Point cloud registration method based on novel feature point matching method
By adopting a new matching method based on feature points in point cloud registration, using ISS key points and multiple feature descriptions to calculate the optimal matching of feature points, combined with the SVD algorithm to obtain the transformation matrix, the problem of insufficient accuracy and efficiency of traditional point cloud registration methods is solved, and high-precision and efficient point cloud registration are achieved.
Patent Information
- Application Number
- CN202411980529.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-31
- Publication Date
- 2025-05-02
AI Technical Summary
The matching of traditional point cloud registration methods has problems of insufficient accuracy and insufficient efficiency.
A new matching method based on feature points is adopted to calculate the ISS key points of the point cloud through ISS detection method, and a feature description such as FPFH, 3DSC or SHOT is used to calculate the feature points, and a unique optimal matching feature point of each feature point is calculated based on the optimal feature matching value. Then, the local and global distance ratios are calculated, the optimal feature points are selected, and the optimal conversion matrix is obtained through the SVD algorithm to achieve optimal registration of point clouds.
High-precision point cloud registration is achieved, which significantly reduces errors, improves registration accuracy, and achieves efficient point cloud registration without sacrificing too much efficiency.
Smart Images
Figure CN119919464A_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the technical field of laser point cloud application, and in particular relates to a point cloud registration method based on a novel feature point matching method. Background Art
[0002] Point cloud registration is a core problem in computer vision and graphics, which involves aligning or matching one set of point cloud data (source point cloud) with another set of point cloud data (target point cloud) so that they correspond in 3D space. Point cloud registration is crucial in many applications, such as 3D modeling, robot navigation, virtual reality and augmented reality, etc.
[0003] 3D laser scanning is limited by the measurement method and the shape of the object. It can only scan and obtain point cloud data of a limited range of the measured object at a time. It is necessary to scan the point cloud from multiple perspectives for coordinate conversion and unify it into a global coordinate system to achieve point cloud registration. Many scholars at home and abroad have conducted in-depth research on point cloud registration and have produced widely used algorithms such as iterative nearest point, feature point matching, and normal distribution transformation. However, the traditional point cloud registration method has problems of insufficient accuracy and efficiency. Therefore, it is urgent to design a point cloud registration method based on a new feature point matching method to solve the above problems. Summary of the invention
[0004] The purpose of the present invention is to provide a point cloud registration method based on a new feature point matching method to solve the problems of insufficient accuracy and efficiency in the traditional point cloud registration method proposed in the above background technology.
[0005] A point cloud registration method based on a novel feature point matching method comprises the following steps:
[0006] S1: Calculate the ISS key points of the source point cloud and the target point cloud by the ISS detection method, calculate the corresponding feature points using feature descriptions such as FPFH, 3DSC or SHOT, and calculate the unique optimal matching feature point of each feature point in the source point cloud in the target point cloud according to the optimal feature matching value;
[0007] S2: Calculate the local distance ratio and f of the feature points of the source point cloud, and select several optimal local feature points with the largest f value;
[0008] S3: Calculate the global distance ratio and F of the optimal local feature point, select several optimal global feature points with the largest F value from step S2, group them in descending order of F value, and obtain the optimal transformation matrix through the SVD algorithm based on each group of optimal global feature points of the source point cloud and the feature points of the target point cloud that match the features, so as to achieve the optimal registration of the point cloud;
[0009] in,
[0010] ISS is (Intrinsic Shape Signatures), internal shape feature detection algorithm;
[0011] FPFH is (Fast Point Feature Histograms), fast point feature histogram;
[0012] 3DSC stands for (3D Shape Context), 3D shape context feature;
[0013] SHOT stands for Signature of Histogram of Orientation, direction histogram.
[0014] Preferably, in step S2, based on the optimal feature (FPFH, 3DSC, SHOT, etc.), in the source point cloud, a, b, c, d are the n0 feature points closest to o, and based on the optimal feature matching value, the optimal matching points of the feature points a, b, c, d of the source point cloud in the target point cloud are A, B, C, D, respectively. Let li (1≤i≤n0) be the distance between the feature point o of the source point cloud and the nearest feature points a, b, c, d, and Li (1≤i≤n0) be the distance between the feature point O of the target point cloud and the feature points A, B, C, D, and the local distance ratio and f of the feature point o of the source point cloud are calculated as follows:
[0015]
[0016] in,
[0017] n0 is the number of feature points closest to the feature point o of the source point cloud, n0 is 150, and the threshold α is 0.9.
[0018] Preferably, in step S2, a small root heap with m1 elements is established, and the information of each element includes the number of the feature point of the source point cloud, the local distance ratio and f, etc., f is the key value, and the logical structure of the heap is a complete binary tree. The key value of any root node of the small root heap is smaller than the key values of the left and right subtrees. The number of feature points of the source point cloud is m0 (m0>m1), and the relevant information of m1 feature points in the source point cloud is taken and stored in the small root heap. The small root heap is adjusted according to the key value f of the feature point so that any root node The key values of the points are all smaller than the key values of the left and right subtrees. The other (m0-m1) feature points of the source point cloud are inserted into the small root heap according to the size of f, so that the key value of any root node is smaller than the key value of the left and right subtrees; the m1 elements in the small root heap correspond to the m1 feature points with the largest (optimal) key value f among the m0 feature points of the source point cloud. The target point cloud has m1 feature points that match its features (FPFH, 3DSC, SHOT, etc.), and the time complexity is O(m0log2m1).
[0019] Preferably, in step S3, based on the m1 optimal local feature points of the source point cloud, let li be the distance between the optimal local feature point o and the other (m1-1) optimal local feature points mentioned above; the optimal local feature point o of the source point cloud is optimally matched with the feature point O of the target point cloud (FPFH, 3DSC, SHOT, etc.), and the other (m1-1) optimal local feature points of the source point cloud are optimally matched with the (m1-1) feature points of the target point cloud, respectively, and Li is the distance between the feature point O of the target point cloud and the other (m1-1) feature points. The global distance ratio and F of the optimal local feature points of the source point cloud are calculated as follows:
[0020]
[0021] in:
[0022] Usually m1 is set to 100 and the threshold α is set to 0.9.
[0023] Preferably, in step S3, the point cloud registration based on the novel feature point matching method comprises the following steps:
[0024] Step 1: Based on the new feature point matching method, obtain the m2 optimal global feature points of the source point cloud and the feature points of the target point cloud that match the feature points. Usually, m2 is set to 50.
[0025] Step 2: Based on the descending order of the global distance ratio and , the optimal global feature points of the source point cloud are divided into n (n=5) groups of feature points consisting of the first 10, first 20, first 30, first 40, and first 50 (m2=50) points;
[0026] Step 3: Based on the optimal global feature points of n groups and the feature points of the target point cloud that match the feature points, the SVD algorithm calculates the transformation matrix T of the corresponding group of the above feature points, calculates the error RMSE of the corresponding group of the source point cloud and the source point cloud transformed by the matrix T, and selects the minimum error RMSE and the corresponding optimal transformation matrix T. ok , the source point cloud passes through the matrix T ok After conversion, the optimal point cloud registration is achieved;
[0027] in,
[0028] RMSE is the root mean square error (Root Mean Square Error, RMSE);
[0029] SVD stands for singular value decomposition (SVD).
[0030] Compared with the prior art, the beneficial effects of the present invention are as follows: the point cloud registration method based on the new feature point matching method realizes high-precision point cloud registration by comprehensively using ISS key point detection, multiple feature descriptions, efficient local and global feature point screening, corresponding relationship determination based on feature matching, small root heap algorithm optimization and SVD transformation matrix solution. Compared with the traditional SAC-IA and K-4PCS algorithms, the method performs better in RMSE evaluation, significantly reduces the error, and improves the accuracy of registration. At the same time, the design of the algorithm takes into account the limitation of computing resources, balances the computational complexity and registration accuracy through effective feature point and feature description selection, so that the algorithm can realize efficient point cloud registration without sacrificing too much efficiency. In addition, the effectiveness and practicability of the method are verified by testing the actual point cloud model, which provides new ideas and technical solutions for the field of point cloud registration. BRIEF DESCRIPTION OF THE DRAWINGS
[0031] Figure 1 is an overall block diagram of the present invention;
[0032] Figure 2 It is a schematic diagram of feature point matching based on the optimal feature matching value;
[0033] Figure 3 Schematic diagram of feature points and distances between source point cloud and target point cloud;
[0034] Figure 4 It is a flowchart of point cloud registration based on a new feature point matching method;
[0035] Figure 5 Schematic diagram of 5 groups of point cloud models for point cloud registration experiment;
[0036] Figure 6 Schematic diagram of the number of points and ISS key points of 5 groups of point cloud models;
[0037] Figure 7 A schematic diagram showing the RMSE data comparison between the new matching algorithm and related algorithms for point cloud registration;
[0038] Figure 8 This is a schematic diagram of the RMSE and time (seconds) of the 5 point cloud registrations of the K-4PCS algorithm;
[0039] Fig. 9 Schematic diagram of the time comparison (seconds) between the new matching algorithm and related methods for point cloud registration;
[0040] Fig.10 This is a schematic diagram comparing the point cloud registration effects of the ISS-FPFH-SAC-IA algorithm and the ISS-FPFH new matching algorithm;
[0041] Fig.11This is a schematic diagram comparing the point cloud registration effects of the ISS-3DSC-SAC-IA algorithm and the ISS-3DSC new matching algorithm;
[0042] Fig.12 This is a schematic diagram comparing the point cloud registration effects of the ISS-SHOT-SAC-IA algorithm and the ISS-SHOT new matching algorithm;
[0043] Fig.13 Schematic diagram comparing the point cloud registration effects of the K-4PCS algorithm and the ISS-FPFH new matching algorithm. DETAILED DESCRIPTION
[0044] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0045] Please see attached Figure 1-13 , the embodiments and related experiments provided by the present invention:
[0046] Example 1
[0047] In an embodiment of the present invention, point cloud registration based on a novel feature point matching method includes the following steps:
[0048] S1, after calculating the ISS key points of the source point cloud and the target point cloud by the ISS detection method, the ISS key points are calculated by one of the features such as FPFH, 3DSC, SHOT, etc. to obtain the corresponding feature points, and based on the optimal feature matching value between the feature points of the source point cloud and the target point cloud, the unique optimal matching feature point of each feature point of the source point cloud in the target point cloud is calculated;
[0049] S2, calculate the local distance ratio and f of the feature points of the source point cloud, and select several optimal local feature points with the largest f value;
[0050] Specifically, based on the optimal feature (FPFH, 3DSC, SHOT, etc.) matching value, the feature point o of the source point cloud and the feature point O of the target point cloud are the optimal matching points; Figure 3As shown, in the source point cloud, a, b, c, d are the n0 feature points closest to o; based on the optimal feature matching value, the optimal matching points of the feature points a, b, c, d of the source point cloud in the target point cloud are A, B, C, D respectively. Let li (1≤i≤n0) be the distance between the feature point o of the source point cloud and the nearest feature point a, b, c, d, and Li (1≤i≤n0) be the distance between the feature point O of the target point cloud and the feature points A, B, C, D. The local distance ratio and f of the feature point o of the source point cloud are calculated as follows:
[0051]
[0052] in:
[0053] n0 is the number of feature points closest to the feature point o of the source point cloud. Usually n0 is set to 150 and the threshold α is set to 0.9;
[0054] Specifically, a small root heap with m1 elements is established, and the information of each element includes the number of the feature points of the source point cloud, the local distance ratio and f, etc., where f is the key value; the logical structure of the heap is a complete binary tree, and the key value of any root node of the small root heap is smaller than the key values of the left and right subtrees; the number of feature points of the source point cloud is m0 (m0>m1), and the relevant information of m1 feature points in the source point cloud is taken and stored in the small root heap. The small root heap is adjusted according to the key value f of the feature point, so that the key value of any root node is smaller than the key value of the left and right subtrees; the other (m0-m1) feature points of the source point cloud are inserted into the small root heap according to the size of f, so that the key value of any root node is smaller than the key value of the left and right subtrees; finally, the m1 elements in the small root heap correspond to the m1 feature points with the largest (optimal) key value f among the m0 feature points of the source point cloud, and the target point cloud has m1 feature points that match its features (FPFH, 3DSC, SHOT, etc.), and the time complexity is O(m0log2m1);
[0055] S3, calculating the global distance ratio and F of the optimal local feature point, selecting several optimal global feature points with the largest F value from step S2, grouping them in descending order of F value, and obtaining the optimal transformation matrix through the SVD algorithm based on each group of optimal global feature points of the source point cloud and the feature points of the target point cloud with matching features, so as to achieve the optimal registration of the point cloud;
[0056] Specifically, based on the m1 optimal local feature points of the source point cloud, let li be the distance between the optimal local feature point o and the other (m1-1) optimal local feature points mentioned above; the optimal local feature point o of the source point cloud is the optimal feature (FPFH, 3DSC, SHOT, etc.) match with the feature point O of the target point cloud, and the other (m1-1) optimal local feature points of the source point cloud are the optimal feature matches with the (m1-1) feature points of the target point cloud, respectively, and Li is the distance between the feature point O of the target point cloud and the other (m1-1) feature points. The global distance ratio and F of the optimal local feature points of the source point cloud are calculated as follows:
[0057]
[0058] in:
[0059] Usually m1 is set to 100 and the threshold α is set to 0.9;
[0060] Specifically, the m1 best local feature points of the source point cloud are sorted in descending order of the global distance ratio and F, and the first m2 best local feature points are taken as the best global feature points to be aligned with the feature points matching the features in the target point cloud. Usually, m2 is set to 50.
[0061] Specifically, the point cloud registration based on the new feature point matching method has the following steps:
[0062] Step 1: Based on the new feature point matching method, obtain the m2 optimal global feature points of the source point cloud and the feature points of the target point cloud that match the feature points. Usually, m2 is set to 50.
[0063] Step 2: Based on the descending order of the global distance ratio and , the optimal global feature points of the source point cloud are divided into n (n=5) groups of feature points consisting of the first 10, first 20, first 30, first 40, and first 50 (m2=50) points;
[0064] Step 3: Based on the optimal global feature points of n groups and the feature points of the target point cloud that match the feature points, the SVD algorithm calculates the transformation matrix T of the corresponding group of the above feature points, and calculates the error RMSE of the corresponding group of the source point cloud and the source point cloud transformed by the matrix T; select the smallest error RMSE and the corresponding optimal transformation matrix T ok , the source point cloud passes through the matrix T ok After conversion, the optimal point cloud registration is achieved.
[0065] Experiment 1
[0066] The present invention conducts relevant experiments on five groups of point cloud models, including rabbit, pig, cow, dragon, and ruins. Figure 5As shown, the source point cloud is the green part, and the target point cloud is the blue part; rabbit, pig, cow, and dragon are point cloud models from public databases such as Stanford University, and the site is a corresponding simplified point cloud model obtained by grid downsampling the three-dimensional laser point cloud of the West Gate Site of Tiewengcheng (the point cloud registration part of the key R&D project in Zhenjiang City that the inventor is responsible for). The hardware and software platform of the present invention: processor Intel (R) Core (TM) i7-7700HQ CPU @ 2.80GHz 2.80GHz, memory 8.00GB, 64-bit Windows 10 operating system, Visual Studio 2022.
[0067] Figure 6 It is the number of points and ISS key points of 5 groups of source point clouds and target point clouds. ISS key points are the key points generated by the ISS detection algorithm of the point cloud model. The above ISS key points use the corresponding algorithm to generate feature points such as FPFH, 3DSC, SHOT, etc., as the input feature points of the SAC-IA (Sample Consensus Initial Alignment) algorithm and the new matching algorithm based on feature points of the present invention. Based on the above feature points, the SAC-IA algorithm is abbreviated as ISS-FPFH-SAC-IA algorithm, ISS-3DSC-SAC-IA algorithm, and ISS-SHOT-SAC-IA algorithm; the algorithm of the present invention is abbreviated as ISS-FPFH new matching algorithm, ISS-3DSC new matching algorithm, and ISS-SHOT new matching algorithm. The above 6 algorithms and 7 algorithms such as K-4PCS (4-Points Congruent Sets) algorithm are used as examples to compare the quality and efficiency of point cloud registration using 5 groups of point cloud models.
[0068] The ISS-FPFH, ISS-3DSC, and ISS-SHOT feature point matching schemes used in the common point cloud registration algorithm are the ISS-FPFH-SAC-IA algorithm, ISS-3DSC-SAC-IA algorithm, and ISS-SHOT-SAC-IA algorithm in the experimental part of the present invention; the K-4PCS scheme used in the common point cloud registration algorithm corresponds to the K-4PCS algorithm in the experimental part of the present invention. Therefore, the novel matching algorithm of the present invention is comparable to the related schemes of the above-mentioned common point cloud registration algorithm.
[0069] Root Mean Square Error (RMSE) is a commonly used indicator for evaluating the quality of point cloud registration. It measures the average error between the source point cloud and the registered source point cloud (after optimal matrix conversion), and provides a statistical measure of the error between point clouds. The calculation formula of RMSE is as follows:
[0070] The present invention conducts registration experiments on 5 groups of source point clouds and target point clouds. Figure 7 and Fig. 9 These are the RMSE and time data comparisons of the above 7 point cloud registration algorithms. Figure 8 As shown in Figure 2, the RMSE results of the five point cloud registrations show that the point cloud registration results of the K-4PCS algorithm are random. Therefore, the present invention selects the K-4PCS algorithm with the smallest (optimal) RMSE data and the relevant RMSE and time data (in bold) to be filled in respectively. Figure 7 and Fig. 9 This makes the comparison between algorithms objective.
[0071] like Figure 7 As shown, based on the same ISS key points and the same features (FPFH, 3DSC, SHOT) of 5 groups of point cloud models, the RMSE of the ISS-FPFH new matching algorithm, the ISS-3DSC new matching algorithm, and the ISS-SHOT new matching algorithm are respectively reduced by 36.24%, 22.85%, and 11.44% on average compared with the ISS-FPFH-SAC-IA algorithm, the ISS-3DSC-SAC-IA algorithm, and the ISS-SHOT-SAC-IA algorithm; the RMSE of the ISS-FPFH new matching algorithm is reduced by 41.31% on average compared with the K-4PCS algorithm. Therefore, based on an objective comparison of RMSE data, the new matching method based on feature points (FPFH, 3DSC, SHOT) of the present invention is better than the corresponding SAC-IA algorithm, and the ISS-FPFH new matching algorithm of the present invention is better than the K-4PCS algorithm.
[0072] like Figure 10-13 As shown, the green part of the point cloud model is the source point cloud, the blue part is the target point cloud, and the red part is the source point cloud after registration. Based on subjective comparison of point cloud registration, the novel matching method based on feature points (FPFH, 3DSC, SHOT) of the present invention is better than the corresponding SAC-IA algorithm, and the novel ISS-FPFH matching algorithm of the present invention is better than the K-4PCS algorithm.
[0073] like Fig. 9 As shown, the ISS-3DSC based algorithms (SAC-IA, new matching) take longer to align than the corresponding ISS-FPFH or SHOT based algorithms, because the 3DSC feature calculation time is much longer than the FPFH or SHOT feature calculation time.
[0074] Based on the comparison of RMSE objective data, subjective effect, execution time, etc. of the five groups of point cloud model registrations, the ISS-FPFH new matching algorithm of the present invention is better than the other six point cloud registration algorithms mentioned above (ISS-FPFH-SAC-IA algorithm, ISS-3DSC-SAC-IA algorithm, ISS-SHOT-SAC-IA algorithm, ISS-3DSC new matching algorithm, ISS-SHOT new matching algorithm, K-4PCS algorithm), that is, the algorithm of the present invention is better than the point cloud registration based on ISS-FPFH, ISS-3DSC, ISS-SHOT feature point matching scheme, and K-4PCS scheme in the conventional point cloud registration algorithms.
[0075] The present invention uses the ISS (Intrinsic Shape Signatures) detection algorithm to extract ISS key points from the source point cloud or the target point cloud. The ISS detection algorithm detects whether the point pi in the point cloud is a key point, which mainly includes the following steps:
[0076] (1) Based on the neighboring points in the bounding sphere with pi as the center and radius r, establish the covariance matrix, calculate the corresponding eigenvectors and eigenvalues, and obtain the eigenvalues from large to small and
[0077] (2) If pi is an ISS key point, then:
[0078]
[0079] Among them, γ 21 and γ 32 The threshold is usually less than 1.
[0080] Based on the ISS key points pi of the source point cloud or the target point cloud, the present invention calculates feature descriptors (FPFH, 3DSC, SHOT, etc.). The three feature descriptors used in the present invention are as follows:
[0081] (1) FPFH feature descriptor
[0082] FPFH (Fast Point Feature Histograms) is a robust multi-dimensional feature descriptor. Its main calculation steps are as follows:
[0083] 1) The neighboring point in the bounding sphere with radius R and centered at ISS key point pi is qj. In the bounding spheres with radius R and centered at pi and qj respectively, the tuple (α, φ, θ) of the normal vector relationship between pi and qj and the neighboring points in the corresponding bounding spheres is calculated. The three features of the tuple are counted into corresponding histograms. Each histogram contains 11 intervals. The number of points in these 11 intervals is counted in each histogram. The three histograms are connected to obtain a 33-dimensional feature vector SPFH (Simple Point Feature Histograms) for pi and qj:
[0084] [α1Lα 11 ]+[φ1Lφ 11 ]+[θ1Lθ 11 ]
[0085] 2) Combine the ISS key point pi and the neighboring point qk to determine the final FPFH of pi:
[0086]
[0087] Where: j is the distance between pi and qk.
[0088] (2) 3DSC feature descriptor
[0089] 3DSC (3D Shape Context) is directly extended from 2DSC. Its main calculation steps are as follows:
[0090] 1) Two concentric spheres are established with the ISS key point pi as the center, with the minimum and maximum radii rmin and rmax respectively, and the normal vector ni of pi is the north pole direction of the sphere. The two concentric spheres are divided into J concentric spheres according to radius, K parts according to longitude, and L parts according to latitude, resulting in J*K*L subspaces.
[0091] 2) The point between the two concentric spheres of the minimum and maximum radius is qj, and its weight is calculated as follows:
[0092]
[0093] Where: V(j,k,l) is the subspace volume of the jth concentric sphere, the kth longitude direction, and the lth latitude direction, ρ j is the density of qj.
[0094] 3) The area of the histogram corresponds to J*K*L subspaces, and the value of the area is obtained by summing the weights of each point qj in the corresponding subspace. Each equal angle in the longitude direction acts as 0o, generating K shape contexts. When using 3DSC feature descriptors for comparison, select the closest one.
[0095] (3) SHOT feature descriptor
[0096] SHOT (Signature of Histogram of Orientation) is a descriptor based on local features. The main calculation steps are as follows:
[0097] 1) Based on all the points in the bounding sphere with the ISS key point pi as the center and the radius R, a covariance matrix is established, and the eigenvalues λ1, λ2 and λ3 from large to small are calculated. The corresponding eigenvectors v1, v2 and v3 correspond to the x, y and z coordinate axes of the bounding sphere respectively.
[0098] 2) Based on the above enclosing sphere coordinate system, the three dimensions of radius, longitude and latitude are divided into 32 subspaces. That is, the radius is divided into two parts, the inner and outer spheres, the latitude is divided into two parts, and the longitude is divided into eight parts.
[0099] 3) In each of the above subspaces, calculate the cosine value of the angle between each neighborhood point qj and the normal vector v3 of pi. Statistical angle cosine values generate an 11-dimensional histogram. Finally, the histograms of each subspace are combined to form a 352-dimensional high-order histogram feature. To eliminate the boundary effect, the interpolation method is used to accumulate each point in the adjacent subspace to a specific cell of the local histogram.
[0100] Based on the above feature points, the SAC-IA (Sample Consensus Initial Alignment) algorithm is abbreviated as ISS-FPFH-SAC-IA algorithm, ISS-3DSC-SAC-IA algorithm, and ISS-SHOT-SAC-IA algorithm; the algorithm of the present invention is abbreviated as ISS-FPFH new matching algorithm, ISS-3DSC new matching algorithm, and ISS-SHOT new matching algorithm. The above 6 algorithms and 7 algorithms including K-4PCS (4-Points Congruent Sets) algorithm are compared in the following experimental part.
[0101] Singular Value Decomposition (SVD) is an important matrix decomposition method in linear algebra. It is a generalization of eigendecomposition on arbitrary matrices and has important applications in signal processing, statistics and other fields.
[0102] Root Mean Square Error (RMSE) is a commonly used indicator for evaluating the quality of point cloud registration. It measures the average error between the source point cloud and the registered source point cloud (after optimal matrix conversion), and provides a statistical measure of the error between point clouds. The calculation formula of RMSE is as follows:
[0103]
[0104] 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. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, 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 point cloud registration method based on a new feature point matching method, characterized in that The following steps are involved: S1: Calculate the ISS key points of the source point cloud and the target point cloud by the ISS detection method, calculate the corresponding feature points using feature descriptions such as FPFH, 3DSC or SHOT, and calculate the unique optimal matching feature point of each feature point in the source point cloud in the target point cloud according to the optimal feature matching value; S2: Calculate the local distance ratio and f of the feature points of the source point cloud, and select several optimal local feature points with the largest f value; S3: Calculate the global distance ratio and F of the optimal local feature point, select several optimal global feature points with the largest F value from step S2, group them in descending order of F value, and obtain the optimal transformation matrix through the SVD algorithm based on each group of optimal global feature points of the source point cloud and the feature points of the target point cloud that match the features, so as to achieve the optimal registration of the point cloud; in, ISS is (Intrinsic Shape Signatures), internal shape feature detection algorithm; FPFH is (Fast Point Feature Histograms), fast point feature histogram; 3DSC stands for (3D Shape Context), 3D shape context feature; SHOT stands for Signature of Histogram of Orientation, direction histogram.
2. The point cloud registration method based on a novel feature point matching method according to claim 1, characterized in that: In step S2, based on the optimal features (FPFH, 3DSC, SHOT, etc.), in the source point cloud, a, b, c, d are the n0 feature points closest to o, and based on the optimal feature matching value, the optimal matching points of the feature points a, b, c, d of the source point cloud in the target point cloud are A, B, C, D, respectively. Let li (1≤i≤n0) be the distance between the feature point o of the source point cloud and the nearest feature points a, b, c, d, and Li (1≤i≤n0) be the distance between the feature point O of the target point cloud and the feature points A, B, C, D, and the local distance ratio and f of the feature point o of the source point cloud are calculated as follows: in, n0 is the number of feature points closest to the feature point o of the source point cloud, n0 is 150, and the threshold α is 0.
9.
3. The point cloud registration method based on the novel feature point matching method according to claim 2, characterized in that: In the step S2, a small root heap with m1 elements is established, and the information of each element includes the number of the feature points of the source point cloud, the local distance ratio and f, etc., where f is the key value. The logical structure of the heap is a complete binary tree. The key value of any root node of the small root heap is smaller than the key values of the left and right subtrees. The number of feature points of the source point cloud is m0 (m0>m1). The relevant information of m1 feature points in the source point cloud is taken and stored in the small root heap. The small root heap is adjusted according to the key value f of the feature point, so that the key value of any root node is smaller than the key value of the left and right subtrees. The other (m0-m1) feature points of the source point cloud are inserted into the small root heap according to the size of f, so that the key value of any root node is smaller than the key value of the left and right subtrees; the m1 elements in the small root heap correspond to the m1 feature points with the largest (optimal) key value f among the m0 feature points of the source point cloud. The target point cloud has m1 feature points that match its features (FPFH, 3DSC, SHOT, etc.), and the time complexity is O(m0log2 m1).
4. The point cloud registration method based on a novel feature point matching method according to claim 1, characterized in that: In step S3, based on the m1 optimal local feature points of the source point cloud, let li be the distance between the optimal local feature point o and the other (m1-1) optimal local feature points mentioned above; the optimal local feature point o of the source point cloud is optimally matched with the feature point O of the target point cloud (FPFH, 3DSC, SHOT, etc.), and the other (m1-1) optimal local feature points of the source point cloud are optimally matched with the (m1-1) feature points of the target point cloud, respectively, and Li is the distance between the feature point O of the target point cloud and the other (m1-1) feature points. The global distance ratio and F of the optimal local feature points of the source point cloud are calculated as follows: in: Usually m1 is set to 100 and the threshold α is set to 0.
9.
5. The point cloud registration method based on the novel feature point matching method according to claim 4, characterized in that: In step S3, the point cloud registration based on the novel feature point matching method is performed as follows: Step 1: Based on the new feature point matching method, obtain the m2 optimal global feature points of the source point cloud and the feature points of the target point cloud that match the feature points. Usually, m2 is set to 50. Step 2: Based on the descending order of the global distance ratio and , the optimal global feature points of the source point cloud are divided into n (n=5) groups of feature points consisting of the first 10, first 20, first 30, first 40, and first 50 (m2=50) points; Step 3: Based on the optimal global feature points of n groups and the feature points of the target point cloud that match the feature points, the SVD algorithm calculates the transformation matrix T of the corresponding group of the above feature points, calculates the error RMSE of the corresponding group of the source point cloud and the source point cloud transformed by the matrix T, and selects the minimum error RMSE and the corresponding optimal transformation matrix T. ok , the source point cloud passes through the matrix T ok After conversion, the optimal point cloud registration is achieved; in, RMSE is the root mean square error (Root Mean Square Error, RMSE); SVD stands for singular value decomposition (SVD).