraints according to claim 1, characterized in that, The method for constructing multi-view global topography closure constraints in step 3 is as follows: Suppose that a total of n point clouds are collected from different viewpoints, and the transformation matrix between adjacent viewpoints is T. 12 T 23 ... T (n -1 )n Then the global closure constraint is ,in This is the transformation matrix from the first viewpoint to the nth viewpoint.

7. The tool point cloud fine registration method based on topographic closure constraints according to claim 1, characterized in that, The method for solving the globally optimal transformation matrix in step 3, which involves optimizing rotation parameters individually, optimizing translation parameters individually, jointly optimizing global parameters, and using the LM algorithm, is as follows: Rotation parameters optimized separately: all translation vectors are fixed, and only the rotation matrices R1, R2, ..., R are optimized. (n-1) Construct the rotational residual function L R The LM algorithm is used to solve for the optimal set of rotation matrices; Translation parameters are optimized separately: the optimal rotation matrix is ​​fixed, and only the translation vectors t1, t2, ..., t are optimized. (n-1) Construct the translation residual function L G The LM algorithm is used to solve for the optimal set of translation vectors; Overall parameter joint optimization: Using the above results as initial values, all rotation matrices and translation vectors are optimized simultaneously.

8. A tool point cloud fine registration system based on topographic closure constraints, characterized in that, The system includes: The balanced feature point set acquisition module is used to remove outliers from the original point cloud using statistical filtering, and to filter overlapping areas of point clouds from adjacent viewpoints using a dual threshold strategy. It also performs segmented sampling of the farthest point along the tool axis to obtain a balanced feature point set. The initial transformation matrix module is used to complete the pairwise registration of point clouds from adjacent viewpoints based on the distributed balanced feature point set obtained by the distributed balanced feature point set acquisition module, using an alternating iterative strategy of matching point pair search and transformation parameter solution, to obtain the initial transformation matrix set. The optimization module is used to construct multi-view global topography closure constraints based on the initial transformation matrix set obtained from the initial transformation matrix module. It solves the global optimal transformation matrix through individual optimization of rotation parameters, individual optimization of translation parameters, joint optimization of global parameters, and the LM algorithm.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-7.

Images