A Robust Point Cloud Registration Method for Industrial Component Pose Estimation

Abstract

The present invention is a robust point cloud registration method for industrial component pose estimation, which relates to the technical field of industrial component pose estimation. By combining the geometric curvature features of the point cloud with the local descriptor FPFH features, a robust similarity measurement method is constructed, effectively enhancing the expression ability of the local features of the point cloud. A point cloud registration strategy that combines local features with global distribution is adopted. By constructing a point cloud bipartite graph and a similarity matrix, and using the KM algorithm to achieve global optimal matching, the false matching rate is effectively reduced. Compared with other feature-based registration methods, the registration accuracy is improved. Under multi-view conditions, this method can effectively handle industrial components with complex geometric structures and delicate surface textures in single-scene, simple-scene, and complex-scene scenarios, significantly improving the accuracy and robustness of registration.

Description

Technical Field

[0001] The present invention relates to the technical field of industrial component pose estimation, and specifically to a robust point cloud registration method for industrial component pose estimation. Background Art

[0002] With the wide application of 3D sensors such as lidar and Kinect in various industries, point clouds have become the main data format for describing the three-dimensional world. And they are widely used in many fields such as unordered grasping, industrial inspection, 3D reconstruction, autonomous driving, robot navigation, remote sensing, and geographic information systems. In these applications, point cloud registration, as a key technology in computer vision, aims to align point cloud data from different perspectives or different times into a unified coordinate system, providing accurate data support for target recognition tasks, and ultimately achieving precise recognition and positioning of the target.

[0003] Currently, most traditional methods achieve point cloud registration by first performing rough registration and then fine registration. The most common combination method is Sample Consensus Initial Alignment (SAC-IA) and Iterative Closest Point (ICP). SAC-IA samples multiple points from the source point cloud and identifies similar subsets in the target point cloud, and finally selects a point as the corresponding point. This selection is repeated to optimize the transformation with the minimum error. However, affected by outlier points, especially when there are few overlapping regions in the point cloud, it is difficult for traditional methods to obtain the global optimal solution. The ICP algorithm achieves fine matching by minimizing the Euclidean distance, but it depends on good initialization, otherwise it may lead to a local optimal solution.

[0004] However, affected by factors such as acquisition devices, industrial scenarios, and target volumes, the point cloud of an object cannot obtain complete information through a single perspective. Therefore, it is necessary to collect point cloud data from multiple angles. When an object is photographed from multiple perspectives, due to the differences in camera perspectives, the overlapping region of the point cloud may decrease. For data with insufficient overlap, since there are relatively few similar feature values in the matching pairs, it is difficult for the classic SAC-IA method to obtain a good solution consistent with the data. Especially when dealing with industrial components with characteristics such as complex geometric shapes and delicate surface textures, problems such as insufficient feature extraction and lack of global expression ability are often faced, resulting in low registration efficiency or even failure.

[0005] In recent years, deep learning-based point cloud registration methods, such as PointNet, PointNet++, and 3DMatch, have demonstrated high accuracy in specific scenarios. These methods achieve the association between point clouds through an end-to-end learning framework or by learning descriptors and show potential in large-scale point cloud data processing. However, the application of deep learning methods in industrial component point cloud registration still faces many challenges. First, these methods usually require a powerful computing platform and a large amount of training data, increasing the cost and complexity of implementation. Second, deep learning algorithms often lack sufficient robustness and accuracy when dealing with industrial components with complex geometries and fine surface textures, especially in the case of insufficient data. Summary of the Invention

[0006] To solve the above problems, the present invention provides a robust point cloud registration method for industrial component pose estimation, including the following steps:

[0007] S1. Extract key points from the point cloud through the 3D-SITF feature point extraction algorithm;

[0008] S2. Use the CCFP weighted fusion algorithm and the PCA principal component analysis method to extract the principal curvature , , the Gaussian curvature K, and the mean curvature H of the point cloud. Construct the cosine similarity of point pairs through the principal curvature , , the Gaussian curvature K, and the mean curvature H, and construct a weighted similarity metric function in combination with the FPFH descriptor similarity; ;

[0009] S3. Combine the point cloud matching strategy of local features and global distribution, use a bipartite graph to construct a similarity matrix, optimize the matching through the KM algorithm, set a threshold to filter out false matches, and calculate the rotation matrix R and the translation matrix T through SVD to ensure the consistency of the matching result in global distribution and local features.

[0010] In a preferred embodiment, the 3D-SITF feature point extraction algorithm in step S1 is as follows: Through the scale space, a point cloud voxel pyramid is established, and the function in the three-dimensional scale space is obtained by the convolution of the Gaussian kernel function and the three-dimensional image :

[0011] ;

[0012] where represents the result of the original point cloud data smoothed by the Gaussian kernel with the scale parameter , is the standard deviation of the Gaussian kernel, is the coordinate of a certain point in the point cloud, is the normalization coefficient, is the Gaussian attenuation term.

[0013] In the preferred embodiment, in the step S2, set as the distance between the target point A and its neighborhood points , and weight the final histogram of the target point A based on the SPFH of the neighborhood points to obtain the weighted result :

[0014] ;

[0015] wherein, is the local geometric feature of point A, is the geometric feature of the neighborhood points of point A;

[0016] wherein, the d-dimensional feature histogram is , and each represents the weighted frequency distribution within the neighborhood of this point. Normalize the i-th descriptor vector to obtain the normalized descriptor vector :

[0017] ;

[0018] For any point in the key point cloud M extracted from the source point cloud P, and for any point in the key point cloud N extracted from the target point cloud Q, the FPFH feature descriptor vectors and , calculate the cosine similarity :

[0019] .

[0020] In the preferred embodiment, use PCA to calculate the normal vector of the point cloud to obtain the curvature of the point cloud;

[0021] During the normal vector estimation process, calculate the covariance matrix of the neighborhood :

[0022] ;

[0023] wherein, is the neighborhood point, is the centroid of the neighborhood of point , is the neighborhood of point ;

[0024] For the covariance matrix Perform eigenvalue decomposition to obtain eigenvalues and eigenvectors:

[0025] ;

[0026] Among them, is the eigenvalue, is the corresponding eigenvector.

[0027] In a preferred embodiment, the principal curvature is calculated by the following formula:

[0028] ;

[0029] Among them, and are the maximum and minimum principal curvatures respectively; the smallest eigenvalue corresponds to the normal vector direction, and the largest eigenvalue corresponds to the curvature direction perpendicular to the normal vector direction;

[0030] The Gaussian curvature and the mean curvature of each point in the point cloud are calculated from the principal curvature:

[0031] .

[0032] In a preferred embodiment, the curvature feature similarity of the point pair is expressed as:

[0033] ;

[0034] Among them, represents the product of the components of the vector and the components and of the vector

[0035] in the k-th dimension; A weighted similarity metric function

[0036] is constructed through the curvature feature cosine similarity and the FPFH descriptor similarity:

[0037] Among them is the weight coefficient.

[0038] In a preferred embodiment, in step S3, during the registration process, by constructing a bipartite graph, all points of the key point cloud M in the source point cloud P and the key point cloud N in the target point cloud Q are placed at both ends of the bipartite graph, and each pair of points and By calculating the CCFP weighted similarity , to connect points and , and assign weights to the edges to construct a similarity matrix of size m×n, where m is the number of points in the key point cloud M and n is the number of points in the key point cloud N. Each element of the matrix is the value of each pair of points;

[0039] Use the KM algorithm to find the set of optimally matched point pairs in the similarity matrix, obtaining a set of optimal matched point pairs. For the i-th row in the similarity matrix, only keep the column corresponding to the position of the maximum similarity, set a minimum threshold for the combination of point pairs with the maximum similarity , compare the value of this point pair with the value. If the value is lower than , then discard the corresponding point pair and only keep the point pairs greater than this minimum threshold.

[0040] In the preferred embodiment, for the filtered set of point pairs , use SVD to solve the formula as follows:

[0041] ;

[0042] Among them, represents the mass center of the key point , represents the center of the key point cloud ; The matrix , where is an orthogonal matrix, is a diagonal matrix;

[0043] Then the rotation matrix R is , and the translation matrix T is .

[0044] Compared with the prior art, the present invention has the following beneficial technical effects:

[0045] (1) Introduce the 3DScale invariant feature transform (3D-SIFT) algorithm to extract key points with rotational invariance to improve the efficiency of point cloud data processing and the ability of feature expression.

[0046] (2) Propose the CCFP weighted fusion method, combine the geometric curvature feature of the point cloud with the local descriptor FPFH feature, construct a robust similarity metric function, and effectively improve the local feature expression ability of the point cloud.

[0047] (3) A point cloud matching strategy combining local features and global distribution is proposed. By constructing a bipartite graph of the source point cloud and the target point cloud and using the Kuhn-Munkres (KM) algorithm to achieve global optimal matching, the problem of false matching is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the robust point cloud registration method for industrial component pose estimation of the present invention;

[0049] Figure 2 It is a structural diagram of the FPFH network;

[0050] Figure 3 It is a schematic diagram of complex components in the ROBI dataset;

[0051] Figure 4 It is a schematic diagram of a simple scene of each component;

[0052] Figure 5 It is a schematic diagram of a complex scene of each component;

[0053] Figure 6 It is a single scene of a piston from different perspectives;

[0054] Figure 7 It is a box plot of RMSE under different scale factors for the simple scene of Chrome_screw;

[0055] Figure 8 It is a box plot of RMSE under different scale factors for the complex scene of Chrome_screw;

[0056] Figure 9 It is the registration result of the simple scene of different components;

[0057] Figure 10 It is the registration result of the complex scene of different components;

[0058] Figure 11 It is the registration result of the low bin and full bin scenes of Chrome_screw;

[0059] Figure 12 It is the registration result of different algorithms from multiple perspectives. DETAILED DESCRIPTION OF THE INVENTION

[0060] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0061] As Figure 1 shown, a robust point cloud registration method for industrial component pose estimation includes the following steps:

[0062] S1. Extract key points from the point cloud through 3D-SITF to enhance the expression ability and processing efficiency of the point cloud information.

[0063] Specifically, the 3D-SIFT feature point extraction method is as follows:

[0064] 3D-SIFT (3D Scale-Invariant Feature Transform) is used to extract rotation, scale, and curvature invariant key points and feature descriptors from 3D point clouds, and has been explored in the research of the Difference of Gaussian (DOG) pyramid.

[0065] The scale space is used to establish a point cloud voxel pyramid. The scale space can be obtained by the convolution of the Gaussian kernel function and the three-dimensional image. The calculation method is as follows:

[0066] (2);

[0067] Among them, is a function in the three-dimensional scale space, representing the result of the original point cloud data after being smoothed by the Gaussian kernel with a scale parameter of . is the standard deviation of the Gaussian kernel, which controls the smoothing degree. is the coordinate of a certain point in the point cloud. is the normalization coefficient, ensuring that the integral sum of the Gaussian kernel is 1. is the Gaussian attenuation term, which controls the attenuation speed of the weight with distance.

[0068] In the constructed scale space, 3D-SIFT determines key points by finding local extreme points of each point in the spatial neighborhood and scale neighborhood. These local extreme points are rotation and scale invariant at different scales.

[0069] To detect the local extreme values of the Difference of Gaussian pyramid in the scale space, points with significant differences in space and scale are found in the DoG layer at each scale. The DoG values of the K nearest neighboring points around each point are compared. If the DoG value of a point is the smallest or largest among all its neighbors, then this point is regarded as a candidate key point.

[0070] S2. Adopt the CCFP weighted fusion method and use PCA to extract the principal curvature of the point cloud , , the Gaussian curvature and the average curvature , construct the cosine similarity of point pairs through curvature features , and construct a weighted similarity metric function by combining the FPFH descriptor similarity, which improves the comprehensive measurement ability of the overall similarity, thereby enhancing the local feature expression ability of the point cloud.

[0071] Specifically, the CCFP weighted fusion method is as follows:

[0072] (1) Calculation of FPFH descriptor similarity

[0073] FPFH is an efficient local descriptor of point cloud, mainly used to describe the local geometric properties of points. When calculating the FPFH of each point, the number of neighborhood points of the feature vector needs to be considered. FPFH only calculates the relationship between a point and its k neighborhood points as Figure 2 shown, forming a Simplified Point Feature Histograms (SPFH).

[0074] In the point cloud dataset, the SPFH value of each point needs to be calculated. By analyzing the weights between each point and its adjacent points, the FPFH value is determined.

[0075] Let be the distance between the target point A and its neighborhood point , and the SPFH value of the neighborhood point is used to weight the final histogram of the target point A, as shown in formula (3):

[0076] (3);

[0077] where, is the local geometric feature of point A. is the geometric feature of the neighborhood point of point A.

[0078] Its histogram form can be regarded as a vector in the feature space. For the FPFH descriptor of a given point , a -dimensional feature histogram , where each represents the weighted frequency distribution of a certain geometric relationship (such as angle or distance) within the neighborhood of the point. To ensure the consistency of the similarity measurement, the descriptor vector is normalized as follows:

[0079] (4);

[0080] where, is the modulus of the descriptor vector. d is the descriptor vector The dimension represents the number of intervals for quantifying geometric relationships.

[0081] For any point in the key point cloud M extracted from the source point cloud P by 3D-SIFT , and any point in the key point cloud N extracted from the target point cloud Q by 3D-SIFT The FPFH feature descriptor vector and , the remaining rotation similarity is defined as:

[0082] (5);

[0083] (2) Calculation of point pair curvature similarity

[0084] The surface normal is a key feature in geometric surface analysis. For a given point, the surface normal is usually approximated as the normal of the local surface at that point, which can be achieved by fitting a plane using the least squares method. Essentially, it is an eigenvalue decomposition problem of the covariance matrix, that is, principal component analysis (PCA). Therefore, the present invention uses PCA to calculate the normal vector of the point cloud, thereby obtaining the curvature of the point cloud.

[0085] During the normal vector estimation process, to calculate the curvature feature of a point, first consider its neighborhood points and calculate the covariance matrix of the neighborhood , the calculation method is as follows:

[0086] (6);

[0087] Among them, is the neighborhood point, is the neighborhood centroid of the point , and is the neighborhood of the point .

[0088] Then perform eigenvalue decomposition on the covariance matrix , as shown in formula (7), to obtain the eigenvalues and eigenvectors:

[0089] (7);

[0090] Among them, are the eigenvalues, are the corresponding eigenvectors. The smallest eigenvalue corresponds to the normal vector direction, and the largest eigenvalue corresponds to the curvature direction perpendicular to the normal vector direction. The principal curvature is calculated by formula (8):

[0091] (8);

[0092] Among them, and are the maximum and minimum principal curvatures respectively.

[0093] The Gaussian curvature and the mean curvature of each point in the point cloud can be calculated from the principal curvatures and by the following calculation method:

[0094] (9);

[0095] For any point in the key point cloud M extracted from the source point cloud P by 3D-SIFT, its multi-feature vector is established according to the curvature characteristics of the point. These feature vectors are composed of the principal curvatures , the Gaussian curvature and the principal curvatures and . At the same time, for any point in the key point cloud N extracted from the target point cloud Q by 3D-SIFT, its multi-feature vector is established according to the curvature characteristics of the point.

[0096] According to the cosine formula of the vector angle, the curvature feature similarity of the point pair can be expressed as:

[0097] (10);

[0098] Among them, represents the product of the components of the vector and the vector in the k-th dimension and ; is the Euclidean norm of the vector . is the Euclidean norm of the vector .

[0099] The weighted similarity metric function is constructed through the curvature feature cosine similarity and the FPFH descriptor similarity as follows:

[0100] (11);

[0101] Among them is the weight coefficient, satisfying 1.

[0102] S3. Point cloud matching strategy combining local features and global distribution, constructing a similarity matrix using a bipartite graph, optimizing the matching through the KM algorithm, setting a threshold to filter out false matches, and finally calculating the rotation matrix R and translation matrix T through SVD to ensure a high degree of consistency in global distribution and local features of the matching result.

[0103] Specifically, the point cloud matching strategy is as follows:

[0104] During the registration process, by constructing a bipartite graph, all points of the key point cloud M in the source point cloud P and the key point cloud N in the target point cloud Q are placed at both ends of the bipartite graph, and each pair of points and can be connected by calculating the CCFP weighted similarity , and weights are assigned to the edges. A similarity matrix of size m×n is constructed, where m is the number of points in the key point cloud M and n is the number of points in the key point cloud N. Each element of the matrix is the value of each pair of points.

[0105] Use the KM algorithm to find the set of point pairs with the optimal match in the similarity matrix. A set of optimal matching point pairs is obtained. In order to eliminate false matching point pairs, for any i-th row in the matrix, only the column corresponding to the position of the maximum similarity is retained .

[0106] At the same time, a minimum threshold is set for the combination of point pairs with the maximum similarity. The value of the point pair is compared with the value. If the value is lower than , the corresponding point pair is discarded, and only the point pairs greater than this threshold are retained. Finally, for the filtered set of point pairs , the solution using SVD is as follows:

[0107] (12);

[0108] Among them, represents the mass center of the key point cloud , represents the center of the key point cloud . The matrix , where is an orthogonal matrix, is a diagonal matrix. In addition, when the results of two adjacent iterations are less than the set threshold, the calculation terminates. Then the rotation matrix R is , and the translation matrix T is .

[0109] Example 2

[0110] In this embodiment, registration experiments were respectively carried out on the public dataset ROBI and the actual dataset.

[0111] The ROBI dataset contains 7 metal industrial components with different reflectivity levels (sized 12.91 mm - 59.4 mm). Four low - bin scenarios and five full - bin scenarios were constructed for each type of object (low - bin: fewer objects stored in the box, i.e., a simple scenario; full - bin: more objects stored in the box, with occlusion between objects, i.e., a complex scenario).

[0112] The robot was used to move the sensor to different viewpoints to capture high - precision depth maps of the scenes. Among the seven components in the ROB dataset, five components, namely Chrome_screw, DIN_connector, DSub_connector, Gear, and Tube_fitting, have complex geometric structures and delicate surface textures as Figure 3 shown.

[0113] To verify the feasibility of the method of the present invention, the depth maps of the above - mentioned five components were converted into point clouds, and registration experiments for simple scenarios and complex scenarios were carried out. The simple scenarios of each component are shown in Figure 4, and the complex scenarios are as Figure 5 shown.

[0114] The present invention selected a piston as experimental data. The piston also has a complex geometric structure and delicate surface texture. A structured - light camera (PhoXI 3D - M structured - light camera) was fixed at a certain scanning height using a fixed bracket. The piston was placed on the scanning platform, and the height between the camera and the piston was approximately 650 mm. Then, the structured - light camera was rotated 45° each time for point - cloud shooting, and 4 groups of point clouds of a single piston scenario were obtained, as Figure 6 shown.

[0115] The registration accuracy is the error between two registered point clouds. In this paper, the root - mean - square error (RMSE) is used to evaluate the quality of point

[0116] cloud registration. RMSE measures the registration accuracy by calculating the root - mean - square value of the Euclidean distances of corresponding point pairs between two point clouds. The calculation method is as follows:

[0117] (13);

[0118] where n is the number of registered corresponding point pairs, R and T are the rotation and translation matrices, and are the corresponding point pairs before and after registration. The smaller the value of RMSE, the higher the registration accuracy.

[0119] To optimize the weighted similarity metric function in the CCFP weighted fusion method Parameter ratio. In view of the characteristics of the research object of the present invention, namely complex geometric structures and delicate surface textures, the parameters are adjusted to explore their optimal configurations. Specifically, the present invention sets to 0.8, 0.7, 0.6, 0.5, and the corresponding values are 0.2, 0.3, 0.4, 0.5 to ensure that the weighted ratio of the two always maintains the normalization of the total weight. Such a setting not only ensures a reasonable proportion of the curvature feature and the FPFH descriptor in the similarity calculation, but also enables the comprehensive evaluation of the matching accuracy under different weight allocations, thereby optimizing the local feature expression ability of the point cloud.

[0120] Registration experiments were carried out on the simple scenario of the Chrome_screw component. Twenty groups of registrations were performed by selecting two adjacent point clouds with different viewpoints from the multi-viewpoint clouds. The RMSE box plots under different proportionality coefficients are as Figure 7 shown. It can be seen that when is 0.2, the RMSE values are mainly distributed between 2.05 mm and 2.45 mm, and the average value is about 2.23 mm. When is 0.3, the RMSE values are mainly distributed between 1.65 mm and 2.15 mm, and the average value is about 1.95 mm. When is 0.4, the RMSE values are mainly distributed between 2.15 mm and 2.55 mm, and the average value is about 2.36 mm. When is 0.5, the RMSE values are mainly distributed between 2.5 mm and 2.9 mm, and the average value is about 2.72 mm.

[0121] Registration experiments were carried out on the complex scenario of the Chrome_screw component, and then 100 groups of registrations were performed by selecting two adjacent point clouds with different viewpoints from the multi-viewpoint clouds. The RMSE box plots under different proportionality coefficients are as Figure 8 shown. It can be seen that when is 0.2, the RMSE values are mainly distributed between 1.95 mm and 2.55 mm, and the average value is about 2.28 mm. When is 0.3, the RMSE values are mainly distributed between 1.7 mm and 2.2 mm, and the average value is about 1.91 mm. When is 0.4, the RMSE values are mainly distributed between 2.05 mm and 2.9 mm, and the average value is about 2.49 mm. When is 0.5, the RMSE values are mainly distributed between 2.45 mm and 3.25 mm, and the average value is about 2.76 mm.

[0122] In summary, when is 0.7, When it is 0.3, for the Chrome_screw component, it is superior to other parameter ratios in both simple and complex scenarios and has good registration accuracy. Therefore, in subsequent experiments, and values are set to 0.7 and 0.3.

[0123] The registration results are as follows:

[0124] 1) Registration results of the public dataset ROBI

[0125] The present invention conducts registration experiments on simple scenarios of DIN_connector, DSub_connector, Gear, and Tube_fitting. Twenty groups of registration experiments are carried out by selecting two adjacent point clouds with different viewpoints from the multi-viewpoint cloud. The registration accuracy results of comparing the method of the present invention with the 4PCS algorithm, RANSAC algorithm, and SAC-IA algorithm are as Figure 9 shown. The results show that, compared with other algorithms, this method has better accuracy and robustness.

[0126] Table 1 shows the registration accuracies of the 4PCS, RNASAC, SAC-IA, and Ours methods for simple scenarios of DIN_connector, DSub_connector, Gear, and Tube_fitting components at different viewpoints. This table is a statistical analysis of the maximum value, minimum value, average value, and standard deviation of the RMSE of 20 registration results for simple scenarios. In the registration experiments for simple scenarios of these four components, the algorithm of the present invention shows excellent results, and the average value and standard deviation of the RMSE are significantly lower than those of the other three algorithms. This indicates that the method of the present invention significantly improves the robustness and accuracy for simple scenarios of industrial components with complex geometric structures and delicate surface textures under multi-viewpoints.

[0127] Table 1 Statistical table of registration accuracy for simple scenarios of different components Unit: mm

[0128]

[0129] In the complex scenario, the present invention also conducts registration experiments on DIN_connector, DSub_connector, Gear, and Tube_fitting. In the experiment, two adjacent point clouds with different viewpoints are selected from the multi-viewpoint cloud, and a total of 100 groups of registration experiments are carried out, and the algorithm of the present invention is compared with the 4PCS, RANSAC, and SAC-IA algorithms. The registration accuracy results of the experiment are as Figure 10 shown. The experimental results show that in the complex scenario, the method of the present invention can still maintain high accuracy and robustness and is significantly superior to other comparison algorithms.

[0130] Table 2 shows the registration accuracies of the 4PCS, RNASAC, SAC-IA, and Ours methods for the above four components in complex scenarios from different perspectives. Compared with the simple scenarios in Table 1, the registration in complex scenarios involves more interference factors, such as occlusion, noise, and local missing, so the registration difficulty is higher. This table is a statistical analysis of the maximum, minimum, average, and standard deviation of the RMSE of 100 registration results in complex scenarios. The experimental results show that in complex scenarios, the method of the present invention still maintains a relatively low average RMSE and standard deviation, and shows stronger robustness and stability compared with other algorithms. It further verifies the advantages of the improved SAC-IA algorithm of the present invention in multi-perspective and complex scenarios, and can more reliably cope with the registration challenges of industrial components with complex geometric structures and delicate surface textures.

[0131] Table 2 Statistical Results of Registration Accuracy for Complex Scenarios of Different Components Unit: mm

[0132]

[0133] Based on the above analysis, for industrial components with complex geometric structures and delicate surface texture features, in multi-perspective scenarios, whether in simple scenarios or complex scenarios, compared with the 4PCS, RANSAC, and SAC-IA methods, the proposed method has achieved significant improvements in the accuracy and robustness of point cloud registration.

[0134] To further verify the feasibility of the proposed algorithm, the present invention conducts registration experiments in simple and complex scenarios of the Chrome_screw dataset. Specifically, in the multi-viewpoint cloud data of the same scenario, two adjacent point clouds with different perspectives are selected as experimental objects, and comparative experiments are carried out using 4PCS+ICP, RANSAC+ICP, SAC-IA+ICP, and Ours+ICP respectively. The registration results are as Figure 11 shown. It can be clearly seen from the visualization results that the registration effect of the Ours+ICP algorithm is the best. This result indicates that using the proposed method for rough registration can effectively provide a more superior initial alignment, enabling the subsequent ICP algorithm to converge faster and obtain a more accurate alignment result in the fine registration stage, thus significantly improving the overall registration accuracy and stability.

[0135] In the piston data, the present invention compares the 4PCS algorithm, RNASAC algorithm, SAC algorithm, and Ours algorithm. Figure 12Six groups of registration results of point clouds under different methods are respectively shown from different perspectives. The red point clouds from top to bottom are the target point clouds at 0°, 45°, 90°, 135°, 0°, and 45°. The blue point clouds from top to bottom are the point cloud data registered with them at 45°, 90°, 135°, 0°, 90°, and 135°. It can be seen from the figure that there are large deviations in the registration results of the 4PCS algorithm under different perspectives, and there are even registration failure phenomena in the registration perspectives of 0° and 90°. The RNASAC algorithm and the SAC-IA algorithm have better registration effects than 4PCS, but there are also small deviations. For example, there are small deviations in both algorithms in the registration perspectives of 0° and 45° and 0° and 90°. It can be seen from the figure that the point clouds of the algorithm of the present invention almost completely overlap under different perspectives. The registration effect is the best compared with the other three algorithms. Therefore, the method of the present invention has good accuracy and robustness.

[0136] In order to better show the accuracy of the above four algorithms, the present invention statistically analyzes the registration accuracy and efficiency of the above 6 groups of data under different algorithms as shown in Table 3. From the perspective of the execution efficiency of the algorithm, the present invention can draw the conclusion that 4PCS has the highest efficiency, SAC-IA has the lowest efficiency, and the efficiency of the algorithm of the present invention is relatively ordinary. From the aspect of algorithm accuracy, it can be seen that the 4PCS algorithm has the worst registration accuracy. The average accuracy of the 6 registration results is 2.034 mm, the average accuracy of the RNASAC algorithm is 1.249 mm, the average accuracy of the SAC-IA algorithm is 1.124 mm, and the algorithm of the present invention shows high accuracy, with an average accuracy of 0.700 mm. In summary, the algorithm of the present invention is superior to the other three algorithms in a single scene of industrial components with complex geometric structures and delicate surface textures under multiple perspectives.

[0137] Table 3 Statistics of registration accuracy and efficiency of different algorithms for piston data

[0138]

[0139] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and can be implemented in other specific forms without departing from the spirit or basic characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claimed rights.

Claims

1. A robust point cloud registration method for industrial component pose estimation, characterized in that: The steps include: S1. Extract key points from point cloud using 3D-SITF feature point extraction algorithm; S2. CCFP weighted fusion algorithm and PCA principal component analysis are used to extract the principal curvature K1, K2, Gaussian curvature K and mean curvature H of the point cloud. The curvature feature cosine similarity of the point pair is constructed through the principal curvature K1, K2, Gaussian curvature K and mean curvature H. The weighted similarity measurement function S is constructed by combining the FPFH descriptor similarity. CCFP ; Point pair (m i ,n j )’s curvature feature cosine similarity S c_cos (m i ,n j ) is expressed as: Among them, m ik ×n jk Represents vector m i and vector n j In the k-th dimension component m ik and n jk The product of For any point m in the key point cloud M extracted from the source point cloud P j , for any point n in the key point cloud N extracted from the target point cloud Q j The FPFH feature descriptor vector f i and f j , calculate the FPFH descriptor similarity S fpfh_cos (m i ,n j ): S3, combining the local features with the global distribution of the point cloud matching strategy, using the bipartite graph to build the similarity matrix, and optimizing the matching through the KM algorithm, setting the threshold to filter out the false matches, and calculating the rotation matrix R and the translation matrix T through SVD to ensure the consistency of the matching results in the global distribution and local features; In the registration process, by constructing a bipartite graph, all the points of the key point cloud M in the source point cloud P and the key point cloud N in the target point cloud Q are placed at the two ends of the bipartite graph. i and n j By calculating the CCFP weighted similarity S ij , to connect point m i and n j , and assign weights to the edges to construct a similarity matrix of size m×n, where m is the number of points in the key point cloud M, n is the number of points in the key point cloud N, and each element of the similarity matrix is ​​the S of each pair of points. ij value; Use the KM algorithm to find the best matching point pair set in the similarity matrix and obtain a set of optimal matching point pairs. For the i-th row in the similarity matrix, only the corresponding maximum S ij The position column of the value j max , the S ij Value and minimum threshold low The values ​​are compared, if S ij The value is lower than the threshold low , the corresponding point pairs are abandoned and only the point pairs greater than the minimum threshold are retained.

2. The robust point cloud registration method for industrial component pose estimation according to claim 1, characterized in that: The 3D-SITF feature point extraction algorithm in step S1 is as follows: a point cloud voxel pyramid is established through the scale space, and the function L(x, y, z, σ) in the three-dimensional scale space is obtained by convolution of the Gaussian kernel function G(x, y, z, σ) and the three-dimensional image I(x, y, z): L(x,y,z,σ)=G(x,y,z,σ)*I(x,y,z) DOG(x,y,z,σ)=L(x,y,z,kσ)-L(x,y,z,σ) Among them, L(x, y, z, kσ) represents the result of the original point cloud data being smoothed by a Gaussian kernel with a scale parameter kσ, σ is the standard deviation of the Gaussian kernel, x, y, z are the coordinates of a point in the point cloud, is the normalization coefficient, is the Gaussian attenuation term.

3. The robust point cloud registration method for industrial component pose estimation according to claim 2, characterized in that: In step S2, it is assumed that w i For the target point A and its neighboring point A i The distance based on the neighborhood point A i The SPFH of the target point A is weighted to obtain the weighted result FPFH(A): Among them, SPFH(A) is the local geometric feature of point A, SPFH(A i ) is the geometric feature of the neighborhood points of point A; Among them, the k-dimensional feature histogram is Each Represents the weighted frequency distribution in the neighborhood of the point. For the i-th descriptor vector f i Normalize and get the normalized descriptor vector f' i :

4. The robust point cloud registration method for industrial component pose estimation according to claim 3, characterized in that: PCA is used to calculate the normal vector of the point cloud and obtain the curvature of the point cloud; During the normal vector estimation process, the covariance matrix C of the neighborhood is calculated: Among them, B j is the neighborhood point, Point B i The neighborhood centroid, N(B i ) is point B i Neighborhood of; Perform eigenvalue decomposition on the covariance matrix C to obtain eigenvalues ​​and eigenvectors: Cv k =λ k v k Among them, λ k is the eigenvalue, v k is the corresponding feature vector.

5. The robust point cloud registration method for industrial component pose estimation according to claim 4, characterized in that: The principal curvature is calculated using the following formula: Among them, K1 and K2 are the maximum and minimum principal curvatures respectively; the smallest eigenvalue λ1 corresponds to the direction of the normal vector, and the largest eigenvalue λ2 corresponds to the curvature direction perpendicular to the normal vector; The Gaussian curvature K and mean curvature H of each point in the point cloud are calculated from the principal curvatures:

6. The robust point cloud registration method for industrial component pose estimation according to claim 5, characterized in that: The weighted similarity measurement function S is constructed by using the curvature feature cosine similarity and the FPFH descriptor similarity. CCFP (m i ,n j ): S CCFP (m i ,n j )=αS fpfh_cos (m i ,n j )+βS c_cos (m i ,n j ) Among them, α and β are weight coefficients.

7. The robust point cloud registration method for industrial component pose estimation according to claim 6, characterized in that: For the filtered point pair set {(p,q)}, the solution formula using SVD is as follows: in, represents the mass center of the key point cloud M, Represents the center of the key point cloud N; matrix H = UVD T , where U and D are orthogonal matrices, and V is a diagonal matrix; Then the rotation matrix R is UD T , the translation matrix T is

Images