CN-122023470-A - Super-point matching pair geometric optimization method in point cloud registration and point cloud registration method

CN122023470ACN 122023470 ACN122023470 ACN 122023470ACN-122023470-A

Abstract

The invention discloses a geometric optimization method of super-point matching pairs in point cloud registration and a point cloud registration method, and belongs to the technical field of three-dimensional computer vision and point cloud processing. And then, accurately reasoning an overlapping region from the original point cloud by using the expanded inner point set as a seed through a region growing algorithm, and resampling the region to improve the overlapping quality of the super-point neighborhood blocks. According to the invention, through the collaborative optimization of the interior point generation and the overlap region resampling, the double improvement of the pair number and the quality of the super-point matching is realized, and the registration success rate and the precision of the low-overlap point cloud can be obviously improved.

Inventors

YANG LIPING
YANG ZHIHUI
HUANG HONG

Assignees

重庆大学

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (7)

1. A geometric optimization method of super-point matching pairs in point cloud registration is characterized by comprising the following steps: 1) For each candidate super point in the initial super point matching pair set in the source point cloud, which is not covered by the initial inner point subset, calculating the distance vector between the candidate super point and all the inner points in the source point cloud, and simultaneously calculating the distance vector between each super point in the initial super point matching pair set in the target point cloud, which is not covered by the initial inner point subset, and all the inner points in the target point cloud; If the minimum distance difference is smaller than a preset distance error threshold, judging that the candidate superpoints in the source point cloud are correctly matched with the matching points with the minimum distance difference in the target point cloud and form candidate pairs, and adding all the candidate pairs obtained by matching into an initial interior point subset to obtain an expanded interior point set; 2) And deducing an overlapping region from the source point cloud and the target point cloud by using the extended inner point set as a seed point through a region growing algorithm, and then downsampling the overlapping region to generate a new super point set and a new dense point set.
2. The method for optimizing geometry of a super point matching pair in point cloud registration as recited in claim 1, wherein step 1) is performed by obtaining said extended set of interior points as follows, 1.1 For a candidate superpoint in the initial set of superpoint matching pairs in the source point cloud that is not covered by the initial interior point subset Calculate it to an initial interior point subset in the source point cloud All interior points of (3) To obtain a distance vector: 1.2 For each superpoint in the set of initial superpoint matching pairs in the target point cloud that is not covered by the initial interior point subset Respectively calculating initial inner point subsets of the target point cloud All interior points of (3) Is a distance vector of (a): All are put together Arranged in a matrix: 1.3 Calculation of (c) Each row and Differences of (2) The difference matrix is obtained as follows: 1.4 For each row vector, calculate its euclidean norm, resulting in an error matrix: selecting the value with the smallest Euclidean norm And candidate superpoints in the corresponding target point cloud: Will be And a distance error threshold Comparing if And judging that the candidate super point in the source point cloud is correctly matched with the matching point with the minimum Euclidean norm in the target point cloud, forming candidate pairs, and adding all the candidate pairs obtained by matching into an initial inner point subset, thereby obtaining an expanded inner point set.
3. The method for optimizing geometry of a super-point matching pair in point cloud registration as recited in claim 2, wherein said distance error threshold is set to be equal to The dynamic setting is carried out by the method, Indicating the number of pairs of super point matches.
4. The method for optimizing geometry of a super point matching pair in point cloud registration according to claim 1, wherein the specific process of step 2) is as follows, 2.1 Using all the super points in the extended inner point set as initial seed points, using a region growing algorithm to diffuse in a source point cloud and target point cloud space, and polymerizing to obtain a complete overlapping region point set; 2.2 Sampling the overlapping region point set by adopting a uniform voxel downsampling method to generate a new super point set and a new dense point set with higher neighborhood overlapping degree; 2.3 Assigning the nearest new super point for each new dense point in the dense point set, and reestablishing the neighborhood block association relation of dense point-super point.
5. The method for optimizing geometry of a super point matching pair in point cloud registration as recited in claim 4, wherein the specific process of step 2.1) is as follows, Firstly, taking the whole super point set of the source point cloud as a search space, and taking all super points of the extended inner point set at one side of the source point cloud as an initial seed point set; the method comprises the steps of initializing, adding all unviewed seed points into a final overlapping area point set and a boundary point set for iterative expansion at the same time, then entering an iterative growth stage, traversing each point in a current boundary point set, taking the point as a center, searching all adjacent points within a set neighborhood distance threshold, adding each adjacent point into the overlapping area point set if the adjacent point is not visited, simultaneously adding the adjacent point into the boundary point set as a new starting point of next round expansion, continuing the process, continuously pushing the boundary outwards until the boundary point set is empty, indicating that a new expandable point cannot be found under the constraint of the preset neighborhood distance threshold, and ending the iteration at the moment, and outputting the overlapping area point set finally obtained by reasoning by an algorithm.
6. The method of claim 5, wherein the neighborhood distance threshold is set to a voxel size at which the downsampling forms the super-points.
7. A point cloud registration method is characterized by obtaining an initial super-point matching pair of a source point cloud and a target point cloud, then optimizing the initial super-point matching pair by using the super-point matching pair geometry optimization method according to any one of claims 1 to 6, and executing a point cloud registration algorithm based on a new super-point set and a dense point set generated after optimization to obtain a final rigid body transformation matrix.

Description

Super-point matching pair geometric optimization method in point cloud registration and point cloud registration method Technical Field The invention relates to an improvement of a point cloud registration technology, in particular to a geometric optimization method of super-point matching pairs in point cloud registration and a point cloud registration method, and belongs to the technical field of three-dimensional computer vision. Background Point cloud registration is a core task in the three-dimensional vision field, and aims to align a plurality of point clouds to the same coordinate system through feature extraction, matching and pose estimation, wherein the accuracy and the robustness of the point cloud are directly related to the performance of downstream application. The technology has wide and key application value in the fields of large-scale scene imaging reconstruction, high-precision detection of industrial small-size parts, automatic driving environment sensing, medical image fusion, remote sensing monitoring and the like. In practical application, three-dimensional point clouds are mostly collected by scanning from different visual angles, and only partial areas overlap. According to the space structure, the point cloud can be divided into overlapping area points and non-overlapping area points, wherein the overlapping area generally corresponds to a continuous point set on the surface of the same object to form an effective basis for registration, and the non-overlapping area is derived from an object with shielding, noise or unique visual angle and belongs to interference information. The interference of non-overlapping areas is particularly remarkable in low-overlapping-rate scenes such as large-range scene imaging or tiny industrial part scanning, and the matching process is seriously interfered by a large number of irrelevant structures, so that accurate registration faces a great challenge. In a real scene with low overlapping rate, the initial super-point matching pair set generated by the existing various registration methods has double defects of quantity and quality, so that the performance is drastically reduced. The concrete steps are as follows: The correct matching is scarce in number, because of the limited view angle of the sensor, severe occlusion and scene structure characteristics, the effective overlap area between the two point clouds is itself narrow, which fundamentally limits the number of correct matching pairs that may exist. Any matching method inevitably generates a large number of mismatching under the condition, and sparse correct matching cannot provide robust and sufficient geometric constraint for subsequent pose estimation. The local neighborhood overlap is too low, namely even a few correctly matched super-point pairs are positioned between local neighborhood blocks, and the effective overlap area is often seriously insufficient due to the point cloud discreteness, downsampling deviation and interference of non-overlap areas. Dense point matching is performed in the local area with low overlapping degree, so that ambiguity is easy to generate, and the local features are difficult to establish an accurate one-to-one correspondence, so that the matching error of the front end is propagated and amplified to a final registration result. Currently, efforts by the research community to address the above issues have mostly focused on designing more complex network architectures or more powerful feature descriptors in order to obtain better initial results at the matching front-end. However, such methods tend to be computationally expensive and fail to fundamentally solve the problem inherent to the initial matching set in low overlap scenarios. The prior art obviously lacks a post-processing mechanism capable of systematically correcting and optimizing a generated initial matching pair set, and particularly can simultaneously realize a collaborative optimization scheme of increasing the correct matching number and improving the matching pair local neighborhood quality. Therefore, the geometrical correction method which does not depend on a specific front-end matcher and can universally optimize the initial matching pair set is developed, and has important research significance and engineering value for improving the robustness and precision of the point cloud registration technology in complex real scenes, especially the application of large-range dynamic scene imaging and high-precision industrial microscale detection. Disclosure of Invention Aiming at the defects of the prior art, the invention aims to provide a geometric optimization method of super-point matching pairs based on point cloud registration, which can solve the problem of registration performance degradation caused by insufficient number of correct matching pairs and low overlapping degree of neighborhood blocks in a low-overlapping scene in the prior art, and remarkably improves registration preci