Railway wide-range inspection-oriented strong generalization point cloud registration method and system

Abstract

The application provides a strong generalization point cloud registration method and system for railway wide-range inspection, and belongs to the technical field of image data processing. The registration method can meet the case that the railway and surrounding environment point cloud data set collected in practice is insufficient for training, and directly use the trained model for feature extraction, so as to achieve high registration accuracy. Through the point cloud registration technology, the data obtained through the previous inspection is compared and analyzed with the reference data, so that the monitoring of the railway and the surrounding equipment can be realized. Moreover, the fusion of multiple railway point cloud data can obtain complete railway line information and surrounding environment information, and provide the optimal route for manual inspection, hidden danger elimination and fault elimination. Therefore, the point cloud registration technology has great potential in the railway inspection process.

Description

Technical Field

[0001] This invention relates to the field of image data processing technology, specifically to a highly generalized point cloud registration method and system for large-scale railway inspection. Background Technology

[0002] Long-distance railway lines, under natural conditions such as temperature, precipitation, and natural disasters, or human factors (such as wear and tear, and derailment), can develop safety hazards and even cause accidents. To ensure transportation safety and the normal use of the tracks, railway lines often require regular inspections. Currently, railway maintenance and management mainly rely on manual inspections or video recordings from onboard or airborne cameras, followed by manual review. However, railway lines are often large in scale and long in length, making long-distance closed inspections impossible; manual inspections are often time-consuming and inefficient; and the increased labor costs also raise railway maintenance and management expenses. Due to the needs of modern development, large-scale railway daily inspections are beginning to adopt methods such as drone remote sensing and robotic identification for rapid sampling, which minimizes or eliminates track occupation and has the advantage of less impact on normal line operation. After sampling, the data is analyzed to identify anomalies and damage types, completing the inspection. Summary of the Invention

[0003] The purpose of this invention is to provide a highly generalized point cloud registration method and system for large-scale railway inspection, so as to solve at least one of the technical problems existing in the background art.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] In a first aspect, the present invention provides a strongly generalized point cloud registration method for large-scale railway inspection, comprising:

[0006] Acquire point cloud data of the railway and its surrounding environment to form a dataset;

[0007] Read point cloud data from different time points in the dataset and construct source point cloud 1 and source point cloud 2;

[0008] Using source point cloud 2 as target point cloud, denoising and filtering preprocessing are performed on source point cloud 1 and target point cloud;

[0009] The source point cloud 1 and the target point cloud are input into a pre-constructed deep learning-based feature learning network. By adjusting the parameter settings in this network, parameters suitable for processing large-scale railway inspection point cloud data are obtained, ultimately acquiring point cloud feature descriptors with strong generalization capabilities. The deep learning-based feature learning network includes a spatial point rotation module, a mapping process, and a feature extraction module. The spatial point rotation module eliminates the point cloud's degrees of freedom about the Z-axis and about the XY plane through rotation. The mapping process maps the local spherical point cloud into a cylindrical point cloud. The feature extraction module extracts features from the source point cloud 1 and the target point cloud after eliminating degrees of freedom.

[0010] Based on the extracted features, the transformation parameters between the final source point cloud and the target point cloud are obtained using the random sampling consensus algorithm.

[0011] The registration accuracy is measured based on the obtained rigidity change parameters to meet the point cloud registration requirements during large-scale inspection.

[0012] Furthermore, using source point cloud 2 as target point cloud, denoising and filtering preprocessing are performed on source point cloud 1 and target point cloud, including: removing useless noise points in source point cloud 1 and target point cloud; performing statistical filtering on source point cloud 1 and target point cloud; and determining whether to randomly select key points in source point cloud 1 and target point cloud based on the number of points in the point cloud obtained after voxel filtering, finally obtaining the key point cloud of source point cloud 1 and key point cloud of target point cloud.

[0013] Furthermore, obtain the coordinates of all points in the source point cloud 1 key point cloud, and perform a nearest neighbor search in the source point cloud 1 with the key point as the center, retaining points within a certain range around the key point; obtain the coordinates of all points in the target point cloud key point cloud, and perform a nearest neighbor search in the target point cloud with the key point as the center, retaining points within a certain range around the key point.

[0014] Furthermore, the process of the spatial point rotation module includes: rotating the normal vectors of the local spherical point clouds in the source point cloud 1 and the target point cloud to align with the Z-axis; dividing the rotated spherical point cloud into voxels according to the radial distance J, elevation angle K, and azimuth angle L; performing a nearest neighbor search again with the voxel center of each voxel as the sphere center, retaining a certain number of points; and rotating the center of each voxel around the Z-axis to align with the YZ plane.

[0015] Furthermore, the feature extraction module process includes: using a weight-shared multilayer perceptron and a max pooling function to extract shallow features from the local cylindrical point clouds of the source point cloud 1 and the target point cloud; and using cylindrical convolution to extract deep features from the extracted features.

[0016] Furthermore, rotation error and translation error are used to measure registration accuracy.

[0017] Secondly, this invention provides a highly generalized point cloud registration system for large-scale railway inspection, comprising:

[0018] The acquisition module is used to acquire point cloud data of the railway and its surrounding environment to form a dataset;

[0019] The building module is used to read point cloud data from different time points in the dataset and construct the source point cloud;

[0020] The preprocessing module is used to take the source point cloud 2 as the target point cloud and perform noise reduction and filtering preprocessing on the source point cloud 1 and the target point cloud.

[0021] The extraction module is used to input the source point cloud 1 and the target point cloud into a pre-constructed deep learning-based feature learning network. By adjusting the parameter settings in this network, parameters suitable for processing large-scale railway inspection point cloud data are obtained, ultimately acquiring point cloud feature descriptors with strong generalization capabilities. The deep learning-based feature learning network includes a spatial point rotation module, a mapping process, and a feature extraction module. The spatial point rotation module eliminates the point cloud's degrees of freedom about the Z-axis and about the XY plane through rotation. The mapping process maps the local spherical point cloud into a cylindrical point cloud. The feature extraction module extracts features from the source point cloud 1 and the target point cloud after eliminating degrees of freedom.

[0022] The calculation module is used to obtain the transformation parameters between the final source point cloud and the target point cloud using a random sampling consistency algorithm based on the extracted features; the registration accuracy is measured based on the obtained rigid change parameters to complete the point cloud registration requirements in the large-scale inspection process.

[0023] Thirdly, the present invention provides a non-transitory computer-readable storage medium for storing computer instructions, which, when executed by a processor, implement the strongly generalized point cloud registration method for large-scale railway inspection as described in the first aspect.

[0024] Fourthly, the present invention provides a computer device including a memory and a processor, wherein the processor and the memory communicate with each other, the memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the strongly generalized point cloud registration method for large-scale railway inspection as described in the first aspect.

[0025] Fifthly, the present invention provides an electronic device, comprising: a processor, a memory, and a computer program; wherein the processor is connected to the memory, the computer program is stored in the memory, and when the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions for implementing the strongly generalized point cloud registration method for large-scale railway inspection as described in the first aspect.

[0026] The beneficial effects of this invention are as follows: Point cloud data of railways and their surrounding environment is acquired using an airborne lidar from an unmanned aerial vehicle (UAV). Source and target point clouds are obtained through preprocessing of the point clouds. The railway point clouds are then input into a pre-trained model that is completely unknown to the railway point clouds, enabling feature extraction. The extracted features are then processed using a random sample consensus algorithm to obtain the rotation and translation matrices between the source and target point clouds. This method exhibits rotation invariance and good robustness, making it suitable for point cloud registration on railway point cloud datasets.

[0027] The advantages of additional aspects of the invention will be set forth more clearly in the following description or will be learned by practice of the invention. Attached Figure Description

[0028] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart illustrating the strongly generalized point cloud registration method for large-scale railway inspection as described in an embodiment of the present invention.

[0030] Figure 2 This is a visual comparison chart of the registration results described in the embodiments of the present invention. Detailed Implementation

[0031] 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 denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0032] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0033] It should also be understood that terms such as those defined in general dictionaries should be understood to have meanings consistent with their meanings in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless defined as here.

[0034] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, and / or groups thereof.

[0035] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0036] To facilitate understanding of the present invention, the present invention will be further explained and described below with reference to the accompanying drawings and specific embodiments. However, the specific embodiments do not constitute a limitation on the embodiments of the present invention.

[0037] Those skilled in the art should understand that the accompanying drawings are merely schematic diagrams of embodiments, and the components in the drawings are not necessarily essential for implementing the present invention.

[0038] This invention provides a highly generalized point cloud registration method for large-scale railway inspection. Point cloud registration, as a fundamental step in processing point cloud data collected by devices such as lidar, also has potential in railway inspection. By comparing and analyzing data obtained from previous inspections with baseline data through point cloud registration technology, monitoring of the railway and surrounding equipment can be achieved. Furthermore, the fusion of multiple railway point cloud data can obtain complete railway line information and surrounding environmental information, providing the optimal route for manual inspection, troubleshooting, and fault identification.

[0039] Example 1

[0040] In this embodiment 1, a highly generalized point cloud registration system for large-scale railway inspection is first provided, including: an acquisition module for acquiring point cloud data of the railway and its surrounding environment to form a dataset; a construction module for reading point cloud data at different time points in the dataset to construct a source point cloud; a preprocessing module for using the source point cloud 2 as the target point cloud and performing noise reduction and filtering preprocessing on the source point cloud 1 and the target point cloud; and an extraction module for inputting the source point cloud 1 and the target point cloud into a pre-constructed feature learning network based on deep learning, and by adjusting the parameter settings in the network, obtaining parameters suitable for processing point cloud data of large-scale railway inspection, and finally obtaining point cloud feature descriptors with strong generalization ability. The deep learning-based feature learning network includes a spatial point rotation module, a mapping process, and a feature extraction module. The spatial point rotation module eliminates the point cloud's degrees of freedom about the Z-axis and about the XY plane through rotation. The mapping process maps a local spherical point cloud into a cylindrical point cloud. The feature extraction module extracts features from the source and target point clouds after eliminating degrees of freedom. A calculation module uses the extracted features and a random sampling consensus algorithm to obtain the final transformation parameters between the source and target point clouds. The obtained rigidity change parameters are used to measure the registration accuracy, fulfilling the point cloud registration requirements during large-scale inspection.

[0041] In this embodiment 1, the above-described system is used to implement a highly generalized point cloud registration method for large-scale railway inspections, including: acquiring point cloud data of the railway and its surrounding environment to form a dataset; reading point cloud data from different time points in the dataset to construct a source point cloud; using source point cloud 2 as the target point cloud, performing denoising and filtering preprocessing on source point cloud 1 and the target point cloud; inputting source point cloud 1 and the target point cloud into a pre-constructed feature learning network based on deep learning, and by adjusting the parameter settings in the network, obtaining parameters suitable for processing point cloud data of large-scale railway inspections, and finally obtaining point cloud feature descriptors with strong generalization capabilities; wherein, based on The deep learning feature learning network includes a spatial point rotation module, a mapping process, and a feature extraction module. The spatial point rotation module eliminates the point cloud's degrees of freedom about the Z-axis and about the XY plane through rotation. The mapping process maps the local spherical point cloud into a cylindrical point cloud. The feature extraction module extracts features from the source and target point clouds after eliminating degrees of freedom. Based on the extracted features, a random sampling consensus algorithm is used to obtain the final transformation parameters between the source and target point clouds. The registration accuracy is measured based on the obtained rigidity change parameters to fulfill the point cloud registration requirements during the inspection process.

[0042] Specifically, the source point cloud 2 is used as the target point cloud, and the source point cloud 1 and the target point cloud are subjected to denoising and filtering preprocessing. This includes: removing useless noise points from the source point cloud 1 and the target point cloud; performing statistical filtering on the source point cloud 1 and the target point cloud; and determining whether to randomly select key points from the source point cloud 1 and the target point cloud based on the number of points in the point cloud obtained after voxel filtering, so as to finally obtain the key point cloud of the source point cloud 1 and the key point cloud of the target point cloud.

[0043] Obtain the coordinates of all points in the source point cloud 1 key point cloud. In the source point cloud 1, perform a nearest neighbor search centered on the key point and retain points within a certain range around the key point. Obtain the coordinates of all points in the target point cloud key point cloud. In the target point cloud, perform a nearest neighbor search centered on the key point and retain points within a certain range around the key point.

[0044] The process of the spatial point rotation module includes: rotating the normal vectors of the local spherical point clouds in the source point cloud 1 and the target point cloud to align with the Z-axis; dividing the rotated spherical point cloud into voxels according to the radial distance J, elevation angle K, and azimuth angle L; performing a nearest neighbor search again with the voxel center of each voxel as the sphere center, retaining a certain number of points; and rotating the center of each voxel around the Z-axis to align with the YZ plane.

[0045] The feature extraction module process includes: using a weight-shared multilayer perceptron and a max pooling function to extract shallow features from the local cylindrical point clouds of the source point cloud 1 and the target point cloud; and using cylindrical convolution to extract deep features from the extracted features.

[0046] Specifically, in this embodiment, the above-mentioned strong generalization point cloud registration method for large-scale railway inspection includes the following steps: S1: Use an UAV-borne LiDAR to collect point cloud data of the railway and its surrounding environment, and form a dataset based on the collected point cloud data; S2: Read data collected at different time points in the dataset, and construct source point cloud 1 and source point cloud 2 based on the point cloud data; S3: Preprocess the point cloud, including denoising and filtering; S4: Input the source point cloud and target point cloud into a pre-constructed feature learning network based on deep learning, and obtain parameters suitable for processing point cloud data of large-scale railway inspection by adjusting the parameter settings in the network, and finally obtain point cloud feature descriptors with strong generalization ability; S5: Based on the extracted features, use the random sampling consensus algorithm to obtain the transformation parameters between the final source point cloud and target point cloud; S6: Measure the registration accuracy based on the obtained rigid change parameters to complete the point cloud registration requirements in the large-scale inspection process.

[0047] Step S3 includes: S31: Using the source point cloud 2 as the target point cloud, remove useless noise points from the source point cloud 1 and the target point cloud;

[0048] S32: Perform statistical filtering on the source point cloud 1 and the target point cloud.

[0049] In this embodiment, the three-dimensional coordinates and other labels of the point cloud are read from the dataset, and points with fewer than 200 adjacent points within a range of 0.8m are removed.

[0050] S33: Based on the number of points in the point cloud obtained after voxel filtering, determine whether to randomly select key points for the source point cloud 1 and the target point cloud, and finally obtain the source point cloud key point cloud and the target point cloud key point cloud.

[0051] The voxel division radius during voxel filtering is specified to be 0.5m. Point cloud data is read from the dataset after voxel filtering. If each point cloud contains more than 5000 points, 5000 points are randomly selected as the key points of that point cloud. If each point cloud contains less than or equal to 5000 points, the point cloud is directly saved as the key point cloud of that point cloud.

[0052] Before step S4, i.e., before using the deep learning-based point cloud registration network, the point cloud registration method further includes:

[0053] (a1): Obtain the coordinates of all points in the key point cloud of source point cloud 1. In source point cloud 1, with the key point as the center, perform a nearest neighbor search with a search radius of 3.0cm, and retain points within a certain range around the key point.

[0054] (a2): Obtain the coordinates of all points in the target point cloud key point cloud. In the target point cloud, with the key point as the center, perform a nearest neighbor search with a search radius of 3.0cm, and retain points within a certain range around the key point.

[0055] Specifically, the set of points surrounding a keypoint can be defined as:

[0056] P s ={p i :||p i -p|| 2 <R}

[0057] Where p is the coordinate of the key point, and R is the search radius, with a value of 3.0.

[0058] Deep learning-based feature learning networks include a spatial point rotation module, a mapping process, and a feature extraction module.

[0059] The rotation module of the spatial points refers to eliminating the Z-axis degree of freedom and the degree of freedom about the XY plane of the point cloud by rotation; the mapping process is to map the local spherical point cloud into a cylindrical point cloud; the feature extraction module refers to extracting features from the source point cloud 1 and the target point cloud after eliminating the degree of freedom.

[0060] In the strongly generalized point cloud registration method for large-scale railway inspection, the process of the spatial point rotation module includes: (b1): rotating the normal vectors of the local spherical point clouds in the source point cloud 1 and the target point cloud to align with the Z-axis; (b2): dividing the rotated spherical point cloud into voxels according to the radial distance J, elevation angle K, and azimuth angle L; specifically, J = 9, K = 40, L = 80. (b3): ​​performing a nearest neighbor search again with the voxel center of each voxel as the sphere center, setting the search radius to 3.0 cm, and retaining 2048 points in the nearest neighbor set; specifically, the point set P obtained after the nearest neighbor search... jkl It can be defined as:

[0061]

[0062] Among them, v jkl Let R be the coordinates of the voxel center. V To find the radius, the value is equal to 0.5. This represents a local point cloud in a spherical shape.

[0063] Further, search for P jkl If the number of midpoints is greater than 2048, randomly select 2048 points from the point set; if the number is less than or equal to 2048, randomly copy points from the point set so that the number of points in the point set is equal to 2048.

[0064] (b4): Rotate the center of each voxel about the Z-axis until it is aligned with the YZ plane;

[0065] Specifically, the rotation matrix of the voxel at position (j,k,l) ​​can be defined as:

[0066]

[0067] In the strong generalization point cloud registration method for large-scale railway inspection, the feature extraction module process includes: (c1): using a multilayer perceptron of size 16 and a max pooling function to extract shallow features from the local cylindrical point clouds of the source point cloud 1 and the target point cloud.

[0068] Specifically, the extracted shallow features can be defined as:

[0069] L x =Max(Relu(MLPs(x)))

[0070] Where Max is the max pooling function, ReLU is the activation function, MLPs is a multilayer perceptron with weight sharing of size 16, and x is the input point cloud.

[0071] (c2): The extracted features are used to extract deeper features using cylindrical convolution;

[0072] Specifically, for the d-th dimension feature at position (j,k,l) ​​in the s-th layer, the cylindrical convolution can be defined as follows:

[0073]

[0074] Among them, R s It is the core size along the radial dimension; Y s and X s These are the height and width of the kernel, respectively; w is the parameter to be learned.

[0075] In the strongly generalized point cloud registration method for large-scale railway inspection, step S6 includes:

[0076] Registration accuracy is measured using rotation error and translation error.

[0077] For the corresponding two local point cloud images P h and Q h Rotational error can be defined as:

[0078]

[0079] Among them, R h These are the actual rotation matrices of two point clouds; This is the rotation matrix estimated using the random sampling consensus algorithm.

[0080] Translation error can be defined as:

[0081]

[0082] Among them, t h This represents the actual translation distance between the two point cloud images; The translational distance is estimated using the random sample consensus algorithm. During the registration process using a self-collected railway dataset, the registration results are visualized as follows: Figure 2 As shown.

[0083] Example 2

[0084] This embodiment 2 provides a non-transitory computer-readable storage medium for storing computer instructions. When the computer instructions are executed by a processor, they implement the strongly generalized point cloud registration method for large-scale railway inspection as described above.

[0085] Example 3

[0086] This embodiment 3 provides a computer device, including a memory and a processor. The processor and the memory communicate with each other. The memory stores program instructions that can be executed by the processor. The processor calls the program instructions to execute the strongly generalized point cloud registration method for large-scale railway inspection as described above.

[0087] Example 4

[0088] This embodiment 4 provides an electronic device, including: a processor, a memory, and a computer program; wherein, the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory, so that the electronic device executes instructions to implement the strongly generalized point cloud registration method for large-scale railway inspection as described above.

[0089] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0090] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0091] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0092] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment, whereby a series of operational steps are performed to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0093] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that, based on the technical solutions disclosed in the present invention, various modifications or variations that can be made by those skilled in the art without creative effort should be included within the scope of protection of the present invention.

Claims

1. A strongly generalized point cloud registration method for large-scale railway inspection, characterized in that, include: Acquire point cloud data of the railway and its surrounding environment to form a dataset; Read point cloud data from different time points in the dataset and construct the source point cloud; Using source point cloud 2 as target point cloud, denoising and filtering preprocessing are performed on source point cloud 1 and target point cloud; The source point cloud 1 and the target point cloud are input into a pre-constructed deep learning-based feature learning network. By adjusting the parameter settings in this network, parameters suitable for processing large-scale railway inspection point cloud data are obtained, ultimately acquiring point cloud feature descriptors with strong generalization capabilities. The deep learning-based feature learning network includes a spatial point rotation module, a mapping process, and a feature extraction module. The spatial point rotation module eliminates the point cloud's degrees of freedom about the Z-axis and about the XY plane through rotation. The mapping process maps the local spherical point cloud into a cylindrical point cloud. The feature extraction module extracts features from the source point cloud 1 and the target point cloud after eliminating degrees of freedom. The spatial point rotation module includes: rotating the normal vectors of the local spherical point clouds in the source point cloud 1 and the target point cloud to align with the Z-axis; dividing the rotated spherical point cloud into voxels according to radial distance J, elevation angle K, and azimuth angle L; performing a nearest neighbor search again with the voxel center as the sphere center, retaining a certain number of points; and rotating each voxel center around the Z-axis to align with the YZ plane. Based on the extracted features, the transformation parameters between the final source point cloud and the target point cloud are obtained using the random sampling consensus algorithm. The registration accuracy is measured based on the obtained rigidity change parameters to meet the point cloud registration requirements during large-scale inspection.

2. The strongly generalized point cloud registration method for large-scale railway inspection according to claim 1, characterized in that, Using source point cloud 2 as target point cloud, denoising and filtering preprocessing are performed on source point cloud 1 and target point cloud, including: removing useless noise points in source point cloud 1 and target point cloud; performing statistical filtering on source point cloud 1 and target point cloud; and determining whether to randomly select key points in source point cloud 1 and target point cloud based on the number of points in the point cloud obtained after voxel filtering, finally obtaining the key point cloud of source point cloud 1 and key point cloud of target point cloud.

3. The strongly generalized point cloud registration method for large-scale railway inspection according to claim 2, characterized in that, Obtain the coordinates of all points in the source point cloud 1 key point cloud. In the source point cloud 1, perform a nearest neighbor search centered on the key point and retain points within a certain range around the key point. Obtain the coordinates of all points in the target point cloud key point cloud. In the target point cloud, perform a nearest neighbor search centered on the key point and retain points within a certain range around the key point.

4. The strongly generalized point cloud registration method for large-scale railway inspection according to claim 1, characterized in that, The feature extraction module process includes: using a weight-shared multilayer perceptron and a max pooling function to extract shallow features from the local cylindrical point clouds of the source point cloud 1 and the target point cloud; and using cylindrical convolution to extract deep features from the extracted features.

5. The strongly generalized point cloud registration method for large-scale railway inspection according to claim 1, characterized in that, Registration accuracy is measured using rotation error and translation error.

6. A highly generalized point cloud registration system for large-scale railway inspection, characterized in that, include: The acquisition module is used to acquire point cloud data of the railway and its surrounding environment to form a dataset; The building module is used to read point cloud data from different time points in the dataset and construct the source point cloud; The preprocessing module uses source point cloud 2 as the target point cloud and performs noise reduction and filtering preprocessing on source point cloud 1 and target point cloud. The extraction module is used to input the source point cloud 1 and the target point cloud into a pre-constructed deep learning-based feature learning network. By adjusting the parameter settings in this network, parameters suitable for processing large-scale railway inspection point cloud data are obtained, ultimately acquiring point cloud feature descriptors with strong generalization capabilities. The deep learning-based feature learning network includes a spatial point rotation module, a mapping process, and a feature extraction module. The spatial point rotation module refers to eliminating the point cloud's degree of freedom about the Z-axis and its degree of freedom about the XY plane through rotation. The mapping process maps a local spherical point cloud into a cylindrical point cloud; the feature extraction module refers to extracting features from the source point cloud 1 and the target point cloud after eliminating degrees of freedom; the spatial point rotation module includes: rotating the normal vectors of the local spherical point clouds in the source point cloud 1 and the target point cloud to align with the Z-axis; dividing the rotated spherical point cloud into voxels according to the radial distance J, elevation angle K, and azimuth angle L; performing a nearest neighbor search again with the voxel center of each voxel as the sphere center, retaining a certain number of points; and rotating the center of each voxel around the Z-axis to align with the YZ plane; The calculation module is used to obtain the transformation parameters between the final source point cloud and the target point cloud using a random sampling consistency algorithm based on the extracted features; the registration accuracy is measured based on the obtained rigid change parameters to complete the point cloud registration requirements in the large-scale inspection process.

7. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium is used to store computer instructions, which, when executed by a processor, implement the strongly generalized point cloud registration method for large-scale railway inspection as described in any one of claims 1-5.

8. A computer device, characterized in that, The system includes a memory and a processor, which communicate with each other. The memory stores program instructions that can be executed by the processor, and the processor calls the program instructions to execute the strongly generalized point cloud registration method for large-scale railway inspection as described in any one of claims 1-5.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a computer program; wherein the processor is connected to the memory, and the computer program is stored in the memory. When the electronic device is running, the processor executes the computer program stored in the memory to cause the electronic device to execute instructions that implement the strongly generalized point cloud registration method for large-scale railway inspection as described in any one of claims 1-5.

Images