Search

CN-122023486-A - Two-stage point cloud registration method based on angle consistency constraint

CN122023486ACN 122023486 ACN122023486 ACN 122023486ACN-122023486-A

Abstract

The invention discloses a two-stage point cloud registration method based on angle consistency constraint, and belongs to the technical field of three-dimensional computer vision. The method comprises the steps of firstly extracting multi-stage characteristics of point cloud and carrying out rough matching, then constructing a first-order similarity matrix and a second-order similarity matrix by utilizing rotation invariance of vector included angles, accurately distinguishing inner points from outer points through global geometric compatibility, then adopting seed selection, local spectrum matching and weighted singular value decomposition to solve the optimal pose, eliminating error super point pairs, and finally utilizing reserved high-quality super point pairs to guide fine registration. The invention can obviously improve the registration precision and robustness under the scenes of low overlapping, weak textures and repeated structures, and effectively solves the registration difficulty under the complex scene.

Inventors

  • YANG LIPING
  • You Chaozhen
  • HUANG HONG

Assignees

  • 重庆大学

Dates

Publication Date
20260512
Application Date
20260202

Claims (8)

  1. 1. The two-stage point cloud registration method based on the angle consistency constraint is characterized by comprising the following steps of: 1) Coarse matching, namely inputting a source point cloud and a target point cloud to be registered into a neural network for downsampling and feature extraction, calculating feature correlation between the source point cloud and the target point cloud super points, and selecting a plurality of super point pairs with highest correlation to form an initial super point matching pair set; 2) Constructing a first-order angle similarity matrix, namely introducing a third reference matching point into any two pairs of matching points in the initial super-point matching pair set, and constructing triangles based on three points in a source point cloud and a target point cloud respectively; 3) Performing binarization processing on the first-order angle similarity matrix, counting the number of adjacent matched pairs which are commonly compatible in a global range of each matched pair, and converting local first-order angle consistency into a second-order similarity matrix reflecting global geometric distribution consistency; 4) The seed selection and consensus set construction, namely performing spectrum decomposition on the first-order angle similarity matrix subjected to binarization processing to obtain initial confidence coefficient, screening out high-confidence seed matching pairs with uniform space distribution; 5) Calculating the weight of each point in the local consensus set by using a spectrum matching method, performing weighted singular value decomposition based on the weight, and solving a plurality of candidate rigid body transformation matrixes; 6) Performing residual error check on the initial super-point matching pair set by utilizing the candidate rigid body transformation matrix, screening out the candidate rigid body transformation matrix with the largest number of inner points as the optimal transformation, removing all error super-point matching pairs which do not accord with the transformation from the initial super-point matching pair set according to the optimal transformation, and reserving correct super-point matching pairs; 7) And (3) fine matching, namely expanding the reserved correct super-point matching pair into a neighborhood block, carrying out local dense point matching in the neighborhood block, and finally calculating a final high-precision rigid body transformation matrix by adopting a local-global weighted singular value decomposition iteration strategy to obtain a final point cloud registration result.
  2. 2. The two-stage point cloud registration method based on the angular consistency constraint of claim 1, wherein the specific implementation steps of step 1) are as follows: 1.1 Performing multi-stage downsampling on the source point cloud and the target point cloud to respectively obtain a dense point set and a super point set of the source point cloud and the target point cloud, and extracting corresponding dense point features and super point features; 1.2 Geometry embedding construction of the superpoints in the set of the source point cloud and the target point cloud Selecting k nearest neighbor superpoints to jointly form a set For a collection Two super points in (a) And Calculation of And The Euclidean distance between, and And Super point adjacent to it Three-way angle of formation Wherein the neighbors are super points Mapping distance and angle into high-dimensional features by using sine and cosine functions, and calculating distance embedding as follows And angle embedding : In the middle of And Is a sensitivity parameter controlling the distance change and the angle change; And calculating geometry embedding by aggregating pair-wise distance embedding and angle embedding : Wherein the method comprises the steps of And Respectively for distance embedding And angle embedding Is a projection matrix of (a); 1.3 The method comprises the steps of establishing a geometrical self-attention module, utilizing the geometrical structure embedded feature to correct attention weight, learning global relativity of features in single point cloud and geometrical space, establishing a feature-based cross-attention module, and carrying out bidirectional feature interaction between a source point cloud and a target point cloud; 1.4 And (3) super-point matching, namely calculating a Gaussian correlation matrix of the mixed characteristics of the source point cloud and the target point cloud, and selecting a plurality of first matching pairs with highest correlation scores to form an initial super-point matching pair set.
  3. 3. The two-stage point cloud registration method based on the angular consistency constraint of claim 1, wherein the method for constructing the first-order angular similarity matrix in the step 2) is as follows: 2.1 Firstly, randomly selecting a plurality of matching pairs from the initial super-point matching pair set to construct a reference matching pair set M; 2.2 For any two matching pairs i and j of similarity to be calculated in an initial super-point matching pair set, randomly selecting one matching pair from the reference matching pair set M as a reference matching pair M of the calculation, respectively extracting source points corresponding to the matching pair i, the matching pair j and the reference matching pair M in a source point cloud and target points corresponding to the matching pair M in a target point cloud, respectively connecting the source points of the reference matching pair M and the source points of the matching pair j by using the source points of the matching pair i in the source point cloud, and constructing two source vectors And (3) with In the target point cloud, the target points of the matched pair i are respectively connected with the target point of the reference matched pair m and the target point of the matched pair j to construct two target vectors And (3) with Calculating two vectors constructed in the source point cloud And (3) with Included angle of (2) And corresponding vectors in the target point cloud And (3) with Included angle of (2) Calculating the absolute value of the difference between the two sets of vector angles : ; Calculating to obtain first-order similarity values of the matching pair i and the matching pair j by using a monotonically decreasing function : ; Wherein, the Is a linear rectification function; 2.3 Maintaining the matched pair i unchanged, repeating the step 2.2), and traversing all matched pairs in the initial super-point matched pair set to obtain a first-order similarity value of the matched pair i and all matched pairs including the matched pair i; 2.4 Repeating the steps 2.2) -2.3), traversing the initial set of super-point matching pairs to calculate first-order similarity values of each matching pair and all matching pairs including the initial set of super-point matching pairs, thereby obtaining first-order similarity values of all matching pairs in the initial set of super-point matching pairs and constructing to obtain a first-order angle similarity matrix 。
  4. 4. The two-stage point cloud registration method based on angle consistency constraint of claim 3, wherein the method for constructing the second-order similarity matrix in the step 3) comprises the steps of setting an angle threshold value If (1) Then it is determined that the matched pair i and matched pair j are geometrically compatible, order the If not, judging that the matched pair i and the matched pair j are incompatible, and making Constructing a second-order similarity matrix in step 3) The specific formula of (2) is: ; Wherein, the The formula represents that the number of the third matching pair k which are compatible together is counted as a second-order similarity value only when the matching pair i and the matching pair j are compatible.
  5. 5. The two-stage point cloud registration method based on the angle consistency constraint according to claim 1 is characterized in that the seed selection method in the step 4) is that power iteration is carried out on a first-order angle similarity matrix after binarization processing, a main feature vector is solved, element values of the main feature vector are used as initial interior point probabilities of all matching pairs, non-maximum suppression is carried out within a preset space radius R range, a matching pair with the highest local interior point probability is reserved as a seed, and matching pairs with lower probability in a neighborhood are suppressed, so that a seed matching pair set with uniform space distribution and high confidence is screened out.
  6. 6. The two-stage point cloud registration method based on the angular consistency constraint of claim 1, wherein the step 5) specifically comprises: 5.1 Constructing a local graph by taking matching pairs in the local consensus set as nodes and taking second-order similarity values between the matching pairs as edge weights, thereby constructing the local graph based on the local consensus set; 5.2 Performing feature decomposition on the adjacent matrix of the local graph, and taking the main feature vector as the interior point probability weight of each matching pair in the local consensus set; 5.3 Weighting singular value decomposition, namely weighting the coordinates of points in the local consensus set by utilizing the internal point probability weight, and solving a candidate rotation matrix through singular value decomposition Translation vector A plurality of candidate rigid body transformation matrices are obtained.
  7. 7. The two-stage point cloud registration method based on the angle consistency constraint according to claim 1 is characterized in that the specific implementation method of the step 6) comprises the steps of traversing all candidate rigid body transformation matrixes, calculating residual errors of all matching pairs in an initial super-point matching pair set under transformation, counting the number of inner points with the residual errors smaller than a distance threshold, selecting the transformation with the largest number of inner points as the optimal transformation, reserving the matching pair with the residual errors smaller than the threshold under the optimal transformation in the initial super-point matching pair set as a correct super-point matching pair, and eliminating the rest matching pairs.
  8. 8. The two-stage point cloud registration method based on the angle consistency constraint according to claim 1 is characterized in that the specific implementation method of the step 7 is that the correct super-point matching pair reserved in the step 6) is expanded into a local neighborhood block matching pair taking the super-point as a center, the characteristics and coordinates of dense points in the neighborhood block are extracted and input into a dense point matching module to carry out dense point matching, the corresponding relation of the dense points in the neighborhood block is calculated by utilizing weighted singular value decomposition, and the weighted singular value decomposition is carried out again based on the corresponding relation of all the dense points to obtain a final rigid body transformation matrix through solving.

Description

Two-stage point cloud registration method based on angle consistency constraint Technical Field The invention relates to an improvement of a point cloud registration technology, in particular to a method for removing wrong matching pairs by utilizing geometrical constraint of angle consistency so as to improve the registration precision of two-stage point clouds, and belongs to the technical field of three-dimensional computer vision. Background The point cloud registration is a basic technology for realizing three-dimensional reconstruction, automatic driving positioning and robot navigation. The current mainstream registration method mostly adopts a coarse-to-fine two-stage strategy, namely coarse matching of block-level super points (Superpoint) is firstly carried out, and then fine matching of point-level dense points is carried out in the super point neighborhood. The strategy effectively relieves the problem that the traditional method is sensitive to the initial pose. However, when dealing with complex scenes with low overlap rates, weak textures, or repeated geometries, the coarse matching stage tends to produce a large number of false pairs of super-point matches (i.e., outliers). If these outliers cannot be effectively culled, they will mislead the subsequent exact matching process, resulting in registration failure. The existing outlier removal method is mainly divided into a learning-based method and a geometric consistency-based method. Among them, spatial consistency based on Euclidean distance is a common geometric constraint method such as SC2-PCR (A Second Order Spatial Compatibility for EFFICIENT AND Robust Point Cloud Registration). The method assumes that the distance between two points remains unchanged before and after the rigid transformation. However, studies have found that there is a discrimination ambiguity in the distance consistency that in some cases the distance between two incorrect pairs of matching points may be accidentally equal (e.g. incorrect matching in a symmetrical structure) and thus pass the consistency check. Disclosure of Invention Aiming at the problems that the prior art two-stage registration method is easy to generate a large number of mismatching in a rough matching stage and the accuracy of point cloud registration is affected due to the fact that the discrimination ambiguity exists in the outer points is removed based on distance consistency, the invention aims to provide the two-stage point cloud registration method based on angle consistency constraint. The technical scheme of the invention is realized as follows: The two-stage point cloud registration method based on the angle consistency constraint comprises the following steps: 1) Coarse matching, namely inputting a source point cloud and a target point cloud to be registered into a neural network for downsampling and feature extraction, calculating feature correlation between the source point cloud and the target point cloud super points, and selecting a plurality of super point pairs with highest correlation to form an initial super point matching pair set; 2) Constructing a first-order angle similarity matrix, namely introducing a third reference matching point into any two pairs of matching points in the initial super-point matching pair set, and constructing triangles based on three points in a source point cloud and a target point cloud respectively; 3) Performing binarization processing on the first-order angle similarity matrix, counting the number of adjacent matched pairs which are commonly compatible in a global range of each matched pair, and converting local first-order angle consistency into a second-order similarity matrix reflecting global geometric distribution consistency; 4) The seed selection and consensus set construction, namely performing spectrum decomposition on the first-order angle similarity matrix subjected to binarization processing to obtain initial confidence coefficient, screening out high-confidence seed matching pairs with uniform space distribution; 5) Calculating the weight of each point in the local consensus set by using a spectrum matching method, performing weighted singular value decomposition based on the weight, and solving a plurality of candidate rigid body transformation matrixes; 6) Performing residual error check on the initial super-point matching pair set by utilizing the candidate rigid body transformation matrix, screening out the candidate rigid body transformation matrix with the largest number of inner points as the optimal transformation, removing all error super-point matching pairs which do not accord with the transformation from the initial super-point matching pair set according to the optimal transformation, and reserving correct super-point matching pairs; 7) And (3) fine matching, namely expanding the reserved correct super-point matching pair into a neighborhood block, carrying out local dense point matching in the neighborhood block, an