Laser point cloud secondary feature matching method based on geometric consistency

Abstract

The invention discloses a geometric consistency-based laser point cloud secondary feature matching method. The method comprises the following steps of registering laser point clouds of front and rear frames based on a KD-Tree-based nearest laser point matching method; constructing a feature association graph based on the registration relationship, associating each feature point with a nearest neighbor feature, and calculating a residual value by using the feature association graph; minimizing a weighted residual error and solving point cloud initial transformation based on a Gauss-Newton method; applying the initial transformation of the point cloud to the original point cloud to obtain the transformed point cloud, and recalculating the residual parameter of the feature association graph; and finally, based on a geometric consistency thought, comparing residual parameter change conditions in the front and back feature association graphs, and rejecting unreliable point clouds to obtain an accurate laser point cloud registration relationship.

Description

Technical Field

[0001] The present invention relates to the field of laser radar point cloud registration technology, and in particular to a laser point cloud secondary feature matching method based on geometric consistency. Background Art

[0002] LiDAR and other sensors generate high-precision 3D point cloud data in a single scan, which can be used to create 3D maps, object detection and tracking, etc. In these applications, point cloud registration technology that aligns point cloud data obtained from different positions or perspectives is particularly important. LiDAR point cloud registration technology can be divided into point-to-point direct matching method and feature-based feature matching method according to the different matching objects. Point-to-point direct matching methods include the ICP algorithm of iterative nearest neighbor points and the NDT algorithm of normal distribution transformation that represents and registers point clouds through normal distribution. Feature-based feature matching methods use feature points and key points with significant features for matching. This method is more efficient than the direct matching method, but this method is sensitive to initial values ​​and is prone to falling into local optimal solutions, resulting in point cloud matching errors. At the same time, the point cloud data has noise and outliers, which will affect the accuracy of registration. Directly using a high-precision registration method will also significantly increase the computational cost of the registration process due to the large scale of point cloud data, resulting in untimely registration. Therefore, it is crucial to find a method that has high point cloud registration accuracy and does not significantly increase the computational cost. Summary of the invention

[0003] The purpose of the present invention is to provide a laser point cloud secondary feature matching method based on geometric consistency and a point cloud registration method based on geometric consistency to improve the accuracy of point cloud registration.

[0004] To achieve the above object, the present invention provides a laser point cloud secondary feature matching method based on geometric consistency, comprising the following steps:

[0005] Step 1: Align the laser point clouds of the previous and next frames based on the nearest neighbor laser point matching method of KD-Tree;

[0006] Step 2: Construct a feature association graph based on the registration relationship, associate each feature point with the nearest neighbor feature, and use the feature association graph to calculate the residual value;

[0007] Step 3: Minimize the weighted residual and solve the initial transformation of the point cloud based on the Gauss-Newton method;

[0008] Step 4: The initial transformation of the point cloud is applied to the original point cloud to obtain the transformed point cloud and recalculate the residual parameters of the feature association graph;

[0009] Step 5: Based on the idea of ​​geometric consistency, compare the changes in residual parameters in the feature association graphs before and after, eliminate unreliable point clouds, and obtain accurate laser point cloud registration relationships.

[0010] Optionally, during the execution of step 1, the point clouds of the previous and next two frames are firstly distinguished by a smoothness calculation formula to identify different types of feature points, including plane points and edge points. The specific formula is as follows:

[0011]

[0012] in, Indicates i The curvature of a point, Indicates the distance value of the curvature point that needs to be calculated. Indicates at point The distance values ​​of the surrounding 10 points, is a constant.

[0013] Optionally, to further speed up the calculation, the following formula is used instead of the smoothness formula to distinguish feature points:

[0014]

[0015] The highly efficient nearest neighbor search algorithm KD-Tree algorithm is used to initially and quickly align and distinguish the laser point clouds of the front and back frames, and to correspond plane points to plane points and edge points to edge points.

[0016] Optionally, during the execution of step 2, the corresponding residual value is calculated by using the coarse matching registration relationship through the feature residual calculation formula, the matching relationship is associated with the corresponding residual value, and a feature association graph is constructed. The specific formula is as follows:

[0017]

[0018] in, , are the residual values ​​of the plane points and edge points respectively, , , , They are the distances from plane points and edge points to the lines formed by their respective features.

[0019] Optionally, during the execution of step 3, the Gauss-Newton method is used as an optimization algorithm to minimize the sum of the residuals to obtain the pose transformation of the feature point cloud. The specific formula is as follows:

[0020]

[0021] in, represents the total pose change, represents the rotation matrix, Represents the translation matrix.

[0022] Optionally, during the execution of step 4, the point cloud pose transformation calculated in claim 4 is applied to the initial point cloud to obtain the transformed point cloud, and the residual calculation formula obtained in claim 3 is reused to calculate the corresponding residual value in the feature association graph after the transformation.

[0023] Optionally, during the execution of step 5, the idea of ​​geometric consistency is used to define the laser point cloud as a rigid transformation, and the new residual value is compared with the original residual value transformation obtained by rough matching. Most of the point clouds are consistent changes with reduced residual values. A small number of inconsistent feature matching pairs with increased residual values ​​are eliminated to obtain more accurate laser point cloud feature matching pairs.

[0024] The invention provides a laser point cloud secondary feature matching method based on geometric consistency, which aligns the laser point clouds of the front and rear frames based on the nearest neighbor laser point matching method of KD-Tree; then, a feature association graph is constructed based on the registration relationship, each feature point is associated with the nearest neighbor feature, and the residual value is calculated using the feature association graph; then, the weighted residual is minimized and the initial transformation of the point cloud is solved based on the Gauss-Newton method; then, the initial transformation of the point cloud is applied to the original point cloud to obtain the transformed point cloud, and the residual parameters of the feature association graph are recalculated; finally, based on the idea of ​​geometric consistency, the changes of the residual parameters in the front and rear feature association graphs are compared, the unreliable point clouds are eliminated, and the accurate laser point cloud registration relationship is obtained. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0026] Figure 1 It is a flow chart of a laser point cloud secondary feature matching method based on geometric consistency of the present invention.

[0027] Figure 2 It is a schematic diagram of the plane point residual calculation of the present invention.

[0028] Figure 3 It is a schematic diagram of edge point residual calculation of the present invention. DETAILED DESCRIPTION

[0029] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0030] The following are some nouns and terminology descriptions of the present invention:

[0031] KD-Tree: K-Dimensional Tree, fast search of K-dimensional tree.

[0032] See also Figure 1 The object of the present invention is to provide a laser point cloud secondary feature matching method based on geometric consistency, comprising the following steps:

[0033] Step 101, registering the laser point clouds of the front and back frames using the nearest neighbor laser point matching method based on KD-Tree;

[0034] Step 102: construct a feature association graph based on the registration relationship, associate each feature point with the nearest neighbor feature, and calculate the residual value using the feature association graph;

[0035] Step 103, minimizing the weighted residual and solving the initial transformation of the point cloud based on the Gauss-Newton method;

[0036] Step 104, the initial transformation of the point cloud is applied to the original point cloud to obtain the transformed point cloud, and the residual parameters of the feature association graph are recalculated;

[0037] Step 105: Based on the idea of ​​geometric consistency, the residual parameter changes in the feature association graphs before and after are compared, unreliable point clouds are eliminated, and accurate laser point cloud registration relationships are obtained.

[0038] The following is a further explanation based on the specific implementation steps:

[0039] In step 101, specifically, firstly, the point clouds of the two frames before and after are distinguished into different types of feature points using a smoothness calculation formula, including plane points and edge points. The specific formula is as follows:

[0040]

[0041] in, Indicates i The curvature of a point, Indicates the distance value of the curvature point that needs to be calculated. Indicates at point The distance values ​​of the surrounding 10 points, is a constant.

[0042] To further speed up the calculation, the following formula is used instead of the smoothness formula to distinguish feature points:

[0043]

[0044] The highly efficient nearest neighbor search algorithm KD-Tree algorithm is used to initially and quickly align and distinguish the laser point clouds of the front and back frames, and to correspond plane points to plane points and edge points to edge points.

[0045] In step 102, specifically, the corresponding residual value is calculated by using the registration relationship of the rough matching through the feature residual calculation formula, the matching relationship is associated with the corresponding residual value, and a feature association graph is constructed. The specific formula is as follows:

[0046]

[0047] in, , are the residual values ​​of the plane points and edge points respectively, , , , They are the distances from plane points and edge points to the lines formed by their respective features.

[0048] In step 103, specifically, the optimization algorithm Gauss-Newton method is used to minimize the sum of the residuals to obtain the feature point cloud pose transformation. The specific formula is as follows:

[0049]

[0050] in, represents the total pose change, represents the rotation matrix, Represents the translation matrix.

[0051] In step 104, specifically, the point cloud pose transformation calculated in step S3 is applied to the initial point cloud to obtain the transformed point cloud, and the residual calculation formula obtained in step S2 is reused to calculate the corresponding residual value in the feature association graph after the transformation.

[0052] Specifically, in step 105, the idea of ​​geometric consistency is used to define the laser point cloud as a rigid transformation, and the new residual value is compared and calculated with the original residual value transformation obtained by rough matching. Most of the point clouds are consistent changes with reduced residual values. A small number of inconsistent feature matching pairs with increased residual values ​​are eliminated to obtain more accurate laser point cloud feature matching pairs.

[0053] What is disclosed above is only a preferred embodiment of the present invention, and it certainly cannot be used to limit the scope of rights of the present invention. Ordinary technicians in this field can understand that all or part of the processes of the above embodiment and equivalent changes made according to the claims of the present invention still fall within the scope of the invention.

Claims

1. A laser point cloud secondary feature matching method based on geometric consistency, characterized in that: The following steps are involved: Step 1: Align the laser point clouds of the previous and next frames based on the nearest neighbor laser point matching method of KD-Tree; Step 2: Construct a feature association graph based on the registration relationship, associate each feature point with the nearest neighbor feature, and use the feature association graph to calculate the residual value; Step 3: Minimize the weighted residual and solve the initial transformation of the point cloud based on the Gauss-Newton method; Step 4: The initial transformation of the point cloud is applied to the original point cloud to obtain the transformed point cloud and recalculate the residual parameters of the feature association graph; Step 5: Based on the idea of ​​geometric consistency, compare the changes in residual parameters in the feature association graphs before and after, eliminate unreliable point clouds, and obtain accurate laser point cloud registration relationships.

2. The laser point cloud secondary feature matching method based on geometric consistency according to claim 1, characterized in that: During the execution of step 1, the point clouds of the previous and next two frames are firstly distinguished by the smoothness calculation formula to distinguish different types of feature points, including plane points and edge points. The specific formula is as follows: Among them, it means The curvature of a point, Indicates the distance value of the curvature point that needs to be calculated. Indicates at point The distance values ​​of the surrounding 10 points, is a constant; Then, the highly efficient nearest neighbor search algorithm KD-Tree algorithm is used to preliminarily and quickly align and distinguish the laser point clouds of the front and back frames, and to correspond plane points to plane points and edge points to edge points.

3. The laser point cloud secondary feature matching method based on geometric consistency according to claim 1, characterized in that: During the execution of step 2, the corresponding residual value is calculated by using the coarse matching registration relationship through the feature residual calculation formula, the matching relationship is associated with the corresponding residual value, and a feature association graph is constructed.

4. The laser point cloud secondary feature matching method based on geometric consistency according to claim 1, characterized in that: During the execution of step 3, the optimization algorithm Gauss-Newton method is used to minimize the sum of the residuals to obtain the pose transformation of the feature point cloud.

5. The laser point cloud secondary feature matching method based on geometric consistency according to claim 1, characterized in that: During the execution of step 4, the point cloud pose transformation calculated in claim 4 is applied to the initial point cloud to obtain the transformed point cloud, and the residual calculation formula obtained in claim 3 is reused to calculate the corresponding residual value in the transformed feature association graph.

6. The laser point cloud secondary feature matching method based on geometric consistency according to claim 1, characterized in that: During the execution of step 5, the idea of ​​geometric consistency is used to define the laser point cloud as a rigid transformation. The new residual value is compared with the original residual value transformation obtained by rough matching. Most of the point clouds are consistent changes with reduced residual values. A small number of inconsistent feature matching pairs with increased residual values ​​are eliminated to obtain more accurate laser point cloud feature matching pairs.

Images