Robust point cloud registration method for industrial element pose estimation

Abstract

The invention relates to a robust point cloud registration method for industrial component pose estimation, which relates to the technical field of industrial component pose estimation, combines geometric curvature features of point clouds with local descriptor FPFH features, constructs a robust similarity measurement method, effectively enhances the expression ability of the local features of the point clouds, and improves the accuracy of industrial component pose estimation. According to a point cloud registration strategy combining local features and global distribution, global optimal matching is realized by constructing a point cloud bipartite graph and a similarity matrix and adopting a KM algorithm, so that the mismatching rate is effectively reduced, and compared with other feature-based registration methods, the registration precision is improved, and the registration efficiency is improved under the multi-view condition. According to the method, industrial elements with complex geometric structures and fine surface textures can be effectively handled in a single scene, a simple scene and a complex scene, and the accuracy and robustness of registration are remarkably improved.

Description

Technical Field

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

[0002] With the widespread application of 3D sensors such as LiDAR and Kinect in various industries, point cloud has become the main data format for describing the three-dimensional world. It is widely used in many fields such as disordered 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 of computer vision, aims to align point cloud data from different perspectives or different times into a unified coordinate system, provide accurate data support for target recognition tasks, and ultimately achieve accurate recognition and positioning of targets.

[0003] Currently, most traditional methods achieve point cloud registration by coarse registration followed by 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 one point as the corresponding point. This selection is repeated to optimize the transformation with the smallest error. However, affected by outlier points, especially when there are few overlapping areas of 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 relies on good initialization, otherwise it may lead to a local optimal solution.

[0004] However, affected by factors such as acquisition equipment, industrial scenes, and target volume, the point cloud of an object cannot obtain complete information through a single perspective. Therefore, point cloud data must be collected from multiple angles. When an object is photographed from multiple perspectives, the overlapping area of ​​the point cloud may be reduced due to differences in camera perspectives. For data with insufficient overlap, it is difficult for the classic SAC-IA method to obtain a good solution consistent with the data due to the relatively small number of similar feature values ​​in the matching pairs. Especially when dealing with industrial components with complex geometric shapes, delicate surface textures and other characteristics, it is often faced with problems of insufficient feature extraction and lack of global expression capabilities, resulting in inefficient registration 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 associations between point clouds through end-to-end learning frameworks or 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 powerful computing platforms and a large amount of training data, which increases the cost and complexity of implementation. Second, deep learning algorithms often lack sufficient robustness and accuracy when faced with industrial components with complex geometry and detailed 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, comprising the following steps: S1. Extract key points from point cloud using 3D-SITF feature point extraction algorithm; S2, using CCFP weighted fusion algorithm and PCA principal component analysis method to extract the principal curvature of the point cloud , , Gaussian curvature K and mean curvature H, through the principal curvature , , Gaussian curvature K and mean curvature H are used to construct the cosine similarity of point pairs, and the weighted similarity measurement function is constructed by combining the FPFH descriptor similarity. ; S3. Combining the point cloud matching strategy with local features and global distribution, the similarity matrix is ​​constructed using the bipartite graph, and the matching is optimized through the KM algorithm. The threshold is set to filter out false matches, and the rotation matrix R and translation matrix T are calculated through SVD to ensure that the matching results are consistent in global distribution and local features.

[0007] In a preferred embodiment, the 3D-SITF feature point extraction algorithm in step S1 is: through the scale space, a point cloud voxel pyramid is established, and the function in the three-dimensional scale space By Gaussian kernel function With 3D images The convolution of gets: ; in, Indicates that the original point cloud data has been scaled. The result after Gaussian kernel smoothing is is the standard deviation of the Gaussian kernel, is the coordinate of a point in the point cloud, is the normalization coefficient, is the Gaussian attenuation term.

[0008] In a preferred embodiment, in step S2, it is assumed that is the target point A and its neighboring points The distance based on the neighborhood point The SPFH of the target point A is weighted to obtain the weighted result. : ;

[0009] in, is the local geometric feature of point A, is the geometric features of the neighborhood points of point A; in, The feature histogram is , each Represents the weighted frequency distribution in the neighborhood of the point, for the i-th descriptor vector Normalize and get the normalized descriptor vector : ;

[0010] For any point in the key point cloud M extracted from the source point cloud P , for any point in the key point cloud N extracted from the target point cloud Q The FPFH feature descriptor vector and , calculate the cosine similarity : .

[0011] In a preferred embodiment, PCA is used to calculate the normal vector of the point cloud to obtain the curvature of the point cloud; During normal vector estimation, the covariance matrix of the neighborhood is calculated : ;

[0012] in, is the neighborhood point, For point The neighborhood centroid of For point Neighborhood of; Covariance matrix Perform eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors: ;

[0013] in, is the characteristic value, is the corresponding feature vector.

[0014] In a preferred embodiment, the principal curvature is calculated by the following formula: ;

[0015] in, and are the maximum and minimum principal curvatures respectively; the smallest eigenvalue Corresponding to the direction of the normal vector, the largest eigenvalue The curvature direction corresponds to the direction perpendicular to the normal vector; Gaussian curvature of each point in the point cloud and mean curvature Calculated from the principal curvatures: .

[0016] In a preferred embodiment, the point pair The curvature feature similarity Expressed as: ;

[0017] in, Representation vector and vector In the k-th dimension and The product of Construct a weighted similarity measurement function through curvature feature cosine similarity and FPFH descriptor similarity : ;

[0018] in is the weight coefficient.

[0019] In the preferred embodiment, in step S3, during 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, and each pair of points and By calculating the CCFP weighted similarity , to connect the dots 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, n is the number of points in the key point cloud N, and each element of the matrix is ​​the similarity of each pair of points. 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 similarity position column is retained. , set the minimum threshold for the point pair combination with the greatest similarity , the point is Value and The values ​​are compared, if Value lower than , the corresponding point pairs are abandoned and only the point pairs greater than the minimum threshold are retained.

[0020] In a preferred embodiment, for the selected point pair set , the solution formula using SVD is as follows: ;

[0021] in, Representing key points Quality Center, Representing key point cloud Center; Matrix ,in is an orthogonal matrix, is a diagonal matrix; Then the rotation matrix R is , the translation matrix T is .

[0022] Compared with the prior art, the present invention has the following beneficial technical effects: (1) The 3D scale invariant feature transform (3D-SIFT) algorithm is introduced to extract key points with rotation invariance to improve the point cloud data processing efficiency and feature expression capabilities.

[0023] (2) A CCFP weighted fusion method is proposed, which combines the geometric curvature features of the point cloud with the FPFH features of the local descriptor to construct a robust similarity measurement function, effectively improving the local feature expression capability of the point cloud.

[0024] (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 mismatching problem is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A flowchart of a robust point cloud registration method for industrial component pose estimation according to the present invention; Figure 2 This is the FPFH network structure diagram; Figure 3 Schematic diagram of complex components in the ROBI dataset; Figure 4 It is a simple scene diagram of each component; Figure 5It is a schematic diagram of the complex scene of each component; Figure 6 A single scene of a piston from different perspectives; Figure 7 This is the RMSE box plot for different scale factors of the Chrome_screw simple scenario; Figure 8 This is the RMSE box plot of different scale factors of the Chrome_screw complex scene; Fig. 9 Registration results for simple scenes with different components; Fig.10 Registration results for complex scenes with different components; Fig.11 Registration results for Chrome_screw low warehouse and full warehouse scenes; Fig.12 Registration results of different algorithms under multiple views. DETAILED DESCRIPTION

[0026] The technical scheme in the embodiment of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiment of the present invention. Obviously, the described embodiment is only a part of the embodiment of the present invention, not all of the embodiments. Based on the embodiment of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0027] like Figure 1 As shown, a robust point cloud registration method for industrial component pose estimation includes the following steps: S1. Extract key points of point cloud through 3D-SITF, so as to enhance the expression ability and processing efficiency of point cloud information.

[0028] Specifically, the 3D-SIFT feature point extraction method is as follows: 3D-SIFT (3D Scale-Invariant Feature Transform) is used to extract key points and feature descriptors that are invariant to rotation, scale, and curvature from 3D point clouds, and has been explored in the study of Difference of Gaussian (DOG) pyramids.

[0029] By using the scale space, a point cloud voxel pyramid is established. The Gaussian kernel function It is obtained by convolution with the three-dimensional image. The calculation method is as follows: (2);

[0030] in, It is a function in the three-dimensional scale space, which means that the original point cloud data has a scale parameter of The result after Gaussian kernel smoothing. is the standard deviation of the Gaussian kernel and controls the degree of smoothing. is the coordinate of a point in the point cloud. is the normalization coefficient, ensuring that the integral of the Gaussian kernel sums to 1. It is a Gaussian attenuation term, which controls the decay speed of the weight with distance.

[0031] 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.

[0032] In order to detect local extreme values ​​of the Gaussian difference pyramid in scale space, points with significant differences in space and scale are found in the DoG layer at each scale. The DoG values ​​of the nearest K neighboring points around each point are compared. If the DoG value of a point is the smallest or largest among all neighbors, the point is considered a candidate key point.

[0033] S2, using CCFP weighted fusion method, using PCA to extract the principal curvature of the point cloud , , Gaussian curvature And the mean curvature , through the curvature feature Construct the cosine similarity of point pairs and construct a weighted similarity measurement function based on the FPFH descriptor similarity , which improves the comprehensive measurement ability of the overall similarity, thereby improving the local feature expression ability of the point cloud.

[0034] Specifically, the CCFP weighted fusion method is as follows: (1) Calculation of FPFH descriptor similarity FPFH is an efficient local descriptor of point cloud, which is mainly used to describe the local geometric characteristics of points. When calculating the FPFH of each point, the number of neighboring points of the feature vector needs to be considered. FPFH only calculates the relationship between the point and its k neighboring points. Figure 2 As shown, a simplified point feature histogram SPFH (Simplified Point Feature Histograms) is formed.

[0035] In the point cloud dataset, each point needs to calculate its SPFH value. The FPFH value is determined by weight analysis between each point and its adjacent points.

[0036] is the target point A and its neighboring points Distance, neighborhood point The SPFH value is used to weight the final histogram of the target point A, as shown in formula (3): (3); in, is the local geometric feature of point A. is the geometric features of the neighborhood points of point A.

[0037] Its histogram form can be regarded as a vector in the feature space. The FPFH descriptor, a -dimensional feature histogram , each of which Represents the weighted frequency distribution of a certain geometric relationship (such as angle or distance) in the neighborhood of the point. In order to ensure the consistency of the similarity measure, the descriptor vector The normalization formula is as follows: (4); in, is the modulus of the descriptor vector. d is the descriptor vector The dimension represents the number of intervals for quantizing geometric relationships.

[0038] For any point in the key point cloud M extracted by 3D-SIFT from the source point cloud P , the target point cloud Q is any point in the key point cloud N extracted by 3D-SIFT The FPFH feature descriptor vector and , the corotation similarity is defined as: (5); (2) Calculation of point pair curvature similarity The surface normal is a key feature in geometric surface analysis. The surface normal of a given point is usually approximated as the normal of the local surface of the point, which can be achieved by fitting a plane using the least squares method, which is essentially the eigenvalue decomposition problem of the covariance matrix, namely, 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.

[0039] In the normal vector estimation process, we need to calculate the point Curvature features, first consider its neighborhood points and calculate the neighborhood covariance matrix , the calculation method is as follows: (6); in, is the neighborhood point, For point The neighborhood centroid of For point 's neighborhood.

[0040] Then the covariance matrix Perform eigenvalue decomposition, as shown in formula (7), to obtain eigenvalues ​​and eigenvectors: (7); in, is the characteristic value, is the corresponding eigenvector. The smallest eigenvalue Corresponding to the direction of the normal vector, the largest eigenvalue The curvature direction corresponds to the direction perpendicular to the normal vector. The principal curvature is calculated by formula (8): (8); in, and are the maximum and minimum principal curvatures respectively.

[0041] Gaussian curvature of each point in the point cloud and mean curvature The principal curvature and The calculation method is as follows: (9); For any point in the key point cloud M extracted by 3D-SIFT from the source point cloud P , establish its multiple eigenvectors based on the curvature characteristics of the point , these eigenvectors are given by the principal curvatures , Gaussian curvature And the principal curvature and At the same time, for any point in the key point cloud N extracted by 3D-SIFT for the target point cloud Q , establish its multiple eigenvectors based on the curvature characteristics of the point .

[0042] According to the cosine formula of vector angle, the point pair The curvature feature similarity It can be expressed as: (10); Among them, among them, Representation vector and vector In the k-th dimension and The product of For vector The Euclidean norm of . For vector The Euclidean norm of .

[0043] Through the curvature feature cosine similarity and FPFH descriptor similarity, the weighted similarity measurement function is constructed as follows: (11); in is the weight coefficient, satisfying 1.

[0044] S3. Combining the point cloud matching strategy with local features and global distribution, the similarity matrix is ​​constructed using the bipartite graph, and the matching is optimized through the KM algorithm. The threshold is set to filter out false matches, and finally the rotation matrix R and translation matrix T are calculated through SVD to ensure the high consistency of the matching results in global distribution and local features.

[0045] Specifically, the point cloud matching strategy is as follows: 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. and The CCFP weighted similarity can be calculated , to connect the two nodes and assign weights to the edges. 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 matrix is ​​the similarity of each pair of points. value.

[0046] Use the KM algorithm to find the best matching point pair set in the similarity matrix. Get a set of optimal matching point pairs. In order to eliminate mismatched point pairs, for any i-th row in the matrix, only keep the corresponding maximum similarity position column. .

[0047] At the same time, a minimum threshold is set for the point pair combination with the greatest similarity , the point is Value and The values ​​are compared, if Value lower than , then the corresponding point pairs are abandoned and only the point pairs larger than the threshold are retained. Finally, for the filtered point pair set , the solution formula using SVD is as follows: (12); in, Representing key point cloud Quality Center, Representing key point cloud Center. Matrix ,in is an orthogonal matrix, is a diagonal matrix. In addition, when the result of two consecutive iterations is less than the set threshold, the calculation is terminated. Then the rotation matrix R is , the translation matrix T is .

[0048] Example 2 This embodiment conducts registration experiments on the public dataset ROBI and the actual dataset.

[0049] The ROBI dataset contains 7 metal industrial components with different reflectivity levels (size ranges from 12.91 mm to 59.4 mm). Four low-bin scenes and five full-bin scenes are constructed for each type of object (low-bin: the number of objects stored in the box is small, that is, a simple scene; full-bin: the number of objects stored in the box is large, and there are occlusions between the objects, that is, a complex scene).

[0050] The robot is used to move the sensor to different viewpoints to capture high-precision depth maps of the scene. Among the seven components in the ROB dataset, five components, Chrome_screw, DIN_connector, DSub_connector, Gear, and Tube_fitting, have complex geometric structures and detailed surface textures such as Figure 3 shown.

[0051] In order to verify the feasibility of the method of the present invention, the present invention converts the depth map of the above five components into point cloud, and performs registration experiments of simple scenes and complex scenes. The simple scene of each component is shown in Figure 4, and the complex scene is shown in Figure 5. Figure 5 shown.

[0052] The present invention selects pistons as experimental data, which also have complex geometric structures and delicate surface textures. Use a fixed bracket to fix the structured light camera (PhoXI 3D-M structured light camera) at a certain scanning height. Place the piston on the scanning platform, and the height between the camera and the piston is about 650mm. Then rotate the structured light camera 45° each time to take point cloud photos, and obtain 4 groups of point clouds of a single scene of the piston, such as Figure 6 shown.

[0053] Registration accuracy is the error between two registered point clouds. This paper uses the root mean square error (RMSE) to evaluate the point Cloud registration quality, RMSE measures the registration accuracy by calculating the root mean square value of the Euclidean distance between corresponding point pairs between two point clouds. The calculation method is as follows: (13); Where n is the number of corresponding point pairs, R and T are rotation and translation matrices, and is the pair of corresponding points before and after registration. The smaller the RMSE value, the higher the registration accuracy.

[0054] In order to optimize the weighted similarity measurement function in CCFP weighted fusion method Parameter ratio, the present invention targets the characteristics of the research object, namely the complex geometric structure and the delicate surface texture, and adjusts the parameter To make adjustments to explore its optimal configuration. Set to 0.8, 0.7, 0.6, 0.5, correspondingly The values ​​are 0.2, 0.3, 0.4, and 0.5 to ensure that the weighted ratio of the two always maintains the normalization of the total weight. This setting not only ensures the reasonable proportion of curvature features and FPFH descriptors in similarity calculations, but also enables a comprehensive evaluation of the matching accuracy under different weight allocations, thereby optimizing the local feature expression capabilities of point clouds.

[0055] A registration experiment was conducted on a simple scene of the Chrome_screw component. Two adjacent point clouds with different perspectives were selected from the multi-view point cloud for 20 groups of registrations. The RMSE box plots under different scale factors are shown in the figure below. Figure 7 As shown. It can be seen that when When is 0.2, the RMSE value is mainly distributed between 2.05mm and 2.45mm, and the average value is about 2.23mm. When it is 0.3, the RMSE value is mainly distributed between 1.65mm and 2.15mm, and the average value is about 1.95mm. When it is 0.4, the RMSE value is mainly distributed between 2.15mm and 2.55mm, and the average value is about 2.36mm. When it is 0.5, the RMSE value is mainly distributed between 2.5mm and 2.9mm, and the average value is about 2.72mm.

[0056] In the registration experiment of the complex scene of the Chrome_screw component, two adjacent point clouds with different perspectives were selected from the multi-view point cloud for 100 groups of registration. The RMSE box plots under different scale coefficients are shown as follows: Figure 8 As shown. It can be seen that when When it is 0.2, the RMSE value is mainly distributed between 1.95mm and 2.55mm, and the average value is about 2.28mm. When it is 0.3, the RMSE value is mainly distributed between 1.7mm and 2.2mm, and the average value is about 1.91mm. When it is 0.4, the RMSE value is mainly distributed between 2.05mm and 2.9mm, and the average value is about 2.49mm. When it is 0.5, the RMSE value is mainly distributed between 2.45mm and 3.25mm, and the average value is about 2.76mm.

[0057] In summary, when is 0.7, When is 0.3, for the Chrome_screw component, both simple and complex scenes are better than other parameter ratios and have good registration accuracy. and The values ​​are set to 0.7 and 0.3.

[0058] The registration results are as follows: 1) Public dataset ROBI registration results The present invention performs registration experiments on simple scenes of DIN_connector, DSub_connector, Gear and Tube_fitting. Two adjacent point clouds with different viewing angles are selected from the multi-view point cloud to perform 20 groups of registration experiments. The registration accuracy results of the method of the present invention are compared with the 4PCS algorithm, the RANSAC algorithm and the SAC-IA algorithm. Fig. 9 The results show that this method has better accuracy and robustness than other algorithms.

[0059] Table 1 shows the registration accuracy of 4PCS, RNASAC, SAC-IA and Ours methods for simple scenes of DIN_connector, DSub_connector, Gear and Tube_fitting components at different viewing angles. This table is a statistical table of the maximum, minimum, average and standard deviation of RMSE for 20 registration results of simple scenes. In the registration experiment of simple scenes of these four components, the algorithm of the present invention shows excellent results, and the average and standard deviation of RMSE are significantly lower than those of the other three algorithms. This shows that the method of the present invention significantly improves the robustness and accuracy for simple scenes of industrial components with complex geometric structures and delicate surface textures under multiple viewing angles.

[0060] Table 1. Statistical data of simple scene registration accuracy of different components Unit: mm

[0061] In complex scenes, the present invention also performs registration experiments on DIN_connector, DSub_connector, Gear and Tube_fitting. The experiment selects two adjacent point clouds with different perspectives from the multi-view point cloud, performs a total of 100 groups of registration experiments, and compares the algorithm of the present invention with the 4PCS, RANSAC and SAC-IA algorithms. The experimental registration accuracy results are shown in Figure 2. Fig.10 The experimental results show that in complex scenarios, the method of the present invention can still maintain high accuracy and robustness, and is significantly better than other comparison algorithms.

[0062] Table 2 shows the registration accuracy of 4PCS, RNASAC, SAC-IA and Ours methods for complex scenes of the above four components at different viewing angles. Compared with the simple scenes in Table 1, the registration of complex scenes contains more interference factors, such as occlusion, noise and local loss, so the registration is more difficult. This table is a statistical table of the maximum, minimum, average and standard deviation of the RMSE of 100 registration results for complex scenes. The experimental results show that in complex scenes, the method of the present invention still maintains a low RMSE average and standard deviation, and shows stronger robustness and stability compared to other algorithms. It further verifies the advantages of the improved SAC-IA algorithm of the present invention in multiple viewing angles and complex scenes, and can more reliably cope with the registration challenges of industrial components with complex geometric structures and delicate surface textures.

[0063] Table 2 Statistics of registration accuracy of different components in complex scenes Unit: mm

[0064] Based on the above analysis, for industrial components with complex geometric structures and fine surface texture features, in multi-view scenarios, whether in simple or complex scenes, the proposed method has achieved significant improvements in the accuracy and robustness of point cloud registration compared with the 4PCS, RANSAC and SAC-IA methods.

[0065] In order to further verify the feasibility of the proposed algorithm, the present invention conducted registration experiments in simple and complex scenes of the Chrome_screw dataset. Specifically, in the multi-view point cloud data of the same scene, two adjacent point clouds with different perspectives were selected as experimental objects, and 4PCS+ICP, RANSAC+ICP, SAC-IA+ICP and Ours+ICP were used for comparative experiments. The registration results are shown in Figure 2. Fig.11As shown in the figure, it can be clearly seen from the visualization results that the Ours+ICP algorithm has the best registration effect. This result shows that the rough registration using the proposed method can effectively provide a better initial alignment, allowing the subsequent ICP algorithm to converge faster in the fine registration stage and obtain more accurate alignment results, thereby significantly improving the overall registration accuracy and stability.

[0066] In the piston data, the present invention compares the 4PCS algorithm, RNASAC algorithm, SAC algorithm and Ours algorithm. Fig.12 The six groups of registration results of point clouds under different viewing angles under different methods are shown respectively. The red point clouds from top to bottom are the target point clouds of 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 the registration results of the 4PCS algorithm under different viewing angles have large deviations, and even registration failure occurs in the registration viewing angles 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 the two algorithms at the registration viewing angles 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 under different viewing angles are almost completely overlapped. Compared with the other three algorithms, the registration effect is the best. Therefore, the method of the present invention has good accuracy and robustness.

[0067] 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 sets of data under different algorithms as shown in Table 3. From the perspective of algorithm execution efficiency, the present invention can conclude 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 analysis of algorithm accuracy, it can be seen that the 4PCS algorithm has the worst registration accuracy, with an average accuracy of 2.034mm for 6 registration results, the average accuracy of the RNASAC algorithm is 1.249mm, and the average accuracy of the SAC-IA algorithm is 1.124mm. The algorithm of the present invention shows higher accuracy with an average accuracy of 0.700mm. In summary, the algorithm of the present invention is superior to the other three algorithms for single scenes of industrial components with complex geometric structures and fine surface textures under multiple perspectives.

[0068] Table 3. Registration accuracy and efficiency statistics of different algorithms for piston data

[0069] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. The embodiments should therefore be considered exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

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, using CCFP weighted fusion algorithm and PCA principal component analysis method to extract the principal curvature of the point cloud , , Gaussian curvature K and mean curvature H, through the principal curvature , , Gaussian curvature K and mean curvature H are used to construct the cosine similarity of point pairs, and the weighted similarity measurement function is constructed by combining the FPFH descriptor similarity. ; S3. Combining the point cloud matching strategy with local features and global distribution, the similarity matrix is ​​constructed using the bipartite graph, and the matching is optimized through the KM algorithm. The threshold is set to filter out false matches, and the rotation matrix R and translation matrix T are calculated through SVD to ensure that the matching results are consistent in global distribution and local features.

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: through the scale space, a point cloud voxel pyramid is established, and the function in the three-dimensional scale space By Gaussian kernel function With 3D images The convolution of gets: ； in, Indicates that the original point cloud data has been scaled. The result after Gaussian kernel smoothing is is the standard deviation of the Gaussian kernel, is the coordinate 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 is the target point A and its neighboring points The distance based on the neighborhood point The SPFH of the target point A is weighted to obtain the weighted result. : ； in, is the local geometric feature of point A, is the geometric features of the neighborhood points of point A; in, The feature histogram is , each Represents the weighted frequency distribution in the neighborhood of the point, for the i-th descriptor vector Normalize and get the normalized descriptor vector : ； For any point in the key point cloud M extracted from the source point cloud P , for any point in the key point cloud N extracted from the target point cloud Q The FPFH feature descriptor vector and , calculate the cosine similarity : 。 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 normal vector estimation, the covariance matrix of the neighborhood is calculated : ； in, is the neighborhood point, For point The neighborhood centroid of For point Neighborhood of; Covariance matrix Perform eigenvalue decomposition to obtain eigenvalues ​​and eigenvectors: ； in, is the characteristic value, 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: ； in, and are the maximum and minimum principal curvatures respectively; the smallest eigenvalue Corresponding to the direction of the normal vector, the largest eigenvalue The curvature direction corresponds to the direction perpendicular to the normal vector; Gaussian curvature of each point in the point cloud and mean curvature Calculated from the principal curvatures: 。 6. The robust point cloud registration method for industrial component pose estimation according to claim 5, characterized in that: Point pair The curvature feature similarity Expressed as: ； in, Representation vector and vector In the k-th dimension and The product of Construct a weighted similarity measurement function through curvature feature cospin similarity and FPFH descriptor similarity : ； in is the weight coefficient.

7. The robust point cloud registration method for industrial component pose estimation according to claim 3, characterized in that: In step S3, during 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 respectively. and By calculating the CCFP weighted similarity , to connect the dots 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, n is the number of points in the key point cloud N, and each element of the similarity matrix is ​​the value; Use the KM algorithm to find the best matching point pair set in the similarity matrix to obtain a set of optimal matching point pairs. For the i-th row in the similarity matrix, only the corresponding maximum Position column of the value , the Value and minimum threshold The values ​​are compared, if Value lower than , the corresponding point pairs are abandoned and only the point pairs greater than the minimum threshold are retained.

8. The robust point cloud registration method for industrial component pose estimation according to claim 7, characterized in that: For the filtered point pair set , 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 ,in is an orthogonal matrix, is a diagonal matrix; Then the rotation matrix R is , the translation matrix T is .

Images