Steel rail three-dimensional profile reconstruction method and system based on point cloud automatic registration

Abstract

The invention provides a steel rail three-dimensional profile reconstruction method and system based on point cloud automatic registration, and relates to the technical field of rail traffic detection.The method comprises the steps that first information of the surface of a steel rail is obtained in the direction of the steel rail, and the first information comprises three-dimensional point cloud data and two-dimensional texture images of multiple continuous frames; data preprocessing is carried out on the first information to obtain second information, and the second information comprises first point cloud data and texture information; performing semantic segmentation and feature extraction based on the second information to obtain category information and a feature descriptor of each point cloud; performing point cloud matching based on the category information and the feature descriptors to obtain a point cloud matching result; and performing three-dimensional profile reconstruction of the steel rail based on a point cloud matching result to obtain a three-dimensional curved surface model of the steel rail. The problem that existing steel rail profile detection cannot comprehensively reflect three-dimensional profile features is solved.

Description

Technical Field

[0001] This invention relates to the field of rail transit inspection technology, and more specifically, to a method and system for reconstructing the three-dimensional profile of rails based on automatic point cloud registration. Background Technology

[0002] As a core component of modern transportation, the safety and reliability of rail transit are of paramount importance. As a key load-bearing component of the track structure, the profile accuracy of the rail directly affects the smoothness and safety of train operation. Traditional rail profile inspection methods mainly rely on manual measurement or two-dimensional inspection using a single sensor, which suffers from drawbacks such as low inspection efficiency, insufficient accuracy, and inability to fully reflect three-dimensional profile features.

[0003] In existing technologies, two-dimensional detection based on a single laser scan is difficult to construct a complete three-dimensional model, and the fusion of multi-sensor data suffers from large registration errors and high computational complexity, making it impossible to achieve high-precision reconstruction of the rail profile and failing to meet the needs of high-speed railways and heavy-haul railways for refined inspection of rail condition. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for reconstructing the three-dimensional profile of rails based on automatic point cloud registration, in order to improve the aforementioned problems. To achieve the above objective, the technical solution adopted by this invention is as follows:

[0005] Firstly, this application provides a method for reconstructing the three-dimensional profile of a rail based on automatic point cloud registration, including:

[0006] First information about the rail surface is obtained along the rail direction, the first information including three-dimensional point cloud data and two-dimensional texture images of multiple consecutive frames;

[0007] The first information is preprocessed to obtain the second information, which includes the first point cloud data and texture information.

[0008] Based on the second information, semantic segmentation and feature extraction are performed to obtain the category information and feature descriptor of each point cloud;

[0009] Point cloud matching is performed based on category information and feature descriptors to obtain point cloud matching results;

[0010] Based on the point cloud matching results, the three-dimensional profile of the rail is reconstructed to obtain the three-dimensional curved surface model of the rail.

[0011] Secondly, this application also provides a rail three-dimensional profile reconstruction system based on point cloud automatic registration, including:

[0012] The acquisition module is used to acquire first information about the rail surface along the rail direction. The first information includes three-dimensional point cloud data and two-dimensional texture images of multiple consecutive frames.

[0013] The preprocessing module is used to preprocess the first information to obtain the second information, which includes the first point cloud data and texture information.

[0014] The segmentation and extraction module is used to perform semantic segmentation and feature extraction based on the second information to obtain the category information and feature descriptor of each point cloud;

[0015] The matching module is used to perform point cloud matching based on category information and feature descriptors to obtain point cloud matching results.

[0016] The reconstruction module is used to reconstruct the three-dimensional profile of the rail based on the point cloud matching results, and obtain the three-dimensional curved surface model of the rail.

[0017] The beneficial effects of this invention are as follows: This invention improves rail inspection results through multi-stage collaborative automatic point cloud registration. It automatically filters point clouds through semantic segmentation and feature extraction, reducing data interference and accelerating processing speed. Simultaneously, it achieves high-precision point cloud alignment by adjusting the matching strategy according to the point cloud curvature using an adaptive distance threshold algorithm, overcoming the limitations of traditional algorithms. Furthermore, during 3D profile reconstruction, it uses moving least squares to fuse point clouds and eliminate inter-frame gaps, and when using NURBS surface fitting, it accurately restores the rail profile by leveraging control points and weight parameters. Therefore, the method of this invention can comprehensively guarantee detection accuracy and efficiency, providing strong support for the safe operation of rail transit.

[0018] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a schematic diagram of the process of the rail three-dimensional profile reconstruction method based on point cloud automatic registration as described in the embodiments of the present invention;

[0021] Figure 2This is a schematic diagram of the three-dimensional profile reconstruction of the rail in an embodiment of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0023] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0024] Example 1:

[0025] This embodiment provides a method for reconstructing the three-dimensional profile of a rail based on automatic point cloud registration.

[0026] See Figure 1 The figure shows that the method includes steps S1, S2, S3, S4, and S5.

[0027] Step S1: Obtain first information about the rail surface along the rail direction. The first information includes three-dimensional point cloud data and two-dimensional texture images of multiple consecutive frames.

[0028] In this embodiment, a high-precision line laser 3D camera (such as the Keyence LJ-V7000 series) is used to acquire 3D point cloud data of the rail surface at a sampling frequency of not less than 1kHz, and the scanning accuracy reaches the sub-millimeter level (X-axis accuracy ≤0.1mm, Z-axis accuracy ≤0.05mm).

[0029] Equipped with an industrial-grade CCD camera (resolution ≥ 5 million pixels), it synchronously acquires two-dimensional texture images of the rail surface to assist in semantic segmentation and feature matching of three-dimensional point cloud data.

[0030] It integrates a GNSS positioning module (accuracy ≤0.1m), an inertial navigation system (INS, attitude angle accuracy ≤0.1°) and an odometer to achieve spatiotemporal synchronization of multi-sensor data, ensuring accurate correspondence between point cloud data and image data.

[0031] When acquiring the first information about the rail surface along the rail direction, a multi-source data acquisition module can be mounted on a portable inspection trolley or rail inspection train, moving along the rail direction at a speed of 1-5 m / s, and simultaneously acquiring three-dimensional point cloud data, two-dimensional image data and positioning information of the rail surface.

[0032] Step S2: Perform data preprocessing on the first information to obtain the second information, which includes the first point cloud data and texture information;

[0033] Step S2 includes:

[0034] Step S21: Use bilateral filtering algorithm and statistical filtering to remove random noise and outliers in the 3D point cloud data, retain effective feature points, and obtain filtered point cloud data;

[0035] In this embodiment, the original three-dimensional point cloud data is subjected to noise filtering. A bilateral filtering algorithm (filter radius ≤ 2mm) is used to remove high-frequency noise, and then statistical filtering (setting the distance threshold to twice the standard deviation of the mean) is used to remove outliers, resulting in filtered point cloud data.

[0036] Step S22: Based on the spatial location and scanning range of two adjacent frames, determine the overlapping area of ​​the filtered point cloud data of two adjacent frames;

[0037] Step S23: Perform feature matching based on the overlapping region to obtain matching features, and stitch together the filtered point cloud data of each frame based on the matching features and the adaptive distance threshold algorithm to obtain the corresponding first point cloud data;

[0038] In this embodiment, based on feature matching of overlapping regions, the adaptive distance thresholding algorithm (AD-ICP) is used to realize the automatic stitching of point cloud data of adjacent scan frames. Taking the first frame of three-dimensional point cloud data as the reference, the transformation matrix is ​​iteratively optimized to realize the initial stitching of multiple frame point clouds, so that the stitching error is less than 0.2mm.

[0039] Step S24: Extract features from each frame of the two-dimensional texture image to obtain the corresponding texture information.

[0040] Step S3: Based on the second information, perform semantic segmentation and feature extraction to obtain the category information and feature descriptor of each point cloud;

[0041] Step S3 includes:

[0042] Step S31: Construct and train the SegFormer semantic segmentation network;

[0043] In this embodiment, a labeled dataset containing multiple categories such as the top surface, side surface, and arc transition area of ​​the rail is constructed to train the SegFormer semantic segmentation network.

[0044] Step S32: Perform feature fusion on the first point cloud data and the corresponding texture information to obtain the stitched features;

[0045] Step S33: Input the concatenated features into the SegFormer semantic segmentation network for point-by-point classification to obtain the category information of each point cloud;

[0046] In this embodiment, the SegFormer semantic segmentation network is used for point-by-point classification to accurately identify the contours of various rail components (such as switch rails, main rails, and wing rails), achieving a segmentation accuracy of 98%.

[0047] Step S34: Calculate the geometric properties of each point cloud, including three-dimensional coordinates, normal vector, and curvature;

[0048] Step S35: Calculate the feature descriptor of the point cloud based on geometric attributes.

[0049] In this embodiment, feature descriptors of PFH (Point Feature Histogram) or FPFH (Fast Point Feature Histogram) are extracted from the point cloud after semantic segmentation for feature matching in subsequent point cloud registration, thereby improving registration efficiency and accuracy.

[0050] Step S4: Perform point cloud matching based on category information and feature descriptors to obtain point cloud matching results;

[0051] Step S4 includes:

[0052] Step S41: Filter out the point cloud related to the rail body based on the category information to obtain the rail point cloud;

[0053] Step S42: Calculate the Hamming distance based on the feature descriptors of the rail point cloud in two adjacent frames, and perform coarse registration based on the Hamming distance to obtain the initial matching points;

[0054] Step S43: After removing mismatched pairs from the initial matching points based on the random sampling consensus algorithm, calculate the initial transformation matrix;

[0055] In this embodiment, the initial transformation matrix is ​​actually obtained based on the matching method of three-dimensional local features. Through fast matching of feature descriptors (Hamming distance threshold ≤ 0.6), the initial correspondence between different scan frames is established. The Random Sample Consensus (RANSAC) algorithm is used to remove mismatched point pairs, thereby achieving coarse registration of point cloud data and making the initial registration error less than 5mm.

[0056] Step S44: Perform fine registration based on the initial transformation matrix to obtain the point cloud matching result.

[0057] Step S44 includes:

[0058] Step S441: Calculate the local curvature of the point cloud based on the initial transformation matrix;

[0059] Step S442: Dynamically adjust the matching distance threshold by adjusting the local curvature of the point cloud, and optimize the transformation parameters based on the matching distance threshold and the least squares method;

[0060] Step S443: Optimize the initial transformation matrix by transforming parameters to obtain point cloud matching results. The point cloud matching results are multiple consecutive frames of rail point cloud data in which all rail point clouds have been precisely aligned in the same coordinate system.

[0061] In this embodiment, a point cloud fine registration based on an adaptive distance threshold is introduced. The matching distance threshold is automatically adjusted according to the local curvature of the point cloud. The larger the local curvature of the point cloud, the smaller the matching distance threshold, with a value range of 0.1-1mm. Then, by minimizing the sum of squared distances between corresponding point pairs, the initial transformation matrix is ​​iteratively optimized to achieve fine registration of the point cloud, making the fine registration error less than 0.1mm.

[0062] Step S5: Reconstruct the three-dimensional profile of the rail based on the point cloud matching results to obtain the three-dimensional curved surface model of the rail.

[0063] Step S5 includes:

[0064] Step S51: The point clouds in the point cloud matching results are fused using the moving least squares method, and the seams between adjacent point clouds are smoothed by weighted averaging to generate a three-dimensional point cloud model of the rail.

[0065] Step S52: Perform surface fitting based on the 3D point cloud model and NURBS surface fitting algorithm to obtain the initial surface model of the rail;

[0066] Step S53: Optimize the initial surface model by adjusting the control points and weight parameters to obtain the three-dimensional surface model of the rail.

[0067] In this embodiment, a point cloud volume fusion technology based on moving least squares is used to smooth the registered point cloud. The seams between adjacent point clouds are smoothed by weighted averaging to generate a complete and continuous three-dimensional point cloud model of the rail. Then, the NURBS (Non-Uniform Rational B-Spline) surface fitting algorithm is used to perform surface modeling on the fused three-dimensional point cloud model to accurately restore the rail profile curve, so that the fitting error is less than 0.05mm.

[0068] Example 2:

[0069] In this embodiment, a section of rail on the Beijing-Shanghai High-Speed ​​Railway is used as an example to reconstruct the three-dimensional profile of the rail. A Keyence LJ-V7080 line laser 3D camera with a scanning range of 500mm and an accuracy of 0.02mm is used; a Baslerac A2500-14gm CCD camera with a resolution of 2592×1944 pixels is used; and a Trimble BD970 positioning module with a GNSS positioning accuracy of 0.05m and an INS attitude angle accuracy of 0.05° is used. A self-developed portable rail inspection trolley, weighing less than 20kg and with a battery life of approximately 8 hours, is also used.

[0070] A detection trolley travels at a speed of 2 m / s on a section of rail along the Beijing-Shanghai High-Speed ​​Railway, collecting rail data over a length of 10 kilometers. Then, a 3D surface model of the rail is constructed using a 3D rail profile reconstruction method, resulting in the following... Figure 2 The three-dimensional surface model shown is as follows. After testing, the reconstructed three-dimensional profile of the rail obtained from the three-dimensional surface model meets the requirements of high-speed railway line maintenance regulations. Defects such as top surface wear (maximum wear amount 0.3mm) and side cracks (length 2mm) were detected, with an accuracy rate of 98%.

[0071] In summary, the method of this invention significantly improves the rail inspection effect through multi-stage collaboration. Semantic segmentation and feature extraction automatically filter the rail body point cloud, reducing redundant data interference and accelerating processing speed. Simultaneously, the adaptive distance threshold algorithm flexibly adjusts the matching strategy based on the point cloud curvature, enabling high-precision alignment of point clouds across frames, effectively overcoming the limitations of traditional fixed threshold algorithms and laying a solid foundation for subsequent reconstruction. In the 3D profile reconstruction stage, point cloud fusion based on moving least squares seamlessly stitches together multiple frame point clouds, eliminating inter-frame gaps and generating a continuous and complete point cloud model. Furthermore, through NURBS surface fitting and utilizing control points and weight parameters, the complex profile of the rail is accurately reconstructed, constructing a highly realistic 3D surface model. Therefore, it comprehensively ensures the accuracy and efficiency of rail inspection, providing strong support for the safe and stable operation of rail transit.

[0072] Example 3:

[0073] This embodiment provides a rail 3D profile reconstruction system based on point cloud automatic registration, the system comprising:

[0074] The acquisition module is used to acquire first information about the rail surface along the rail direction. The first information includes three-dimensional point cloud data and two-dimensional texture images of multiple consecutive frames.

[0075] The preprocessing module is used to preprocess the first information to obtain the second information, which includes the first point cloud data and texture information.

[0076] The segmentation and extraction module is used to perform semantic segmentation and feature extraction based on the second information to obtain the category information and feature descriptor of each point cloud;

[0077] The matching module is used to perform point cloud matching based on category information and feature descriptors to obtain point cloud matching results.

[0078] The reconstruction module is used to reconstruct the three-dimensional profile of the rail based on the point cloud matching results, and obtain the three-dimensional curved surface model of the rail.

[0079] The preprocessing module includes:

[0080] The filtering unit is used to remove random noise and outliers from the 3D point cloud data by employing bilateral filtering and statistical filtering algorithms, while retaining effective feature points to obtain filtered point cloud data.

[0081] The acquisition unit is used to determine the overlapping area of ​​the filtered point cloud data of two adjacent frames based on the spatial position and scanning range of the two adjacent frames.

[0082] The matching unit is used to perform feature matching based on the overlapping region to obtain matching features, and to stitch together the filtered point cloud data of each frame based on the matching features and the adaptive distance threshold algorithm to obtain the corresponding first point cloud data.

[0083] The extraction unit is used to extract features from each frame of two-dimensional texture image to obtain the corresponding texture information.

[0084] The segmentation and extraction module includes:

[0085] Building units are used to build and train the SegFormer semantic segmentation network;

[0086] The fusion unit is used to fuse the first point cloud data and the corresponding texture information to obtain spliced ​​features;

[0087] The classification unit is used to input the concatenated features into the SegFormer semantic segmentation network for point-by-point classification, thereby obtaining the category information of each point cloud.

[0088] The first computing unit is used to calculate the geometric properties of each point cloud, including three-dimensional coordinates, normal vectors, and curvature.

[0089] The second computational unit is used to compute feature descriptors of the point cloud based on geometric attributes.

[0090] The matching module includes:

[0091] The filtering unit is used to filter out the point cloud related to the rail body based on category information to obtain the rail point cloud.

[0092] The third calculation unit is used to calculate the Hamming distance based on the feature descriptors of the rail point cloud of two adjacent frames, and to perform coarse registration based on the Hamming distance to obtain the initial matching points.

[0093] The fourth calculation unit is used to calculate the initial transformation matrix after removing mismatched point pairs from the initial matching points based on the random sampling consensus algorithm;

[0094] The fine registration unit is used to perform fine registration based on the initial transformation matrix to obtain the point cloud matching result.

[0095] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0096] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for reconstructing the three-dimensional profile of a rail based on automatic point cloud registration, characterized in that, include: First information about the rail surface is obtained along the rail direction, the first information including three-dimensional point cloud data and two-dimensional texture images of multiple consecutive frames; The first information is preprocessed to obtain the second information, which includes the first point cloud data and texture information. Based on the second information, semantic segmentation and feature extraction are performed to obtain the category information and feature descriptor of each point cloud; Point cloud matching is performed based on category information and feature descriptors to obtain point cloud matching results; Based on the point cloud matching results, the three-dimensional profile of the rail is reconstructed to obtain the three-dimensional curved surface model of the rail.

2. The method for reconstructing the three-dimensional profile of rails based on automatic point cloud registration according to claim 1, characterized in that... The step of preprocessing the first information to obtain the second information includes: By employing bilateral filtering and statistical filtering, random noise and outliers in the 3D point cloud data are removed, while effective feature points are retained to obtain filtered point cloud data. Based on the spatial location and scanning range of two adjacent frames, the overlapping area of ​​the filtered point cloud data of two adjacent frames is determined. Feature matching is performed based on overlapping regions to obtain matching features. Then, based on the matching features and an adaptive distance threshold algorithm, the filtered point cloud data of each frame is stitched together to obtain the corresponding first point cloud data. Feature extraction is performed on each frame of the two-dimensional texture image to obtain the corresponding texture information.

3. The method for reconstructing the three-dimensional profile of rails based on automatic point cloud registration according to claim 1, characterized in that... The process of obtaining the category information and feature descriptor for each point cloud includes: Construct and train the SegFormer semantic segmentation network; The first point cloud data and the corresponding texture information are fused to obtain the spliced ​​features; The concatenated features are input into the SegFormer semantic segmentation network for point-by-point classification to obtain the category information of each point cloud. Calculate the geometric properties of each point cloud, including three-dimensional coordinates, normal vector, and curvature; Calculate feature descriptors for point clouds based on geometric attributes.

4. The method for reconstructing the three-dimensional profile of rails based on automatic point cloud registration according to claim 1, characterized in that... The point cloud matching based on category information and feature descriptors, to obtain the point cloud matching result, includes: Based on category information, point clouds related to the rail body are filtered out to obtain rail point clouds; Hamming distance is calculated based on the feature descriptors of rail point clouds from two adjacent frames, and coarse registration is performed based on the Hamming distance to obtain the initial matching points. After removing mismatched pairs from the initial matching points based on the random sampling consensus algorithm, the initial transformation matrix is ​​calculated. Fine registration is performed based on the initial transformation matrix to obtain the point cloud matching results.

5. The method for reconstructing the three-dimensional profile of rails based on automatic point cloud registration according to claim 4, characterized in that... The fine registration based on the initial transformation matrix to obtain the point cloud matching result includes: Calculate the local curvature of the point cloud based on the initial transformation matrix; The matching distance threshold is dynamically adjusted by adjusting the local curvature of the point cloud, and the transformation parameters are optimized based on the matching distance threshold and the least squares method. The initial transformation matrix is ​​optimized by transforming parameters to obtain point cloud matching results. The point cloud matching results are multiple consecutive frames of rail point cloud data in which all rail point clouds have been precisely aligned in the same coordinate system.

6. The method for reconstructing the three-dimensional profile of rails based on automatic point cloud registration according to claim 1, characterized in that... The method of reconstructing the three-dimensional profile of the rail based on point cloud matching results to obtain a three-dimensional curved surface model of the rail includes: The point clouds in the point cloud matching results are fused using the moving least squares method, and the seams between adjacent point clouds are smoothed by weighted averaging to generate a three-dimensional point cloud model of the rail. The initial surface model of the rail is obtained by fitting the surface based on the 3D point cloud model and the NURBS surface fitting algorithm. The initial surface model is optimized by adjusting the control points and weight parameters to obtain the three-dimensional surface model of the rail.

7. A three-dimensional profile reconstruction system for rails based on automatic point cloud registration, characterized in that, include: The acquisition module is used to acquire first information about the rail surface along the rail direction. The first information includes three-dimensional point cloud data and two-dimensional texture images of multiple consecutive frames. The preprocessing module is used to preprocess the first information to obtain the second information, which includes the first point cloud data and texture information. The segmentation and extraction module is used to perform semantic segmentation and feature extraction based on the second information to obtain the category information and feature descriptor of each point cloud; The matching module is used to perform point cloud matching based on category information and feature descriptors to obtain point cloud matching results. The reconstruction module is used to reconstruct the three-dimensional profile of the rail based on the point cloud matching results, and obtain the three-dimensional curved surface model of the rail.

8. The rail three-dimensional profile reconstruction system based on point cloud automatic registration according to claim 7, characterized in that, The preprocessing module includes: The filtering unit is used to remove random noise and outliers from the 3D point cloud data by employing bilateral filtering and statistical filtering algorithms, while retaining effective feature points to obtain filtered point cloud data. The acquisition unit is used to determine the overlapping area of ​​the filtered point cloud data of two adjacent frames based on the spatial position and scanning range of the two adjacent frames. The matching unit is used to perform feature matching based on the overlapping region to obtain matching features, and to stitch together the filtered point cloud data of each frame based on the matching features and the adaptive distance threshold algorithm to obtain the corresponding first point cloud data. The extraction unit is used to extract features from each frame of two-dimensional texture image to obtain the corresponding texture information.

9. The rail three-dimensional profile reconstruction system based on point cloud automatic registration according to claim 7, characterized in that, The segmentation and extraction module includes: Building units are used to build and train the SegFormer semantic segmentation network; The fusion unit is used to fuse the first point cloud data and the corresponding texture information to obtain spliced ​​features; The classification unit is used to input the concatenated features into the SegFormer semantic segmentation network for point-by-point classification, thereby obtaining the category information of each point cloud. The first computing unit is used to calculate the geometric properties of each point cloud, including three-dimensional coordinates, normal vectors, and curvature. The second computational unit is used to compute feature descriptors of the point cloud based on geometric attributes.

10. The rail three-dimensional profile reconstruction system based on point cloud automatic registration according to claim 7, characterized in that, The matching module includes: The filtering unit is used to filter out the point cloud related to the rail body based on category information to obtain the rail point cloud. The third calculation unit is used to calculate the Hamming distance based on the feature descriptors of the rail point cloud of two adjacent frames, and to perform coarse registration based on the Hamming distance to obtain the initial matching points. The fourth calculation unit is used to calculate the initial transformation matrix after removing mismatched point pairs from the initial matching points based on the random sampling consensus algorithm; The fine registration unit is used to perform fine registration based on the initial transformation matrix to obtain the point cloud matching result.

Images