Uniform feature sampling method for 3D point cloud registration

Abstract

The purpose of the present invention is to provide a uniform feature sampling method for three-dimensional point cloud registration, which first arranges the point cloud according to the laser line number and then performs uniform sampling; the sampled point cloud is subjected to local feature extraction and global feature extraction respectively, and then spliced to obtain multi-scale features. According to the characteristic that feature extraction aggregates domain information, the multi-scale features are also uniformly sampled. The uniformly sampled multi-scale features are input into the cross-attention module to extract the change features, and then the transformation features are embedded into the sampled multi-scale features. A matching matrix is constructed based on the difference in feature space to represent the correspondence between each point between the point clouds, and a roughly aligned virtual point cloud is obtained from the matching matrix and the sampled target point cloud. Subsequently, the rigid transformation is estimated by weighted SVD. After solving the transformation, the transformation is applied to the source point cloud after point cloud sampling, and the above operation is repeated for iteration.

Description

Technical Field

[0001] The present invention belongs to the technical field of three-dimensional point cloud registration, and in particular relates to a uniform feature sampling method for three-dimensional point cloud registration. Background Art

[0002] With the rapid development of LiDAR (light detection and ranging), point clouds have become a primary data format for representing the three-dimensional world. Point cloud registration technology, which transforms point cloud data acquired by point cloud scanning devices from different perspectives into a common coordinate system through rotation and translation, is widely used in fields such as 3D reconstruction and mobile robotics.

[0003] LiDAR odometry calculation methods can be roughly divided into two categories: matching-based methods and feature-based methods. Matching-based methods are computationally intensive and lack robustness in the face of noise interference, large pose changes, and object occlusion. Therefore, existing odometry methods mostly use feature-based methods.

[0004] Currently, feature extraction for point clouds is generally divided into two types: manually extracted features and deep learning-based features. With the advancement of deep learning, most feature-based methods now utilize features extracted through deep learning. However, due to the large number of LiDAR point clouds and limited computing resources, most deep learning-based methods require data sampling for large scene point clouds. Therefore, extracting more information from limited data is crucial to algorithm accuracy. Summary of the Invention

[0005] This paper focuses on the research of large-scale point cloud registration algorithms. Based on the data characteristics of LiDAR point clouds and the actual needs of application fields, a uniform feature sampling method for 3D point cloud registration is designed. This method reduces computing resource consumption and improves algorithm accuracy.

[0006] In this scheme, the point cloud is first arranged according to the laser line number, and then uniformly sampled; the sampled point cloud is subjected to local feature extraction and global feature extraction respectively, and then spliced to obtain multi-scale features. According to the characteristic that feature extraction aggregates domain information, the multi-scale features are also uniformly sampled. The uniformly sampled multi-scale features are input into the cross-attention module to extract the change features, and then the transformation features are embedded into the sampled multi-scale features. A matching matrix is constructed based on the difference in feature space to represent the correspondence between each point between the point clouds, and a roughly aligned virtual point cloud is obtained from the matching matrix and the sampled target point cloud. Subsequently, the rigid transformation is estimated by weighted SVD. After solving the transformation, the transformation is applied to the source point cloud after point cloud sampling, and the above operation is repeated for iteration. The present invention reduces the consumption of computing resources and improves the accuracy of the algorithm.

[0007] The technical solution specifically adopted by the present invention to solve the technical problem is:

[0008] A uniform feature sampling method for three-dimensional point cloud registration, characterized by comprising the following steps:

[0009] Step S1: Arrange the source point cloud and the target point cloud according to the laser line number, and then perform uniform sampling;

[0010] Step S2: extract local features and global features from the sampled point cloud, and then splice them to obtain multi-scale features;

[0011] Step S3: Based on the characteristic of feature extraction that aggregates domain information, the multi-scale features are uniformly sampled; the uniformly sampled multi-scale features are input into the cross attention module to extract the change features, and then the transformed features are embedded into the sampled multi-scale features;

[0012] Step S4: Construct a feature matching matrix based on the differences in the feature space to represent the correspondence between each point in the point cloud. The virtual point cloud for this iterative alignment is obtained from the matching matrix and the sampled target point cloud. Subsequently, the rigid transformation is estimated by weighted SVD in the source point cloud branch.

[0013] Step S5: After the source point cloud branch solves the transformation in step S4, the transformation is applied to the source point cloud after point cloud sampling, and then the operations of steps S2 to S4 are repeated iteratively until the registration result is obtained.

[0014] Furthermore, step S1 specifically includes the following steps:

[0015] Step S11: Point cloud For each point, use the coordinate information of the point to calculate and get the laser line number to which the point belongs; where R 3 Represents three-dimensional space;

[0016] Step S12: rearrange the point cloud in order according to the laser line number;

[0017] Step S13: Calculate the quotient of the number of original point clouds N and the number of point clouds input to the network M, and round down to obtain the sampling interval K;

[0018] Step S14: According to the sampling interval, the arranged point cloud is uniformly sampled at equal intervals to obtain

[0019] Step S15: Calculate the difference Q between the number of sampled point clouds N / K and the number of point clouds input to the network M;

[0020] Step S16: Remove the redundant Q points to obtain the sampled point cloud.

[0021] Furthermore, step S2 specifically includes:

[0022] Step S21: Send the sampled point cloud X′ to the global feature extraction network and the local feature extraction network respectively to obtain F Global and F Loacal ;

[0023] Step S22: F Global and F Loacal Splicing to obtain multi-scale features

[0024] Furthermore, in step S3, the multi-scale features are uniformly sampled as follows: according to the hyperparameter λ, the multi-scale features are uniformly sampled to obtain The multi-scale features after uniform sampling are input into the cross attention module as follows: the sampled features of the source point cloud and the target point cloud are respectively fed into the cross attention module to obtain the transformed features Φ(S) and Φ(T); the transformed features are then embedded into the sampled multi-scale features, , = to obtain

[0025] Furthermore, in step S4 and step S5:

[0026] The matching matrix is calculated using the two-frame features corresponding to the source point cloud and the target point cloud, and a relatively certain virtual point cloud is obtained from the matching matrix and the subsampled target point cloud. Finally, the rigid transformation R between the subsampled source point cloud and the virtual point cloud is estimated by weighted SVD. i , t i , where i represents the current iteration number;

[0027] Then use the R calculated last time i-1 , t i-1Act on the source point cloud after point cloud sampling to iterate;

[0028] The subsampling refers to subsampling the point cloud X′ according to the hyperparameter λ to obtain X′ ds .

[0029] Compared with the existing technology, the present invention and its preferred solution greatly reduce the consumption of computing resources and can further improve the algorithm accuracy through multiple iterations. Compared with the existing point cloud registration solution, it has significant advantages. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0031] Figure 1 1 is a schematic diagram of the overall flow of the method in an embodiment of the present invention.

[0032] Figure 2 4 is a flow chart of point cloud sampling in an embodiment of the present invention.

[0033] Figure 3 4 is a flow chart of feature sampling in an embodiment of the present invention. DETAILED DESCRIPTION

[0034] To make the features and advantages of this patent more clearly understood, the following embodiments are specifically described in detail as follows:

[0035] It should be noted that the following detailed description is illustrative and is intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used in this specification have the same meaning as commonly understood by those skilled in the art to which this application belongs.

[0036] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0037] like Figure 1 FIG. 1 is a flow chart of a uniform feature sampling method for 3D point cloud registration provided by the present invention. The method includes the following steps:

[0038] Step S1: Arrange the source point cloud and the target point cloud according to the laser line number, and then perform uniform sampling;

[0039] Step S2: extract local features and global features from the sampled point cloud, and then splice them to obtain multi-scale features;

[0040] Step S3: Based on the characteristic of feature extraction that aggregates domain information, the multi-scale features are uniformly sampled; the uniformly sampled multi-scale features are input into the cross attention module to extract the change features, and then the transformed features are embedded into the sampled multi-scale features;

[0041] Step S4: Construct a feature matching matrix based on the differences in the feature space to represent the correspondence between each point in the point cloud. The virtual point cloud for this iterative alignment is obtained from the matching matrix and the sampled target point cloud. Subsequently, the rigid transformation is estimated by weighted SVD in the source point cloud branch.

[0042] Step S5: After the source point cloud branch solves the transformation in step S4, the transformation is applied to the source point cloud after point cloud sampling, and then the operations of steps S2 to S4 are repeated iteratively until the registration result is obtained.

[0043] The process of point cloud sampling is as follows: Figure 2 As shown, the following steps are included:

[0044] Step E1: Point cloud For each point, use the coordinate information of the point to calculate and get the laser line number to which the point belongs;

[0045] Step E2: rearrange the point cloud in order according to the laser line number;

[0046] Step E3: Calculate the quotient of the number of original point clouds N and the number of point clouds input to the network M, and round down to obtain the sampling interval K;

[0047] Step E4: Perform uniform sampling of the arranged point cloud at equal intervals according to the sampling interval to obtain

[0048] Step E5, calculating the difference Q between the number of sampled point clouds N / K and the number of point clouds input to the network M;

[0049] Step E6: Remove the redundant Q points and get

[0050] Step E7: Subsample the point cloud X′ according to the hyperparameter λ to obtain X′ ds .

[0051] The process of feature sampling is as follows Figure 3 As shown:

[0052] Step M1: Send the sampled point cloud X′ to the global feature extraction network and the local feature extraction network respectively to obtain F Globaland F Loacal ;

[0053] Step M2: F Global and F Loacal Splicing to obtain multi-scale features

[0054] Step M3: According to the hyperparameter λ, the multi-scale features are sampled at equal intervals to obtain

[0055]

[0056] Step M4: Send the sampled features of the two frames of point clouds to the cross attention module respectively to obtain the transformed features Φ(S) and Φ(T);

[0057] Step M5: embed the change feature into the multi-scale feature after sampling to obtain

[0058] The specific execution process of step S4 and step S5 generally includes the following:

[0059] Step F1: Use the features of the two frames to calculate the matching matrix, and obtain a relatively certain virtual point cloud from the matching matrix and the subsampled target point cloud. Finally, use weighted SVD to estimate the rigid transformation R between the subsampled source point cloud and the virtual point cloud. i , t i , where i represents the current iteration number;

[0060] Step F2: Use the R obtained in the previous calculation i-1 , t i-1 Acting on the source point cloud after point cloud sampling, the subsequent steps are the same and iterative.

[0061] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code.

[0062] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the steps in the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0063] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0064] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0065] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other manner. Any person skilled in the art may utilize the above-disclosed technical content to modify or modify the present invention into equivalent embodiments. However, any simple modifications, equivalent variations, and modifications to the above embodiments that do not depart from the technical content of the present invention and are based on the technical essence of the present invention remain within the scope of protection of the present invention.

[0066] This patent is not limited to the above-mentioned optimal implementation method. Anyone can derive various other forms of uniform feature sampling methods for three-dimensional point cloud registration based on the inspiration of this patent. All equal changes and modifications made within the scope of the patent application of this invention should be covered by this patent.

Claims

1. A uniform feature sampling method for 3D point cloud registration, characterized in that: The following steps are involved: Step S1: Arrange the source point cloud and the target point cloud according to the laser line number, and then perform uniform sampling; Step S2: extract local features and global features from the sampled point cloud, and then splice them to obtain multi-scale features; Step S3: Based on the characteristic of feature extraction that aggregates domain information, the multi-scale features are uniformly sampled; the uniformly sampled multi-scale features are input into the cross attention module to extract the change features, and then the transformation features are embedded into the sampled multi-scale features; Step S4: Construct a feature matching matrix based on the differences in the feature space to represent the correspondence between each point in the point cloud. The virtual point cloud for this iterative alignment is obtained from the matching matrix and the sampled target point cloud. Subsequently, the rigid transformation is estimated by weighted SVD in the source point cloud branch. Step S5: After the source point cloud branch solves the transformation in step S4, the transformation is applied to the source point cloud after point cloud sampling, and then the operations of steps S2 to S4 are repeated iteratively until the registration result is obtained; Step S1 specifically includes the following steps: Step S11: Point cloud For each point, use the coordinate information of the point to calculate and get the laser line number to which the point belongs; where R 3 Represents three-dimensional space; Step S12: rearrange the point cloud in order according to the laser line number; Step S13: Calculate the quotient of the number of original point clouds N and the number of point clouds input to the network M, and round down to obtain the sampling interval K; Step S14: According to the sampling interval, the arranged point cloud is uniformly sampled at equal intervals to obtain Step S15: Calculate the difference Q between the number of sampled point clouds N / K and the number of point clouds input to the network M; Step S16: Remove the redundant Q points to obtain the sampled point cloud. Step S2 specifically includes: Step S21: Send the sampled point cloud X′ to the global feature extraction network and the local feature extraction network respectively to obtain F Global and F Loacal ; Step S22: F Global and F Loacal Splicing to obtain multi-scale features In step S3, the multi-scale features are uniformly sampled as follows: according to the hyperparameter λ, the multi-scale features are sampled at equal intervals to obtain The uniformly sampled multi-scale features are input into the cross attention module as follows: the sampled features of the source point cloud and the target point cloud are respectively fed into the cross attention module to obtain the transformed features Φ(S) and Φ(T); the transformed features are then embedded into the sampled multi-scale features to obtain In steps S4 and S5: The matching matrix is calculated using the two-frame features corresponding to the source point cloud and the target point cloud, and the virtual point cloud is obtained by the matching matrix and the subsampled target point cloud. Finally, the rigid transformation R between the subsampled source point cloud and the virtual point cloud is estimated by weighted SVD. i , t i , where i represents the current iteration number; Then use the R calculated last time i-1 , t i-1 Act on the source point cloud after point cloud sampling to iterate; The subsampling refers to subsampling the point cloud X′ according to the hyperparameter λ to obtain X′ ds .

Images